CN118093839B

CN118093839B - Knowledge operation question-answer dialogue processing method and system based on deep learning

Info

Publication number: CN118093839B
Application number: CN202410496215.0A
Authority: CN
Inventors: 李壮壮
Original assignee: Beijing Zhongguancun Kejin Technology Co Ltd
Current assignee: Beijing Zhongguancun Kejin Technology Co Ltd
Priority date: 2024-04-24
Filing date: 2024-04-24
Publication date: 2024-07-02
Anticipated expiration: 2044-04-24
Also published as: CN118093839A

Abstract

The application relates to the technical field of artificial intelligence, and provides a knowledge operation question-answer dialogue processing method and system based on deep learning. Specifically, the method can automatically answer the questions presented by the user, provide personalized knowledge service and perform intelligent interaction with the user. The method not only reduces the burden of manual customer service and improves the service efficiency, but also can more accurately meet the demands of users and improve the satisfaction degree of the users. Meanwhile, by capturing the change of the user demand and adjusting the service strategy in real time, the enterprise can better adapt to market change and keep competitive advantage.

Description

Knowledge operation question-answer dialogue processing method and system based on deep learning

Technical Field

The application relates to the technical field of artificial intelligence, in particular to a knowledge operation question-answer dialogue processing method and system based on deep learning.

Background

With the popularization of the internet and the rapid development of electronic commerce, more and more enterprises begin to pay attention to knowledge operation, namely, by effectively managing and utilizing knowledge resources, user experience and service quality are improved. In this context, how to accurately understand the questions of the user and provide personalized knowledge services becomes a highly urgent problem to be solved.

Traditional knowledge operation methods often rely on manual customer service or simple keyword matching technology, and cannot fully understand the semantics and requirements of users, so that service efficiency is low and user satisfaction is low. In recent years, deep learning technology has made remarkable progress in the field of natural language processing, and has provided possibilities for achieving higher-level semantic understanding and intelligent interaction.

However, existing deep learning systems still face some challenges in handling user questions. First, user questions often contain complex semantic structures and multi-level information requirements, requiring a system with powerful question understanding capabilities. Second, user demand may jump over time and with changes in context, requiring the system to be able to capture and accommodate such changes in real-time. Finally, to implement a personalized knowledge service, the system needs to be able to generate answers matching the questions of the user based thereon.

Disclosure of Invention

In order to improve the problems, the application provides a knowledge operation question-answering dialogue processing method and system based on deep learning.

In a first aspect, an embodiment of the present application provides a knowledge operation question-answer dialogue processing method based on deep learning, which is applied to a deep learning system, and the method includes:

Acquiring a target knowledge operation question text, wherein the target knowledge operation question text comprises a plurality of online user question paragraphs with semantic association;

Determining a corresponding multi-order question sentence of each online user question paragraph in the plurality of online user question paragraphs; each of the multiple-order question sentences is obtained by adjusting text fine granularity of the corresponding online user question paragraph, and the text fine granularity of each of the multiple-order question sentences is in a decreasing trend;

Respectively carrying out problem understanding vector mining on the multi-order question sentences corresponding to each online user question paragraph to obtain corresponding problem understanding characterization vectors of each online user question paragraph; the question understanding characterization vector is used for reflecting the embedded semantic vector of each online user question paragraph under the different linear vector capacity weights and connected with user demand jump;

And carrying out enterprise user demand jump analysis on any two continuous online user question paragraphs in the plurality of online user question paragraphs according to the question understanding characterization vectors respectively corresponding to the two online user question paragraphs under the capacity weight of at least one linear vector.

In some technical solutions, the performing problem understanding vector mining on the multi-order question sentences corresponding to each online user question paragraph to obtain the problem understanding characterization vectors corresponding to each online user question paragraph respectively includes:

Respectively carrying out problem understanding vector mining on the multi-order question sentences corresponding to each online user question paragraph to obtain a multi-order problem understanding vector relation network corresponding to each online user question paragraph; the linear vector capacity weights and the number of semantic vector clusters corresponding to the problem understanding vector relation network of different feature orders in the multi-order problem understanding vector relation network are different;

Obtaining question element linear variables corresponding to semantic vector clusters in each order question understanding vector relation network corresponding to each online user question paragraph according to the multi-order question understanding vector relation network corresponding to each online user question paragraph;

and determining the question understanding characterization vector corresponding to each online user question paragraph according to the question element linear variable corresponding to each online user question paragraph.

In some technical solutions, the obtaining, according to the multi-order problem understanding vector relation network corresponding to each online user question paragraph, a question element linear variable corresponding to each semantic vector cluster in each order problem understanding vector relation network corresponding to each online user question paragraph includes:

Determining basic question element linear variables of each question understanding vector unit in each order of the question understanding vector relation network in the multi-order question understanding vector relation network according to the multi-order question understanding vector relation network corresponding to each online user question paragraph;

Determining a first semantic element linear variable and a second semantic element linear variable of the basic question element linear variable on a target semantic feature coordinate system; the first semantic element linear variable is used for indicating a semantic vector of the problem understanding vector unit under the word level attention index; the second semantic element linear variable is used for indicating a semantic vector of the problem understanding vector unit under the sentence-level attention index;

for each semantic vector cluster, determining a first local problem understanding vector of the semantic vector cluster according to a first semantic element linear variable corresponding to each problem understanding vector unit in the semantic vector cluster and the total number of the problem understanding vector units contained in the semantic vector cluster;

for each semantic vector cluster, determining a second local problem understanding vector of the semantic vector cluster according to a second semantic element linear variable corresponding to each problem understanding vector unit in the semantic vector cluster and the total number of the problem understanding vector units contained in the semantic vector cluster;

And determining question element linear variables corresponding to semantic vector clusters in each order of question understanding vector relation network corresponding to each online user question paragraph according to the first local question understanding vector and the corresponding second local question understanding vector.

In some technical solutions, the question understanding token vector includes word level text semantics for reflecting a word level attention index and sentence level text semantics for reflecting a sentence level attention index;

The step of analyzing the enterprise user demand jump of the two online user question paragraphs under the capacity weight of at least one linear vector according to the respective question understanding characterization vectors of the two online user question paragraphs comprises the following steps:

Determining word level comparison results among the corresponding semantic vector clusters of the two online user question paragraphs under the same linear vector capacity weight according to the word level text semantics of the two online user question paragraphs respectively corresponding to the same linear vector capacity weight;

determining sentence level comparison results among the corresponding semantic vector clusters of the two online user question paragraphs under the same linear vector capacity weight according to the sentence level text semantics of the two online user question paragraphs respectively corresponding to the same linear vector capacity weight;

And carrying out enterprise user demand jump analysis on the two online user question paragraphs under the weight of at least one linear vector capacity according to the word level comparison result and the sentence level comparison result.

In some technical solutions, the performing, according to the word level comparison result and the sentence level comparison result, the enterprise user demand jump analysis on the two online user question segments under at least one linear vector capacity weight includes:

Taking the maximum linear vector capacity weight in the plurality of different linear vector capacity weights as the current linear vector capacity weight;

Comparing the question requirement change weight corresponding to the question requirement change information of each corresponding semantic vector cluster under the current linear vector capacity weight with the corresponding question requirement label value;

if the comparison analysis result shows that the enterprise user demand jump exists under the current linear vector capacity weight, determining that the enterprise user demand jump exists in the analyzed online user question paragraph;

If the comparison and analysis result shows that the enterprise user demand jump does not exist under the current linear vector capacity weight, taking the next linear vector capacity weight in the plurality of different linear vector capacity weights as the current linear vector capacity weight;

and jumping to the step of comparing and analyzing the question requirement change weight corresponding to the question requirement change information of each corresponding semantic vector cluster under the current linear vector capacity weight with the corresponding question requirement label value until the jump capturing termination requirement is met.

In some technical solutions, if the comparison analysis result indicates that there is an enterprise user demand jump under the current linear vector capacity weight, before determining that the analyzed online user question section has an enterprise user demand jump, the method further includes:

Under the current linear vector capacity weight, determining a current quantitative statistical result of which the comparison analysis result accords with a set identification condition;

If the current quantitative statistical result is greater than or equal to a preset statistical threshold corresponding to the current linear vector capacity weight, determining that the comparison analysis result represents that enterprise user demand jump exists under the current linear vector capacity weight;

if the current quantitative statistical result is smaller than a preset statistical threshold corresponding to the current linear vector capacity weight, determining that the comparison analysis result represents that no enterprise user demand jump exists under the current linear vector capacity weight;

The preset statistical threshold is related to the number of semantic vector clusters corresponding to the current linear vector capacity weight.

In some aspects, the comparison analysis results include a first comparison analysis view indicating a word level comparison result and a second comparison analysis view indicating a sentence level comparison result;

under the current linear vector capacity weight, determining a current quantitative statistical result of which the comparison analysis result meets a set identification condition, wherein the method comprises the following steps of:

determining a first quantitative statistical result of the first comparison analysis viewpoint meeting a set identification condition under the current linear vector capacity weight, and determining a second quantitative statistical result of the second comparison analysis viewpoint meeting the set identification condition;

and determining the current quantization statistic according to the first quantization statistic and the second quantization statistic.

In some aspects, the method further comprises:

And if the enterprise user demand jump analysis result represents that the analyzed online user question paragraphs have enterprise user demand jumps, carrying out demand jump annotation on the analyzed online user question paragraphs.

