TWI803093B - Semantic understanding system for rich-text, method and computer readable medium thereof - Google Patents

Abstract

The present invention is a semantic understanding system for rich-text and a method thereof. Multiple named entity recognition methods are integrated to find vocabularies of potential named entities by a potential named entity recognition module. Non-overlapping vocabularies are combined to generate similar sentences according to the system support sentence pattern and the order of appearance of the potential named entities in the sentence by a similar sentence generation module. The closest sentence from a number of similar sentences is selected to estimate the sentence intention and select a combination of named entities by a sentence intention selection module. Therefore, the recognition problem caused by missing or typo in the sentence can be avoided in the present invention, so it can improve the quality and precision of voice control service. The present invention also provides a computer-readable medium for executing the method of the present invention.

Description

Semantic understanding system, method and computer-readable medium for processing rich text

The invention relates to the technology of voice-controlled semantic recognition, especially a semantic understanding system and method for processing rich text and its computer-readable medium.

Voice control services are widely used in scenarios such as home appliance control, streaming media playback, and event reservations, and also provide people with convenient life applications. With the emergence of more and more voice-activated services, how to accurately identify the user's semantics has become more important. For example, users often use voice-activated services without thinking, and the resulting sentences are generally short, sometimes lacking subjects, verbs, or objects, causing system Recognition error, or the user can’t remember the program name that is too long, and the resulting sentence misses some words. For example, the program title of "Classic 101 World Literature Classics for Children" only narrates the segment "Say it for children Literary classics" will also cause the system to be unable to accurately identify, or there are homonyms in the program names, such as "stock cancer", "talking about money", "dujiaoxi" and other speech recognition errors. Missing characters or typos will pose more challenges to the quality of natural language processing of voice-activated services.

In other words, it is known that natural language processing systems such as Dialogflow continue to use fuzzy matching technology. When there are missing characters in the program name in the sentence, for example, the sentence "I want to listen to career", the program name is missing "99" , Dialogflow may mistake the verb "listen" as part of the program title, and the result Mistakenly touching the program "Job Enjoy Listening to You". According to statistics, about 10% of the program's lack of characters will lead to misplaying; also, when the program name has homophony in the sentence, such as the sentence "I want to listen to Emily's explosive report", the homophone The program "Emily Newspaper" could not be played due to the uncorrected "Burst", according to the statistics, about 10% of the similar situations; in addition, regardless of missing characters or typos, there are still 10% of the programs that may not be recognized, the system needs to ask the user's intention again, For example, the sentence "I want to listen to the latest episode of the major league" lacks the word "Hito" or the sentence "I want to listen to Sanjinxiu" has a homophone "jin". It can be seen from the above that the commonly used natural language processing system has an error rate of about 20% when the program name is missing or homonymous, so there is a need for improvement.

In view of the above problems, how to improve the recognition rate of voice control, especially when the user uses the voice control service, the sentence is too simple, the program name is too long and some words are missing, and the program name is homophonic due to typos and mistakes, and the sentence intention can still be accurately estimated , and then provide users with correct voice-activated services, which will become the goal that those skilled in the art are eager to pursue.

In order to solve the above-mentioned problems in the prior art, the present invention discloses a semantic understanding system for processing rich text, which includes: a latent named entity recognition module, which has multiple named entity recognizers of different types and receives text converted from speech Each named entity recognizer is used to calculate and score the longest common subsymbol sequence between the text sentence and the pre-stored vocabulary, so as to select the vocabulary with the highest score and meet the first threshold as a potential named entity; generate similar sentences The module is used to combine the vocabulary corresponding to the potential named entity without repetition to generate similar sentences and score them, so that the similar sentences with the highest score and meeting the second threshold are selected as similar sentences; And the sentence intent selection module is used to provide parameters for triggering corresponding services, generate corresponding reply messages or generate interactive messages according to the number of similar sentences.