In some technical solutions, if the analysis result of the enterprise user demand jump characterizes that the analyzed online user question paragraph has enterprise user demand jump, the method for marking the demand jump of the analyzed online user question paragraph includes:

If the enterprise user demand jump analysis result represents that the analyzed online user question paragraphs have enterprise user demand jumps, carrying out demand jump association on the analyzed online user question paragraphs;

and marking answer retrieval keywords of the analyzed online user question paragraphs related by the demand jump to obtain target online user question paragraphs marked by the demand jump, wherein the question sentence influence weight corresponding to the target online user question paragraphs is smaller than that of the question sentences corresponding to the analyzed online user question paragraphs.

In a second aspect, embodiments of the present application provide a deep learning system comprising at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes computer-executable instructions stored in the memory such that the at least one processor performs the method of the first aspect.

In a third aspect, embodiments of the present application provide a computer readable storage medium having stored thereon a computer program which, when run, implements the method of the first aspect.

The application provides a natural language processing and semantic understanding method based on a deep learning technology. The method comprises the steps of firstly obtaining a target knowledge operation question text, wherein the target knowledge operation question text comprises a plurality of online user question paragraphs with semantic association. And then, carrying out text fine granularity adjustment on each online user question paragraph to generate a multi-order question sentence so as to more comprehensively understand the question content of the user. And then, carrying out problem understanding vector mining on the multi-order question sentences by using a deep learning technology to obtain a problem understanding characterization vector so as to reflect the embedded semantic vectors of the users under different linear vector capacity weights. And finally, carrying out enterprise user demand jump analysis on the continuous user question paragraphs so as to capture dynamic changes of user demands.

By the application of the method and the system, enterprises can obviously improve knowledge operation efficiency and enhance user experience and loyalty. Specifically, the method can automatically answer the questions presented by the user, provide personalized knowledge service and perform intelligent interaction with the user. The method not only reduces the burden of manual customer service and improves the service efficiency, but also can more accurately meet the demands of users and improve the satisfaction degree of the users. Meanwhile, by capturing the change of the user demand and adjusting the service strategy in real time, the enterprise can better adapt to market change and keep competitive advantage.

Drawings

Fig. 1 is a flowchart of a knowledge operation question-answer dialogue processing method based on deep learning according to an embodiment of the present application.

Fig. 2 is a schematic structural diagram of a deep learning system 200 according to an embodiment of the application.

Detailed Description

In order to better understand the above technical solutions, the following detailed description of the technical solutions of the present application is made by using the accompanying drawings and specific embodiments, and it should be understood that the specific features of the embodiments and the embodiments of the present application are detailed descriptions of the technical solutions of the present application, and not limiting the technical solutions of the present application, and the technical features of the embodiments and the embodiments of the present application may be combined with each other without conflict.

Fig. 1 shows a knowledge operation question-answer dialogue processing method based on deep learning, which is applied to a deep learning system and comprises the following steps 101-104.

Step 101, acquiring a target knowledge operation question text, wherein the target knowledge operation question text comprises a plurality of online user question paragraphs with semantic association.

Step 102, determining a corresponding multi-order question sentence of each online user question paragraph in the plurality of online user question paragraphs; each of the multiple-order question sentences is obtained by adjusting text fine granularity of the corresponding online user question paragraph, and the text fine granularity of each of the multiple-order question sentences is in a decreasing trend.

Step 103, respectively carrying out problem understanding vector mining on the multi-order question sentences corresponding to each online user question paragraph to obtain problem understanding characterization vectors corresponding to each online user question paragraph; the question understanding characterization vector is used for reflecting the embedded semantic vector of each online user question paragraph under the different linear vector capacity weights and connected with user demand jump.

Step 104, for any two continuous online user question paragraphs in the plurality of online user question paragraphs, performing enterprise user demand jump analysis on the two online user question paragraphs under the weight of at least one linear vector capacity according to the question understanding characterization vectors respectively corresponding to the two online user question paragraphs.

In order to better understand the above technical solution, the following description is firstly made in its entirety through 3 application scenarios, then the terms related to the steps are explained, and then each step is described in detail.

Application scenario 1: user demand jump analysis in intelligent customer service system

In an intelligent customer service system of a large e-commerce platform, a deep learning system is used to process and analyze online consultation questions of users to more accurately understand user needs and provide personalized services.

The deep learning system captures a series of user question text from the interaction log of the intelligent customer service, which forms the target knowledge operation question text. These text includes continuous questions of the user about various aspects of consulting merchandise information, order status, after-market services, etc., each question containing a certain requirement or point of interest of the user.

For each user's question paragraph, the deep learning system uses natural language processing techniques to conduct fine-grained text analysis. The system carries out layering processing on each question paragraph to generate a multi-order question sentence. These multi-level question sentences are arranged according to the decreasing trend of text granularity, and summarize the question contents of users layer by layer from concrete to abstract.

Next, the deep learning system performs mining of the question understanding vector for each of the multi-order question sentences. Through a trained deep learning model, the system converts each question sentence into a vector in a high-dimensional semantic space, i.e., a question understanding token vector. This vector contains not only the literal meaning of the question sentence, but also the deep semantic information related to the user's demand jump.

When a user continuously presents a plurality of questions, the deep learning system can perform user demand jump analysis on the continuous questions. The system will compare the problem understanding token vectors of two adjacent problems, analyze the differences and links between these two vectors under different linear vector capacity weights. Through the analysis, the system can capture the change and jump of the requirement of the user in the consultation process, thereby providing basis for the follow-up personalized service.

For example, the user may first ask for detailed information of a certain item and then jump to inquire about the shipping time of the order. Through the user demand jump analysis of the deep learning system, the intelligent customer service can rapidly identify the change of the demand and timely provide information required by the user, so that the user experience and satisfaction are improved.

Application scenario 2: student problem answering system of online education platform

In an online educational platform, a deep learning system is used to automatically analyze and solve questions posed by students during learning to provide more timely and personalized learning support.

The deep learning system collects the question text of the student from the forum of the online education platform or the real-time question and answer system, and the text forms the target knowledge operation question text. These question texts typically relate to aspects of course content, homework questions, examination preparation, etc., and these texts often have semantic relevance due to discussions and communications between students.

For each collected question paragraph of each student, the deep learning system uses natural language processing technology to analyze and layer. The system first identifies core questions in the question paragraph and then expands gradually to generate a multi-order question sentence. These multi-level question sentences are arranged in order from concrete to abstract, ensuring that different levels and details of the question can be captured in subsequent processing.

The deep learning system uses a pre-trained model to vectorize each multi-order question sentence, i.e., mine the question understanding characterization vector. These vectors not only contain the literal meaning of the sentence, but also incorporate the context information and semantic relationships of the sentence. By vectorizing the representation, the system can convert complex text information into numerical forms which can be processed by a computer, and provides a basis for subsequent analysis.

In an online education platform, questions of students often change along with the progress of learning, and the characteristics of demand jump are shown. The deep learning system can capture changes and transitions in student demand by comparing and analyzing the question understanding characterization vectors of successive question paragraphs. For example, students may initially ask an explanation of a basic concept, but as understanding proceeds they may present a more advanced or specific problem. The system can recognize such a demand jump and adjust the answer generation strategy accordingly to provide learning support that is more consistent with the current demands of the students.

By the aid of the deep learning system, the questioning text of the students in the online education platform can be effectively processed and analyzed, timely, accurate and personalized learning assistance is provided, and therefore learning experience and effect of the students are improved.

Application scenario 3: customer demand analysis system in enterprise marketing conversion field

In the field of enterprise marketing conversion, deep learning systems are applied to a customer demand analysis system which aims to improve conversion by analyzing the text of online questions of potential customers, understanding their demand changes, and optimizing marketing strategies accordingly.

The deep learning system collects questioning texts of potential clients from official websites, social media platforms, online customer service systems and other channels of enterprises. These texts contain customer inquiries on products, services, prices, promotions, etc., and constitute the targeted knowledge operation questioning text. These texts typically have semantic relevance because they reflect the thinking and decision making processes that customers have in understanding enterprise products and services.

For each collected question paragraph of each client, the deep learning system uses natural language processing technology to analyze and layer. The method comprises the steps of firstly identifying core questions in a question paragraph, then expanding gradually to generate a multi-order question sentence. These multi-level question sentences are arranged in order from concrete to abstract, from detail to whole, in order to more fully capture the customer's questions and needs.

The deep learning system uses a pre-trained model to vectorize each multi-order question sentence, i.e., mine the question understanding characterization vector. These vectors not only contain the literal meaning of the sentence, but also incorporate the context information and semantic relationships of the sentence. By vectorizing the representation, the system can convert complex text information into numerical forms which can be processed by a computer, and a basis is provided for subsequent demand analysis.

In the enterprise marketing transformation scene, the demands of clients often change along with the deep understanding, and the characteristics of demand jump are shown. The deep learning system can capture changes and transitions in customer demand by comparing and analyzing the question understanding token vectors of successive question paragraphs. For example, a customer may initially query the basic functionality of a product, but as the understanding proceeds, they may be concerned with aspects of the product's performance, compatibility, after-market services, and the like. The system can identify such a demand jump and adjust marketing strategies accordingly to provide information and services that better meet the current needs of the customer.

By means of the method, the deep learning system can effectively analyze the demand change of potential clients in the field of enterprise marketing transformation, help enterprises to more accurately locate target client groups, optimize marketing strategies and improve conversion rate and client satisfaction.

Next, the term explanation is made for the technical terms involved in steps 101-104 (taking the intelligent customer service system of a large-scale electronic commerce platform as an example).

Target knowledge operation question text: in the intelligent customer service system of a large-scale electronic commerce platform, the 'target knowledge operation question text' refers to a series of question texts captured from interactions of users with intelligent customer service. These texts are generated by the user during actual shopping or consultation, reflecting the user's needs and questions regarding various aspects of merchandise information, order status, after-sales services, etc. The text data is the basis for the intelligent customer service system to perform deep learning and understand the user demands, and is also the key for optimizing the service quality of the e-commerce platform and improving the user experience.

Several online user question paragraphs with semantic association: in the context of intelligent customer service systems, "several online user question paragraphs with semantic relevance" refers to a set of interrelated question paragraphs identified from the target knowledge operation question text. These paragraphs typically consist of questions of the same user or of different users over a continuous period of time, with some sort of association or logical order semantically between them. For example, a user may first query the size and color of an item, and then further query the item for inventory and shipping time, which may constitute a question paragraph with semantic association.

Each online user asks for a corresponding multi-level question sentence of the paragraph: for each online user question paragraph with semantic association, the intelligent customer service system refines the online user question paragraph into question sentences of multiple layers through natural language processing technology. These multi-level question sentences are arranged in order from concrete to abstract, from detail to whole, in order to more fully capture the user's questions and needs. For example, for a question paragraph in which a user asks for the size of an item of merchandise, a multi-step question sentence may include "what is the specific size of the piece of clothing? "," whether a detailed size table can be referred to? "etc.

Question sentences of each order: the method refers to a question sentence corresponding to each step in the multi-step question sentences. These sentences semantically represent progressive relationships from concrete to abstract, and from detail to whole. Lower order question sentences typically focus on specific detailed information, while higher order question sentences focus more on overall understanding and summarization. This hierarchical question sentence structure helps the intelligent customer service system to understand the user's questions and needs more deeply.

Fine-granularity adjustment of text: the intelligent customer service system adjusts the detail level of the text according to the requirement when the online user question paragraph is processed. Such adjustment may be a finer division of text to extract more specific points of information; text may also be summarized and abstracted to capture higher level topics and concepts. The purpose of the text fine granularity adjustment is to enable the intelligent customer service system to better adapt to the demands and understanding levels of different users and provide more accurate and personalized services.

Text fine granularity is in a decreasing trend: refers to the trend that the text granularity of the question sentences gradually decreases along with the increase of the order in the process of generating the multi-order question sentences. That is, in a question sentence from low order to high order, the description of specific detail information gradually decreases, and the degree of overall generalization and abstraction gradually increases. This trend reflects the understanding of intelligent customer service systems from concrete to abstract, from detail to whole, in handling user questions. This decreasing granularity of text helps the system capture and understand the user's needs at different levels, laying the foundation for providing more comprehensive services.

Problem understanding vector mining: in an intelligent customer service system of a large-scale electronic commerce platform, the 'problem understanding vector mining' refers to a process of performing deep semantic analysis and vectorization representation on a question text of a user by using a deep learning technology. This process aims to extract key information from the user's textual description and convert it into a numerical form, i.e. a vector, that the computer can understand and process. By further analyzing and mining these vectors, the intelligent customer service system can more accurately grasp the actual intent and needs of the user, thereby providing more accurate and personalized services. For example, when the user asks "how does the battery life of the mobile phone? When the system is used, the system can recognize that the user pays attention to the battery performance of the mobile phone through problem understanding vector mining, and further provides relevant detailed information and comparison data.

Problem understanding characterization vector: is a numerical representation generated during the problem understanding vector mining process that reflects the location and direction of user questions in semantic space. Each question understanding token vector is a high abstraction and generalization of a user question, contains key information and semantic features in the question, and can be used for subsequent natural language processing tasks such as similar question matching, answer retrieval and the like. In intelligent customer service systems, a problem understanding characterization vector is the basis for implementing user demand understanding and service response.

Linear vector capacity weight: refers to the importance or contribution of different dimensions or features to the overall vector representation in vector space. In the context of intelligent customer service systems, linear vector capacity weights reflect the importance of different information points in a user's question to overall demand understanding. Through reasonable setting and adjustment of the weights, the system can more accurately capture the core focus of the user, and the accuracy and efficiency of service response are improved. For example, in dealing with merchandise consultation type problems, key information points such as price, quality, after-sales, etc. may be given higher weights to ensure that such information is fully considered in demand understanding and service response.

User demand jump: the method refers to the phenomenon that the user demand changes or shifts in the process of interaction between the user and the intelligent customer service system. Such hopping may be due to a variety of reasons, such as increased knowledge of the goods or services by the user, influence of external environments, changes in personal preferences, etc. In an intelligent customer service system of a large-scale e-commerce platform, user demand jump is a common and complex phenomenon, and the system is required to track and adapt to the change of user demands in real time, so that continuous and consistent service support is provided. For example, a user may initially focus on price factors when browsing goods, but with a deep understanding of the performance of the goods, may gradually shift to focus on quality or after-market services.

Embedding a semantic vector: refers to a technique for mapping words or phrases in text into a low-dimensional vector space that contains semantic information of words or phrases such that semantically similar words are located close together in the vector space. In intelligent customer service systems, embedded semantic vectors are widely used for understanding and presenting user questions. Through converting the user questions into the embedded semantic vectors, the system can more conveniently perform semantic similarity calculation, cluster analysis and other operations, so that more accurate user demand understanding and service response are realized.

And carrying out enterprise user demand jump analysis under the weight of at least one linear vector capacity: by this is meant that the influence of at least one linear vector capacity weight is taken into account when performing the user demand jump analysis. In practical applications, this means that the intelligent customer service system needs to comprehensively consider a plurality of key information points in the user question (the information points correspond to different linear vector capacity weights) and interrelationships and influences among the information points to comprehensively and accurately grasp the change and transfer of the user demand. In this way, the system can understand the shopping decision process and service demand change of the user more deeply, and provide more targeted marketing strategies and service optimization suggestions for the electronic commerce industry.

In step 101, in the operation of a large-scale e-commerce platform, the intelligent customer service system plays a crucial role, which is not only a bridge for communication between a user and the platform, but also a key for improving user experience and promoting transaction conversion. The deep learning system is used as the brain of the intelligent customer service system and is responsible for processing and analyzing massive user question texts so as to accurately understand the user demands and provide personalized services.

In this process, the deep learning system first obtains the target knowledge operation question text. These texts are the actual question records of the user on the e-commerce platform, reflecting various questions and needs of the user during shopping. The questioning texts can come from various links such as commodity detail pages, shopping cart pages, order settlement pages and the like, and cover various aspects such as commodity information, price consultation, sales promotion activities, after-sales services and the like.

In order to acquire the target knowledge operation question text, the deep learning system is connected with a database of the e-commerce platform, and the question record of the user is acquired in real time through a data interface. The records are stored in a database and arranged in a time sequence, so that consistency and time sequence of the questioning text are ensured.

In the process of acquiring the questioning text, the deep learning system can also preprocess the text, including removing irrelevant characters, unifying text formats, identifying and correcting wrongly written characters and the like, so as to ensure the accuracy and normalization of the text. Meanwhile, the system also uses natural language processing technology to process text such as word segmentation and part of speech tagging, and lays a foundation for subsequent semantic analysis and vector mining.

Notably, these target knowledge operation question texts are not isolated, and there is a semantic association between them. Such association may be represented as a logical link between questions of the same user at different points in time, or as similar questions of different users for the same good or service. The deep learning system may utilize such semantic associations to aggregate and group related question text to form a number of online user question paragraphs having semantic associations.

These online user question paragraphs are the basic unit for the deep learning system to follow-up processing and analysis. Through deep mining and understanding of the paragraphs, the system can grasp the requirements and the intentions of the users more comprehensively, and provide more accurate and personalized service support for the e-commerce platform. Meanwhile, the question paragraphs are also important bases for optimizing commodity information display and improving service flows of the e-commerce platform, and are beneficial to improving shopping experience of users and operation efficiency of the platform.

In step 102, the deep learning technique plays a central role in the intelligent customer service system of the large electronic commerce platform, particularly in the process of understanding and parsing the user questions. When users interact with intelligent customer service, their questions often contain multiple levels of information and intent. To more accurately capture this information and intent, the deep learning system may take a series of steps to determine a corresponding multi-level question sentence for each online user question paragraph.

First, the deep learning system obtains online user question paragraphs on the e-commerce platform. These paragraphs are generated by the user during shopping and may contain questions about various aspects of merchandise information, order status, after-market services, etc. After these questioning paragraphs are obtained, the system may perform necessary preprocessing, such as removing irrelevant characters, unifying text formats, etc., to ensure accuracy and normalization of the text.

Next, the deep learning system uses natural language processing techniques to identify semantic associations and classify topics for the pre-processed question paragraphs. By identifying keywords and phrases in paragraphs, the system can determine semantic associations between different paragraphs, as well as the topic categories to which each paragraph belongs. This helps the system more accurately understand the intent and needs of the user.

After determining the semantic association and topic classification, the deep learning system may make text fine-grained adjustments to each online user question paragraph. The purpose of this step is to hierarchically process the user's questions from concrete to abstract, and from details to whole. Through text fine granularity adjustment, the system can generate multiple-order question sentences, wherein each order question sentence corresponds to different levels of information and intent in the original question paragraph.

Specifically, the system will first identify specific details in the question paragraphs, such as specific specifications, colors, etc. of the merchandise, to form a lower-order question sentence. The system then abstracts and summaries this information step by step, forming higher-order question sentences that focus on a wider range of topics or concepts. Thus, the question sentences from low order to high order form a multi-order question sentence structure with the text granularity decreasing.

After the text fine granularity adjustment, the deep learning system can determine corresponding multi-order question sentences of each online user question paragraph. These multi-level question sentences not only contain the original question information of the user, but also reveal different levels of user intent and demand through hierarchical processing. This enables the intelligent customer service system to more fully understand the user's questions and needs, providing powerful support for subsequent responses and services.

In general, the deep learning system achieves deep understanding and parsing of user questions by a series of steps of obtaining and processing online user question paragraphs, identifying semantic associations and topic classifications, performing text fine-grained adjustment, and determining final multi-level question sentences. The method provides a more accurate and personalized service foundation for the e-commerce platform, and is beneficial to improving user experience and platform operation efficiency.

In step 103, in the intelligent customer service system of the large-scale electronic commerce platform, the deep learning technology is a key for realizing accurate user demand understanding. When users interact with intelligent customer service, their questions often contain complex and multi-level semantic information. To accurately capture this information, the deep learning system performs problem understanding vector mining on each online user question paragraph to obtain a problem understanding token vector that reflects the user's needs.

First, the deep learning system obtains online user question paragraphs on the e-commerce platform. These paragraphs are generated by the user during shopping and cover various aspects of merchandise information, order status, after-market services, etc. After these questioning paragraphs are obtained, the system performs necessary preprocessing, such as removing irrelevant characters, unifying text formats, correcting wrongly written characters, etc., to ensure the accuracy and normalization of the text.

Next, the system will make text fine-grained adjustments to each online user question paragraph, generating a multi-order question sentence. These sentences reflect the user's questioning content from concrete to abstract, from detail to whole. Through this step, the system is able to more fully understand the intent and needs of the user.

After generating the multi-order question sentences, the deep learning system performs question understanding vector mining on each multi-order question sentence by using advanced natural language processing technology and deep learning algorithm. The purpose of this step is to convert the text information into a numerical form, i.e. a vector, that the computer can understand and process.

Specifically, the system will first perform semantic analysis on each multi-level question sentence to extract key information and features. These information and features are then learned and represented using a deep learning model (e.g., convolutional neural network, recurrent neural network, etc.) to generate corresponding problem understanding vectors. These vectors capture semantic information and user requirements in the question sentence.

By performing problem understanding vector mining on each multi-order question sentence, the deep learning system can finally obtain a corresponding problem understanding characterization vector for each online user question paragraph. These vectors are deep abstractions and generalizations of the original question paragraphs that contain key information and semantic features in the question while reflecting different levels and aspects of user requirements.

Notably, these problem understanding token vectors are obtained at several different linear vector capacity weights. These weights reflect the importance of different information points to overall demand understanding, helping the system to more accurately grasp the core focus of the user. At the same time, these problem understanding token vectors are also tied to user demand hops. When the user demand changes or shifts, the corresponding problem understanding characterization vector also changes, thereby tracking and adapting to the change of the user demand in real time.

In general, the deep learning system obtains a question understanding characterization vector capable of reflecting the user requirements by performing question understanding vector mining on the online user question paragraphs. The vectors provide a more accurate and personalized service basis for the e-commerce platform, and are beneficial to improving user experience and platform operation efficiency. At the same time, they are also important inputs and bases for subsequent natural language processing tasks (e.g., similar question matching, answer retrieval, etc.).

In step 104, in the operation scenario of the large-scale e-commerce platform, the intelligent customer service system plays a crucial role, which is not only a bridge for communication between enterprises and users, but also a key for improving user experience and promoting transaction conversion. The deep learning system serves as the core of the intelligent customer service system, and processes and analyzes the user question text to accurately understand the user requirements and provide personalized services. When faced with continuous user questions, the deep learning system needs to perform enterprise user demand jump analysis to capture dynamic changes in user demand.

First, the deep learning system obtains two continuous online user question paragraphs on the e-commerce platform. These paragraphs may come from the same user asking questions at different points in time, or from different users asking questions for similar questions. These successive question paragraphs form the basis data for the analysis.

And then, the deep learning system utilizes the trained model to excavate the problem understanding vector of each online user question paragraph so as to obtain a corresponding problem understanding characterization vector. These vectors are deep abstractions and generalizations of the user questions, including key information and semantic features in the questions.

When the enterprise user demand jump analysis is carried out, the deep learning system needs to consider the influence of different linear vector capacity weights on an analysis result. These weights reflect the importance of the different information points to the overall demand understanding. The system can select proper linear vector capacity weight for analysis according to actual conditions and experience.

After determining the problem understanding token vector and the linear vector capacity weight, the deep learning system uses related algorithms and techniques to perform a demand jump analysis on two consecutive online user question paragraphs. The purpose of this step is to capture the user's demand changes at different points in time or on different questions.

Specifically, the system calculates the similarity or variance of the question understanding token vector for the two question paragraphs under the selected linear vector capacity weight. If the similarity is higher, the requirement of the user is not changed obviously; if the difference degree is large, the user needs are indicated to jump.

And finally, the deep learning system outputs a result of the demand jump analysis to the intelligent customer service system. Based on these results, the intelligent customer service system can take corresponding strategies to cope with the change of the user's demands. For example, when a jump in the user demand is detected, the intelligent customer service system may actively guide the user to confirm the new demand or recommend related goods and services to meet the new demand of the user.

In general, the deep learning system can capture dynamic changes of user demands in real time by carrying out enterprise user demand jump analysis, and provides more accurate and personalized service support for an e-commerce platform. This helps to improve the shopping experience of the user and the operating efficiency of the platform.

Therefore, the application realizes natural language processing and semantic understanding by utilizing the deep learning technology, and provides personalized knowledge service by automatically answering the questions presented by the user, thereby realizing intelligent interaction with the user. The method comprises the steps of question understanding, information retrieval, answer generation, dialogue management and the like, knowledge operation efficiency can be comprehensively improved, and user experience and loyalty are enhanced.

Specifically, the beneficial effects of the application are mainly embodied in the following aspects:

Improving accuracy of problem understanding: the deep learning system is used for acquiring the target knowledge operation question text, and the target knowledge operation question text comprises a plurality of online user question paragraphs with semantic association, so that the real intention and the demand of the user can be more comprehensively understood. Meanwhile, text fine granularity adjustment is carried out on each online user question paragraph, a multi-order question sentence is generated, and accuracy and depth of question understanding are further improved;

Realizing personalized knowledge service: the problem understanding characteristic vector is obtained by carrying out problem understanding vector mining on the multi-order question sentences corresponding to each online user question paragraph, and the embedded semantic vector of each user under different linear vector capacity weights can be reflected, so that more personalized knowledge service is provided for the users. The service mode can better meet the demands of users and promote the satisfaction and loyalty of the users;

Capturing user demand jump: the application can also carry out enterprise user demand jump analysis on continuous user question paragraphs, and timely capture dynamic changes of user demands. The method is beneficial to the enterprises to respond to the change of the demands of the users more quickly, adjust the service strategy and improve the service quality and efficiency;

and the knowledge operation efficiency is improved: through the steps of automated question understanding, information retrieval, answer generation, dialogue management and the like, the knowledge operation efficiency can be greatly improved. Enterprises can more efficiently process the problems of users, so that labor cost is saved, and operation efficiency is improved.

In summary, the application realizes natural language processing and semantic understanding by using deep learning technology, and can remarkably improve knowledge operation efficiency, user experience and loyalty of enterprises by providing personalized knowledge service and intelligent interaction. This is significant for enterprises to stay in a leading position in a strong market competition.

In some optional embodiments, the performing the problem understanding vector mining on the multi-order question sentences corresponding to each online user question paragraph to obtain the problem understanding characterization vectors corresponding to each online user question paragraph respectively includes: respectively carrying out problem understanding vector mining on the multi-order question sentences corresponding to each online user question paragraph to obtain a multi-order problem understanding vector relation network corresponding to each online user question paragraph; the linear vector capacity weights and the number of semantic vector clusters corresponding to the problem understanding vector relation network of different feature orders in the multi-order problem understanding vector relation network are different; obtaining question element linear variables corresponding to semantic vector clusters in each order question understanding vector relation network corresponding to each online user question paragraph according to the multi-order question understanding vector relation network corresponding to each online user question paragraph; and determining the question understanding characterization vector corresponding to each online user question paragraph according to the question element linear variable corresponding to each online user question paragraph.

In some alternative embodiments, the deep learning system may perform more in-depth processing on each online user question paragraph to obtain a more accurate question understanding token vector. Specifically, the system may first perform problem understanding vector mining on the corresponding multi-level question sentences of each online user question paragraph. In this step, the deep learning model analyzes the semantics of each question sentence, converts it into a vector form, and captures the key information and meaning in the sentence.

Through the mining process, the system can obtain a corresponding multi-order problem understanding vector relation network of each online user question paragraph. The relation network is a complex structure and comprises problem understanding vectors of different feature orders, and the vector relation network of each feature order has corresponding linear vector capacity weight and semantic vector cluster number. These weights and numbers are determined based on the importance of the vectors and the amount of information, which reflects how much different vectors contribute in understanding the user's questions.

Then, the system further extracts question element linear variables corresponding to each semantic vector cluster in the question understanding vector relation network of each order according to the corresponding multi-order question understanding vector relation network of each online user question paragraph. These question element linear variables are the result of a quantitative representation of key elements in the user question, which contain the core information and requirements of the user question.

Finally, the system determines the question understanding characterization vector corresponding to each online user question paragraph according to the question element linear variable corresponding to each online user question paragraph. The characterization vector is a comprehensive and deep understanding of the user question, integrates information of multiple layers and aspects in the user question, and provides a solid foundation for subsequent demand jump analysis and answer generation.

In this way, the deep learning system can more accurately understand the questions of the user and capture the nuances and changes of the user's needs. This helps the system provide more personalized, more accurate knowledge service, promotes user experience and satisfaction. Meanwhile, the processing method also enhances the flexibility and adaptability of the system, so that the processing method can better cope with the change of requirements of different scenes and users.

In general, the technical scheme of the application carries out multi-level and deep processing and understanding on the online user questions through the deep learning technology, thereby realizing accurate grasp and personalized service of the user demands. This not only improves the efficiency and quality of knowledge operations, but also wins more user loyalty and market competitiveness for the enterprise.

In the next step, according to the multi-order question understanding vector relation network corresponding to each online user question paragraph, obtaining question element linear variables corresponding to semantic vector clusters in each order question understanding vector relation network corresponding to each online user question paragraph, including: determining basic question element linear variables of each question understanding vector unit in each order of the question understanding vector relation network in the multi-order question understanding vector relation network according to the multi-order question understanding vector relation network corresponding to each online user question paragraph; determining a first semantic element linear variable and a second semantic element linear variable of the basic question element linear variable on a target semantic feature coordinate system; the first semantic element linear variable is used for indicating a semantic vector of the problem understanding vector unit under the word level attention index; the second semantic element linear variable is used for indicating a semantic vector of the problem understanding vector unit under the sentence-level attention index; for each semantic vector cluster, determining a first local problem understanding vector of the semantic vector cluster according to a first semantic element linear variable corresponding to each problem understanding vector unit in the semantic vector cluster and the total number of the problem understanding vector units contained in the semantic vector cluster; for each semantic vector cluster, determining a second local problem understanding vector of the semantic vector cluster according to a second semantic element linear variable corresponding to each problem understanding vector unit in the semantic vector cluster and the total number of the problem understanding vector units contained in the semantic vector cluster; and determining question element linear variables corresponding to semantic vector clusters in each order of question understanding vector relation network corresponding to each online user question paragraph according to the first local question understanding vector and the corresponding second local question understanding vector.

In the next step, the deep learning system further goes deep into the details of each online user question paragraph according to the multi-order question understanding vector relation network constructed before, and extracts key information in the question understanding vector relation network, namely question element linear variables corresponding to each semantic vector cluster.

Firstly, the system determines basic question element linear variables of each question understanding vector unit in the multi-order question understanding vector relation network aiming at the multi-order question understanding vector relation network corresponding to each online user question paragraph. These basic question element linear variables are quantized representations of the basic attributes and features of the question understanding vector unit, which form the basis for understanding the user questions.

The system then further determines a first semantic element linear variable and a second semantic element linear variable of the base question element linear variables on the target semantic feature coordinate system. The target semantic feature coordinate system herein is a framework or space for describing and understanding semantic features. The first semantic element linear variable mainly focuses on semantic vectors under the word level attention index, namely the meaning and importance of keywords or phrases in a user question; the second semantic element linear variable focuses on the semantic vector under the sentence-level attention index, namely the meaning and structure of the whole sentence in the user question.

Next, for each semantic vector cluster, the system determines a first local problem understanding vector for the semantic vector cluster based on a corresponding first semantic element linear variable for each of the problem understanding vector units and a total number of problem understanding vector units included in the semantic vector cluster. This first local problem understanding vector is a comprehensive representation of all problem understanding vector units in the semantic vector cluster at word level attention that reflects key information in some aspect of the user question.

Similarly, for each semantic vector cluster, the system determines a second local problem understanding vector for the semantic vector cluster based on a corresponding second semantic element linear variable for each of the problem understanding vector units and the total number of problem understanding vector units included in the semantic vector cluster. This second local problem understanding vector is then a comprehensive representation of all problem understanding vector units in the semantic vector cluster at sentence level attention, supplementing and refining the information of the first local problem understanding vector.

Finally, the system comprehensively determines question element linear variables corresponding to the semantic vector clusters in the respective order question understanding vector relation network of each online user question paragraph according to the first local question understanding vector and the corresponding second local question understanding vector of each semantic vector cluster. These question element linear variables are comprehensive and thorough understanding of each key element in the user question, and they provide important basis for subsequent demand jump analysis and answer generation.

Through the processing flow, the deep learning system can further understand the questions of the user, and capture a plurality of layers and details of the user requirements. The intelligent system not only helps to promote the intelligent level of the system, but also provides more personalized and accurate knowledge service experience for users.

In some preferred embodiments, the question understanding token vector includes word level text semantics for reflecting a word level attention index and sentence level text semantics for reflecting a sentence level attention index; the step of analyzing the enterprise user demand jump of the two online user question paragraphs under the capacity weight of at least one linear vector according to the respective question understanding characterization vectors of the two online user question paragraphs comprises the following steps: determining word level comparison results among the corresponding semantic vector clusters of the two online user question paragraphs under the same linear vector capacity weight according to the word level text semantics of the two online user question paragraphs respectively corresponding to the same linear vector capacity weight; determining sentence level comparison results among the corresponding semantic vector clusters of the two online user question paragraphs under the same linear vector capacity weight according to the sentence level text semantics of the two online user question paragraphs respectively corresponding to the same linear vector capacity weight; and carrying out enterprise user demand jump analysis on the two online user question paragraphs under the weight of at least one linear vector capacity according to the word level comparison result and the sentence level comparison result.

In some preferred embodiments, the deep learning system may focus on word-level text semantics and sentence-level text semantics in the issue understanding token vector particularly when performing enterprise user demand jump analysis. The two text semantics respectively reflect the attention indexes of different layers in the user question, and are important for comprehensively understanding the user requirements.

In particular, word-level text semantics focus primarily on keywords or phrases in user questions, as well as relationships and importance between them. These keywords or phrases are typically direct representations of the user's needs, so word-level text semantics are very sensitive to capturing subtle differences and changes in the user's needs. Sentence-level text semantics focus more on the overall structure and meaning of the user question, including the grammar, semantics, and context information of the sentence. Through sentence-level text semantics, the system can understand the deeper meaning and purpose of a user question.

When the jump analysis of the enterprise user demands is carried out, the system firstly determines word level comparison results among corresponding semantic vector clusters under the same linear vector capacity weight according to the word level text semantics of two online user question paragraphs respectively under the same linear vector capacity weight. This comparison reveals similarities and differences between the two question paragraphs at the keyword or phrase level, as well as possible variations in user demand.

Similarly, the system also determines sentence-level comparison results between corresponding semantic vector clusters under the same linear vector capacity weight according to the sentence-level text semantics of two online user question paragraphs respectively under the same linear vector capacity weight. The comparison result can reveal the similarity and the difference between the two question paragraphs on the overall structure and meaning level, and further supplement and perfect the information of the word level comparison result.

Finally, the system comprehensively considers the word level comparison result and the sentence level comparison result, and performs enterprise user demand jump analysis on two online user question segments under the least linear vector capacity weight. This analysis process will take into account information about multiple levels and aspects of the user's questions, thereby capturing more accurately the changes and jumps in the user's needs. By the processing mode, the system can deeply understand the real demands of the users and provide more accurate and personalized knowledge services for enterprises. Meanwhile, the method is also beneficial to improving the service quality and user satisfaction of enterprises and enhancing the market competitiveness of the enterprises.

Further, the step of performing enterprise user demand jump analysis on the two online user question segments under the least one linear vector capacity weight according to the word level comparison result and the sentence level comparison result includes: taking the maximum linear vector capacity weight in the plurality of different linear vector capacity weights as the current linear vector capacity weight; comparing the question requirement change weight corresponding to the question requirement change information of each corresponding semantic vector cluster under the current linear vector capacity weight with the corresponding question requirement label value; if the comparison analysis result shows that the enterprise user demand jump exists under the current linear vector capacity weight, determining that the enterprise user demand jump exists in the analyzed online user question paragraph; if the comparison and analysis result shows that the enterprise user demand jump does not exist under the current linear vector capacity weight, taking the next linear vector capacity weight in the plurality of different linear vector capacity weights as the current linear vector capacity weight; and jumping to the step of comparing and analyzing the question requirement change weight corresponding to the question requirement change information of each corresponding semantic vector cluster under the current linear vector capacity weight with the corresponding question requirement label value until the jump capturing termination requirement is met.

Further, when the technical scheme is implemented, the deep learning system can conduct enterprise user demand jump analysis according to a series of steps. This process aims to capture the change and jump of user demand by comparing question demand change information under different linear vector capacity weights.

First, the system sets an initial current linear vector capacity weight, and typically selects the largest linear vector capacity weight of a number of different linear vector capacity weights as the starting point. This weight represents the maximum capacity of semantic information in the user question and is an important reference point for analyzing user demand jumps.

Next, the system compares the question requirement change information of each corresponding semantic vector cluster under the current linear vector capacity weight. Specifically, the system calculates a question requirement change weight for each semantic vector cluster and compares the question requirement change weight with a corresponding question requirement tag value. The questioning requirement change weight reflects the requirement change degree of the user on the semantic vector cluster, and the questioning requirement label value is a preset standard value and is used for measuring whether the requirement of the user reaches a jump threshold value.

If the comparison analysis result shows that the enterprise user demand jump exists under the current linear vector capacity weight, that is, the questioning demand change weight of a certain semantic vector cluster or certain semantic vector clusters exceeds the corresponding questioning demand label value, the system can determine that the analyzed online user questioning paragraph has the enterprise user demand jump. This means that the user's needs have changed or shifted significantly, requiring the system to respond and adjust accordingly.

However, if the comparison analysis result shows that there is no jump in the enterprise user demand under the current linear vector capacity weight, i.e. the question demand change weights of all semantic vector clusters do not exceed the corresponding question demand label values, the system will continue to analyze the next linear vector capacity weight. Specifically, the system takes the next linear vector capacity weight from the plurality of different linear vector capacity weights as the new current linear vector capacity weight and jumps back to the step of comparison analysis for recalculation and comparison.

This process continues until the system meets the jump capture termination requirements. The jump capture termination requirement may be that a preset maximum number of iterations is reached, all linear vector capacity weights have been analyzed, or the change rate of the user demand jump is below a certain threshold. Once the termination requirements are met, the system stops analyzing and outputs the final enterprise user demand jump analysis results.

Through the processing mode, the deep learning system can comprehensively and deeply analyze semantic information in the user question and capture the change and jump of the user requirements in different layers and aspects. The method and the system help enterprises to more accurately understand the real demands of users and provide more personalized and accurate knowledge services, so that the user experience and satisfaction are improved.

Under still other exemplary design considerations, if the comparison analysis results characterize that there is an enterprise user demand transition under the current linear vector capacity weight, the method further comprises, prior to the step of determining that the analyzed online user question paragraph has an enterprise user demand transition: under the current linear vector capacity weight, determining a current quantitative statistical result of which the comparison analysis result accords with a set identification condition; if the current quantitative statistical result is greater than or equal to a preset statistical threshold corresponding to the current linear vector capacity weight, determining that the comparison analysis result represents that enterprise user demand jump exists under the current linear vector capacity weight; if the current quantitative statistical result is smaller than a preset statistical threshold corresponding to the current linear vector capacity weight, determining that the comparison analysis result represents that no enterprise user demand jump exists under the current linear vector capacity weight; the preset statistical threshold is related to the number of semantic vector clusters corresponding to the current linear vector capacity weight.

Under still other exemplary design considerations, the deep learning system may introduce additional decision criteria to enhance the accuracy and reliability of the analysis when performing the enterprise user demand jump analysis. The core of the design concept is that before determining whether the user needs to jump, the system firstly carries out quantitative statistics on the comparison analysis result and compares the statistics result with a preset statistics threshold.

Specifically, when the system performs an alignment analysis under a certain current linear vector capacity weight, it calculates a current quantization statistic. The quantitative statistical result reflects the change condition of the semantic vector cluster in the user question under the current weight, and is a quantitative measure for the change of the user demand. This quantization statistic may be an average, maximum, cumulative sum, or other statistic of the varying weights, depending on the system design and analysis goals.

Then, the system acquires a preset statistical threshold corresponding to the current linear vector capacity weight. The preset statistical threshold is a preset threshold for judging whether the user demand change reaches a significant degree. It is noted that this preset statistical threshold is related to the number of semantic vector clusters corresponding to the current linear vector capacity weight. This is because the number of different semantic vector clusters represents the different semantic richness and complexity of the user questions, and therefore the value of the statistical threshold needs to be adjusted according to the actual situation.

The system then compares the current quantized statistics to a preset statistics threshold. If the current quantitative statistical result is greater than or equal to the preset statistical threshold, the change of the user demand exceeds the preset threshold under the current linear vector capacity weight, and the enterprise user demand jump can be considered to exist. At this point, the system performs the corresponding processing steps, such as recording jump information, updating user portraits, etc.

In contrast, if the current quantization statistic result is smaller than the preset statistic threshold, it is indicated that the change of the user demand is not significant under the current linear vector capacity weight, and it can be considered that no jump of the enterprise user demand exists. In this case, the system will continue to select the next linear vector capacity weight as the current weight and perform the next round of comparison analysis according to the procedure described previously.

Through such design thought, the deep learning system can more accurately catch the change and jump of user's demand. The method not only considers semantic information in the user question, but also further verifies and screens the change of the user demand in a quantitative statistics and preset threshold mode. This helps to improve the robustness and accuracy of the system, providing a more reliable user demand analysis result for the enterprise.

In other preferred embodiments, the comparison analysis results include a first comparison analysis view indicating word level comparison results and a second comparison analysis view indicating sentence level comparison results; under the current linear vector capacity weight, determining a current quantitative statistical result of which the comparison analysis result meets a set identification condition, wherein the method comprises the following steps of: determining a first quantitative statistical result of the first comparison analysis viewpoint meeting a set identification condition under the current linear vector capacity weight, and determining a second quantitative statistical result of the second comparison analysis viewpoint meeting the set identification condition; and determining the current quantization statistic according to the first quantization statistic and the second quantization statistic.

In other preferred embodiments, the deep learning system performs more careful processing of analysis results than when performing enterprise user demand transition analysis. Specifically, the comparison analysis result is no longer a single value or conclusion, but includes a first comparison analysis viewpoint for indicating the word level comparison result and a second comparison analysis viewpoint for indicating the sentence level comparison result.

The word level comparison result, i.e., the first comparison analysis viewpoint, mainly focuses on the variation of keywords or phrases in the user questions. The system compares word level text semantics of two online user question paragraphs under the same linear vector capacity weight, and analyzes word level differences and similarities between corresponding semantic vector clusters. Such differences and similarities may be quantitatively represented by calculating an index of cosine similarity, euclidean distance, etc. of the word vector. The first comparative analysis view is a conclusion based on these quantitative indicators to reflect the change in demand of the user questions at the word level.

Sentence level comparison, i.e., the second comparison analysis view, is more focused on the overall structure and meaning of the user question. The system compares sentence-level text semantics of two online user question paragraphs under the same linear vector capacity weight, and analyzes sentence-level differences and similarities between corresponding semantic vector clusters. Such differences and similarities may be represented quantitatively by calculating metrics such as semantic matching of sentence vectors, sentence structure similarity, and the like. The second comparative analysis view is a conclusion based on these quantitative indicators to reflect the change in demand of the user questions at sentence level.

When determining that the comparison analysis result meets the current quantitative statistical result of the set identification condition, the system respectively processes the first comparison analysis view and the second comparison analysis view. Specifically, the system determines a first quantization statistic for which a first comparison analysis view meets a set identification condition under the current linear vector capacity weight, and determines a second quantization statistic for which a second comparison analysis view meets the set identification condition. The set identification condition here may be a preset threshold, rule, model, or the like for judging whether the comparison analysis viewpoint is valid or significant.

The first quantization statistic and the second quantization statistic may be values, statistics, probabilities, etc. of various quantization indexes, depending on the system design and analysis objective. For example, the first quantization statistic may be an average, a maximum, a cumulative sum, or the like of the word level differences; the second quantized statistical result may be a score, ranking or classification result, etc. of sentence-level similarity.

Finally, the system determines the current quantization statistic according to the first quantization statistic and the second quantization statistic. The current quantization statistic is a quantization value or conclusion that integrates word level and sentence level comparisons to reflect the overall degree of change in user demand under the current linear vector capacity weight. The current quantitative statistical result is used as an important basis for judging whether the enterprise user demand jump exists or not.

By the processing mode, the deep learning system can more comprehensively consider information of multiple layers and aspects in the user question, so that changes and jumps of user demands can be more accurately captured. Meanwhile, the flexibility and the expandability of the system are improved, so that the system can adapt to user demand analysis tasks under different scenes and demands.

In an alternative embodiment, the method further comprises: and if the enterprise user demand jump analysis result represents that the analyzed online user question paragraphs have enterprise user demand jumps, carrying out demand jump annotation on the analyzed online user question paragraphs.

In an alternative embodiment, the deep learning system further processes the analysis results after completing the enterprise user demand transition analysis. Specifically, if the analysis result shows that the analyzed online user question paragraph has an enterprise user requirement jump, the system performs an additional step of marking the requirement jump of the question paragraph.

Demand jump annotation is a process of marking or annotating to explicitly indicate the portion of a user question where a demand jump occurred. Such annotations may be in the form of text, symbols, colors, icons, etc., depending on the design of the system and the user's preferences. The labeling aims at facilitating subsequent processing and identification, so that a system or a user can quickly locate and understand the content of the requirement jump.

When the demand jump labeling is executed, the system determines a question paragraph to be labeled according to the previous analysis result of the demand jump of the enterprise user. These paragraphs are the portions of the system that consider there to be a jump in demand, i.e., where the user's question intent or focus has changed significantly.

The specific mode of labeling can be selected according to actual conditions. For example, the system may add a specific label or text next to the question paragraph to indicate that there is a demand jump in the paragraph. Or the system may use different colors or icons to distinguish the marked paragraphs from other parts, making the marking more intuitive and easy to identify.

In addition, the demand hop annotation may also contain some additional information to help the system or user better understand the nature and cause of the demand hop. For example, the callout can contain a hop type, degree of hop, related keywords or phrases, and so forth. The information may be automatically generated by the system or may be manually added by the user.

After the demand jump labeling is completed, the analyzed online user question paragraphs will be marked as containing the content of the demand jump. Thus, in subsequent processing or presentation, the system may focus specifically on these labeled paragraphs to better meet the needs of the user or to provide relevant services.

Through the step of introducing the requirement jump label, the deep learning system can more accurately identify and process the requirement jump condition in the user question. The intelligent level of the system is improved, and more personalized and accurate service experience is provided for the user.

Further, if the analysis result of the enterprise user demand jump characterizes that the analyzed online user question paragraph has enterprise user demand jump, the analysis result of the enterprise user demand jump characterizes that the analyzed online user question paragraph has demand jump, including: if the enterprise user demand jump analysis result represents that the analyzed online user question paragraphs have enterprise user demand jumps, carrying out demand jump association on the analyzed online user question paragraphs; and marking answer retrieval keywords of the analyzed online user question paragraphs related by the demand jump to obtain target online user question paragraphs marked by the demand jump, wherein the question sentence influence weight corresponding to the target online user question paragraphs is smaller than that of the question sentences corresponding to the analyzed online user question paragraphs.

Further, in the process of performing enterprise user demand jump analysis by the deep learning system, if the analysis result shows that the analyzed online user question section has enterprise user demand jump, the system can take a series of steps to carry out special treatment on the question section so as to better understand and meet the demands of users.

First, the system will make a demand jump association. The purpose of this step is to establish a connection between the current question paragraph and the previous or subsequent question paragraphs, thereby identifying the change track of the user's demand. The requirement jump association can be realized by analyzing the time sequence, the topic similarity, the keyword co-occurrence and the like among the question paragraphs. For example, the system may compare keywords of a current question paragraph with keywords of a previous question paragraph, and if a significant change in keywords is found, it may be considered that a demand jump association exists between the two paragraphs.

After the requirement jump association is completed, the system can label answer retrieval keywords for the associated question paragraphs. The purpose of this step is to extract key information in the question paragraph for use in the subsequent answer retrieval process. The answer retrieval keywords may be core words, phrases or concepts in the question paragraph that can accurately reflect the question intent and needs of the user. The system can automatically extract answer retrieval keywords through natural language processing technologies such as word frequency analysis, part-of-speech tagging, named entity recognition and the like. Meanwhile, in order to improve the accuracy and efficiency of labeling, the system can also utilize auxiliary information such as a predefined domain dictionary, a user portrait and the like.

In the process of marking the answer retrieval keywords, the system can adjust the influence weight of the question sentences according to the condition of the jump of the requirements. Specifically, if there is a demand jump, the question sentence impact weight in the associated and annotated target online user question paragraph will be less than the question sentence impact weight in the original analyzed online user question paragraph. This is because the demand jump means that the user's focus or intent has changed, so the new question sentence is of relatively low importance in reflecting the user's demand. By adjusting the impact weight, the system can more accurately evaluate the importance of different question sentences, thereby giving proper attention in the subsequent answer retrieval and recommendation process.

Finally, after the steps of requirement jump association, answer search keyword labeling, influence weight adjustment and the like, the system can obtain a target online user question paragraph marked by the requirement jump. The paragraph not only contains the original questioning information of the user, but also enriches the semantics and the context information of the user in a mode of association, labeling and the like, so that the system can more deeply understand the requirements of the user and provide more accurate services. Meanwhile, the system can also more flexibly cope with the change and jump situation of the user demand by adjusting the influence weight of the questioning sentence.

In some independent embodiments, after performing enterprise user demand jump analysis on any two online user question paragraphs of the plurality of online user question paragraphs under at least one linear vector capacity weight according to the respective question understanding token vector of the two online user question paragraphs, the method further comprises: and generating a question and answer dialogue set based on the question and answer guide graph of the target knowledge operation question text according to the enterprise user demand jump viewpoint obtained by the enterprise user demand jump analysis.

Further, in other independent embodiments, generating a question and answer dialog set based on the question and answer mind map of the target knowledge operation question text according to the enterprise user demand jump viewpoint obtained by the enterprise user demand jump analysis includes:

extracting a key enterprise user demand jump viewpoint according to enterprise user demand jump analysis; the enterprise user demand jump viewpoint reflects the change and transfer of the demands of users among different question paragraphs and is the basis for generating a question thinking guide diagram;

And constructing an initial node of the questioning thinking guide graph based on the extracted enterprise user demand jump point. Each node represents a user demand or point of interest, and the position, size and color attributes of the nodes are set based on the importance and urgency of the demand;

Determining the relation between the thinking guide graph nodes by analyzing the logic relation and the semantic relation between enterprise user demand jump viewpoints; the relationship among the thinking guide graph nodes comprises a hierarchical relationship, a parallel relationship and a causal relationship, and is used for reflecting the inherent relationship and the evolution path among the demands of the users;

generating a visual questioning thinking guide graph by using a graphical technology according to the initial node and node relation; the thinking guide graph is used for showing the overall structure and the change process of the user demand;

Extracting key question-answer pairs from the generated question thought map; each question-answer pair comprises a question and a corresponding answer, wherein the question reflects the specific requirement of a user, and the answer is generated by the system according to the requirement of the user;

Organizing the extracted question-answer pairs into a target question-answer dialog set; the dialogue set contains questions and answers of the system which are presented by the user in different demand stages, and is the comprehensive response of the system to the user demand; the format of the dialogue set is adjustable;

optimizing and updating the questioning thinking figures and the target questioning and answering dialogue sets regularly or in real time; including adding new nodes, adjusting node relationships, and updating question-answer pair operations.

Through the substeps, the system can effectively utilize the result of the enterprise user demand jump analysis to generate the questioning thinking guide graph and the target questioning and answering dialogue set, thereby improving the intelligent level and the user satisfaction of the system.

1. Extracting enterprise user demand jump viewpoint

The system first receives and analyzes a number of online user question paragraphs. These paragraphs may come from customer support forums of businesses, product feedback areas, social media interactions, and so forth. Using natural language processing and machine learning techniques, the system identifies user demand changes between different question paragraphs, which are referred to as "demand hops.

For example, a user may first ask how a certain function of a product is used, then in a subsequent question, move to troubleshooting related to the function, and finally jump to the expectation of a new product function. The system needs to accurately capture these changes in demand and refine them into "enterprise user demand jump views".

2. Initial node for constructing questioning thinking guide graph

Based on the demand jump point extracted in the previous step, the system starts to construct the initial node of the questioning thinking guide. Each node represents a particular user demand or point of interest, such as a certain function of the product, a certain aspect of the service, etc.

The attributes of the nodes (e.g., location, size, and color) are set according to the importance and urgency of the need. For example, for demands that occur frequently or are critical to the user, the system may set its nodes larger, more striking, and placed in the center of the mind map.

3. Determining relationships between mind map nodes

At this step, the system will analyze the logical relationship and semantic links between the demand jump perspectives in depth. This includes identifying which demands are interrelated, which demands are order dependent, which demands are parallel, etc.

For example, if the user first interrogates the basic functions of the product and then interrogates the advanced functions, the system may determine that there is a hierarchical relationship between the two requirements. Also, if the user simultaneously interrogates multiple different functions of the product, the system may determine that there is a parallel relationship between these needs.

4. Generating visual questioning thinking guide graph

Using graphical techniques (e.g., D3.Js, vis. Js, etc.), the system generates a visual questioning and thinking guide based on the initial node-to-node relationship. This mind map not only demonstrates the overall structure of the user's needs, but also reveals the inherent links and evolutionary paths between the needs.

By looking at this mind map, an enterprise can quickly learn about the user's needs changing at different stages and how these needs are interrelated. This helps the enterprise more fully understand the user needs and thereby formulate more efficient product policies and service plans.

5. Extracting key question-answer pairs

The system will extract key question-answer pairs from the generated question-mind map. These question-answer pairs are generated based on answers to questions and systems posed by the user at different stages of demand.

To ensure that the extracted question-answer pairs are representative and accurate, the system may utilize natural language processing techniques to semantically analyze and similarity match the questions and answers. Thus, even if the question expressions are different, the system can classify them as the same question-answer pair as long as the semantics are the same or similar.

6. Organizing a set of target question-answer conversations

Finally, the system organizes the extracted question-answer pairs into a set of target question-answer conversations. The dialog set not only contains questions and answers to the system presented by the user at different stages of demand, but also may contain some additional information such as the time of presentation of the questions, feedback from the user, etc.

The format of the dialog set can be flexibly adjusted as required. For example, it may be a simple text file or a JSON file containing rich metadata. In addition, the system may provide a visual interface to expose this dialog set for ease of viewing and analysis.

7. Optimizing and updating mind map and dialog set

The system needs to optimize and update the questioning mind map and the target question and answer dialog set periodically or in real time as time goes on and the user's needs change. This may include adding new nodes to reflect new user needs, adjusting node relationships to better show the connections between user needs, updating question-answer pairs to ensure that the system's responses remain consistent with the user's current needs at all times, etc.

To achieve this, the system may utilize machine learning and data mining techniques to continuously analyze and learn the historical data. In this way, the system can be continuously self-optimized and perfected to better meet the needs of users.

Fig. 2 is a schematic structural diagram of a deep learning system 200 according to an embodiment of the present application. The deep learning system 200 as shown in fig. 2 includes a processor 210, and the processor 210 may call and run a computer program from a memory to implement the method in an embodiment of the present application.

Optionally, as shown in fig. 2, deep learning system 200 may also include memory 230. Wherein the processor 210 may call and run a computer program from the memory 230 to implement the method in an embodiment of the application.

Wherein the memory 230 may be a separate device from the processor 210 or may be integrated into the processor 210.

Optionally, as shown in fig. 2, the deep learning system 200 may further include a transceiver 220, and the processor 210 may control the transceiver 220 to interact with other devices, and in particular, may transmit information or data to other devices, or receive information or data transmitted by other devices.

Optionally, the deep learning system 200 may implement the storage engine or a component (such as a processing module) in the storage engine or a corresponding flow corresponding to a device in which the storage engine is deployed in each method of the embodiments of the present application, which is not described herein for brevity.

It should be appreciated that the processor of an embodiment of the present application may be an integrated circuit chip having signal processing capabilities. In implementation, the steps of the above method embodiments may be implemented by integrated logic circuits of hardware in a processor or instructions in software form. The Processor may be a general purpose Processor, a digital signal Processor (DIGITAL SIGNAL Processor, DSP), an Application SPECIFIC INTEGRATED Circuit (ASIC), an off-the-shelf programmable gate array (Field Programmable GATE ARRAY, FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware components. The disclosed methods, steps, and logic blocks in the embodiments of the present application may be implemented or performed. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The steps of the method disclosed in connection with the embodiments of the present application may be embodied directly in the execution of a hardware decoding processor, or in the execution of a combination of hardware and software modules in a decoding processor. The software modules may be located in a random access memory, flash memory, read only memory, programmable read only memory, or electrically erasable programmable memory, registers, etc. as well known in the art. The storage medium is located in a memory, and the processor reads the information in the memory and, in combination with its hardware, performs the steps of the above method.

It will be appreciated that the memory in embodiments of the application may be volatile memory or nonvolatile memory, or may include both volatile and nonvolatile memory. The nonvolatile Memory may be a Read-Only Memory (ROM), a Programmable ROM (PROM), an Erasable PROM (EPROM), an Electrically Erasable EPROM (EEPROM), or a flash Memory. The volatile memory may be random access memory (Random Access Memory, RAM) which acts as external cache memory. By way of example, and not limitation, many forms of RAM are available, such as static random access memory (STATIC RAM, SRAM), dynamic random access memory (DYNAMIC RAM, DRAM), synchronous Dynamic Random Access Memory (SDRAM), double data rate Synchronous dynamic random access memory (Double DATA RATE SDRAM, DDR SDRAM), enhanced Synchronous dynamic random access memory (ENHANCED SDRAM, ESDRAM), synchronous link dynamic random access memory (SYNCHLINK DRAM, SLDRAM), and Direct memory bus RAM (DR RAM). It should be noted that the memory of the systems and methods described herein is intended to comprise, without being limited to, these and any other suitable types of memory.

It should be appreciated that the above memory is exemplary but not limiting, and for example, the memory in the embodiments of the present application may also be static random access memory (STATIC RAM, SRAM), dynamic random access memory (DYNAMIC RAM, DRAM), synchronous Dynamic Random Access Memory (SDRAM), double data rate synchronous dynamic random access memory (doubledata RATE SDRAM, DDR SDRAM), enhanced synchronous dynamic random access memory (ENHANCED SDRAM, ESDRAM), synchronous link dynamic random access memory (SYNCH LINK DRAM, SLDRAM), direct Rambus RAM (DR RAM), and the like. That is, the memory in embodiments of the present application is intended to comprise, without being limited to, these and any other suitable types of memory.

On the basis of the above, a computer readable storage medium is provided, on which a computer program is stored, which computer program, when run, implements the method described above.

The foregoing is merely exemplary of the present application and is not intended to limit the present application. Various modifications and variations of the present application will be apparent to those skilled in the art.

Claims

1. A knowledge operation question-answer dialogue processing method based on deep learning, which is characterized by being applied to a deep learning system, the method comprising:

For any two continuous online user question paragraphs in the plurality of online user question paragraphs, carrying out enterprise user demand jump analysis on the two online user question paragraphs under the weight of at least one linear vector capacity according to the question understanding characterization vectors respectively corresponding to the two online user question paragraphs;

The step of carrying out question understanding vector mining on the multi-order question sentences corresponding to each online user question paragraph respectively to obtain question understanding characterization vectors corresponding to each online user question paragraph respectively comprises the following steps: respectively carrying out problem understanding vector mining on the multi-order question sentences corresponding to each online user question paragraph to obtain a multi-order problem understanding vector relation network corresponding to each online user question paragraph; the linear vector capacity weights and the number of semantic vector clusters corresponding to the problem understanding vector relation network of different feature orders in the multi-order problem understanding vector relation network are different; obtaining question element linear variables corresponding to semantic vector clusters in each order question understanding vector relation network corresponding to each online user question paragraph according to the multi-order question understanding vector relation network corresponding to each online user question paragraph; determining the question understanding characterization vector corresponding to each online user question paragraph according to the question element linear variable corresponding to each online user question paragraph;

The obtaining, according to the multi-order question understanding vector relation network corresponding to each online user question paragraph, question element linear variables corresponding to semantic vector clusters in each order question understanding vector relation network corresponding to each online user question paragraph respectively includes: determining basic question element linear variables of each question understanding vector unit in each order of the question understanding vector relation network in the multi-order question understanding vector relation network according to the multi-order question understanding vector relation network corresponding to each online user question paragraph; determining a first semantic element linear variable and a second semantic element linear variable of the basic question element linear variable on a target semantic feature coordinate system; the first semantic element linear variable is used for indicating a semantic vector of the problem understanding vector unit under the word level attention index; the second semantic element linear variable is used for indicating a semantic vector of the problem understanding vector unit under the sentence-level attention index; for each semantic vector cluster, determining a first local problem understanding vector of the semantic vector cluster according to a first semantic element linear variable corresponding to each problem understanding vector unit in the semantic vector cluster and the total number of the problem understanding vector units contained in the semantic vector cluster; for each semantic vector cluster, determining a second local problem understanding vector of the semantic vector cluster according to a second semantic element linear variable corresponding to each problem understanding vector unit in the semantic vector cluster and the total number of the problem understanding vector units contained in the semantic vector cluster; determining question element linear variables corresponding to semantic vector clusters in each order of question understanding vector relation network corresponding to each online user question paragraph according to the first local question understanding vector and the corresponding second local question understanding vector;

The question understanding characterization vector comprises word level text semantics for reflecting word level attention indicators and sentence level text semantics for reflecting sentence level attention indicators; the step of analyzing the enterprise user demand jump of the two online user question paragraphs under the capacity weight of at least one linear vector according to the respective question understanding characterization vectors of the two online user question paragraphs comprises the following steps: determining word level comparison results among the corresponding semantic vector clusters of the two online user question paragraphs under the same linear vector capacity weight according to the word level text semantics of the two online user question paragraphs respectively corresponding to the same linear vector capacity weight; determining sentence level comparison results among the corresponding semantic vector clusters of the two online user question paragraphs under the same linear vector capacity weight according to the sentence level text semantics of the two online user question paragraphs respectively corresponding to the same linear vector capacity weight; carrying out enterprise user demand jump analysis on the two online user question paragraphs under the weight of at least one linear vector capacity according to the word level comparison result and the sentence level comparison result;

And performing enterprise user demand jump analysis on the two online user question segments under the least one linear vector capacity weight according to the word level comparison result and the sentence level comparison result, wherein the analysis comprises the following steps: taking the maximum linear vector capacity weight in the plurality of different linear vector capacity weights as the current linear vector capacity weight; comparing the question requirement change weight corresponding to the question requirement change information of each corresponding semantic vector cluster under the current linear vector capacity weight with the corresponding question requirement label value; if the comparison analysis result shows that the enterprise user demand jump exists under the current linear vector capacity weight, determining that the enterprise user demand jump exists in the analyzed online user question paragraph; if the comparison and analysis result shows that the enterprise user demand jump does not exist under the current linear vector capacity weight, taking the next linear vector capacity weight in the plurality of different linear vector capacity weights as the current linear vector capacity weight; jumping to the step of comparing and analyzing the question requirement change weight corresponding to the question requirement change information of each corresponding semantic vector cluster under the current linear vector capacity weight with the corresponding question requirement label value until the jump capturing termination requirement is met;

If the comparison analysis result indicates that the enterprise user demand jump exists under the current linear vector capacity weight, before the step of determining that the enterprise user demand jump exists in the analyzed online user question paragraph, the method further comprises the following steps: under the current linear vector capacity weight, determining a current quantitative statistical result of which the comparison analysis result accords with a set identification condition; if the current quantitative statistical result is greater than or equal to a preset statistical threshold corresponding to the current linear vector capacity weight, determining that the comparison analysis result represents that enterprise user demand jump exists under the current linear vector capacity weight; if the current quantitative statistical result is smaller than a preset statistical threshold corresponding to the current linear vector capacity weight, determining that the comparison analysis result represents that no enterprise user demand jump exists under the current linear vector capacity weight; the preset statistical threshold is related to the number of semantic vector clusters corresponding to the current linear vector capacity weight.

2. The method of claim 1, wherein the comparison analysis results include a first comparison analysis view indicating word level comparison results and a second comparison analysis view indicating sentence level comparison results;

3. The method according to claim 1, wherein the method further comprises:

4. The method of claim 3, wherein if the analysis result of the enterprise user demand jump characterizes that the analyzed online user question paragraph has enterprise user demand jump, the analyzing online user question paragraph is marked with the demand jump, comprising:

5. A deep learning system comprising at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executing computer-executable instructions stored in the memory causes the at least one processor to perform the method of any one of claims 1-4.