The present invention further discloses a semantic understanding method for processing rich text, which is executed by a computer device. The method includes the following steps: making the latent named entity recognition module receive text sentences converted from speech, so as to use different types of named entity recognition The device calculates the longest common subsymbol sequence between the text sentence and the pre-stored vocabulary for scoring, so as to select the vocabulary with the highest score and meeting the first threshold as a potential named entity; let the similar sentence generation module compare the potential named entity The corresponding vocabulary is combined without repetition to generate similar sentences and score them, so that the similar sentences with the highest score and meeting the second threshold are selected as similar sentences; and the sentence intent selection module is based on the similar sentences The number of , providing parameters for triggering the corresponding service, generating a corresponding reply message or generating an interactive message.

In the aforementioned system and method, when the named entity recognizer is a character named entity recognizer, it uses a Chinese character or an English word as a symbol unit to calculate the matching rate between the symbol of the text sentence and the symbol of the vocabulary, and Obtain the longest common subsymbol sequence.

In the aforementioned system and method, when the named entity recognizer is a musical note named entity recognizer, it uses a phonetic symbol of a Chinese character without tones or an English word stem as a symbol unit, and the symbol of the text statement is related to the The symbols of the vocabulary are calculated to obtain the longest common sub-symbol sequence.

In the foregoing system and method, the latent named entity recognition module finds the longest common subsymbol sequence between the text sentence and the vocabulary in a sliding window manner, and the scoring method is to compare the symbols of the text sentence with the For the symbol of the vocabulary, points are added when the symbol is correct and points are deducted when the symbol is wrong, so that the word with the highest score and meeting the first threshold is used as the potential named entity.

In another embodiment, after the potential named entity recognition module finds out the potential named entity, it uses masking the common subsequence in the text sentence, and then uses the same procedure to find other potential named entities in the text sentence until Until no other potential named entities are found.

In the aforementioned system and method, the similar sentence generating module is based on the sentence patterns supported by the system and the order of appearance of the potential named entity, combining non-overlapping words to generate the similar sentence, and then calculates each similar sentence and the text The longest common sub-symbol sequence of the sentence is scored after comparing the symbols of the literal sentence with the symbols of the similar sentence, so as to add points when the symbols are correct and deduct points when the symbols are wrong, so that the score is the highest and meets the second A similar statement of the threshold value is used as the approximate statement.

In the aforementioned systems and methods, when there is only one approximate sentence, the sentence intent selection module generates parameters to trigger the corresponding service through the parameters of the backend repeatedly passed to the module, or when the approximate sentence is zero or When multiple, the reply message or the interaction message is generated through the reply statement generating module at the back end.

In the aforementioned system and method, the semantic understanding method for processing rich text further includes a voice-activated sentence pattern construction module, which is used to list important Vocabulary combinations and ignoring unimportant words to construct the pre-stored vocabulary.

The present invention further discloses a computer-readable medium, which is used in a computing device or a computer, and stores instructions to execute the aforementioned semantic understanding method for processing rich text.

It can be seen from the above that the semantic understanding system and method for processing rich texts of the present invention integrates the recognition results of several named entity (NE) to exert synergistic effects to cover various situations and effectively reduce the error rate of named entities. Specifically, the present invention constructs sentence patterns supported by voice-activated services in natural language, formulates a set of potential named entity primary selection mechanism to recommend various potential named entities, and then combines potential named entities according to sentence patterns to generate similar sentences. Finally, through approximation Sentence selection formulas to establish sentence intent and select command name entity combination. Compared with the traditional model, which first recognizes the intent of the sentence and then fuzzily compares the potential named entities, the present invention can avoid the preconception of a single model, or the fuzzy comparison of the range of frame errors, so it can effectively improve the recognition rate of named entities for sentences with missing characters and typos.

1: Semantic understanding system for processing rich text

11: Potential Named Entity Recognition Module

111: Character Named Entity Recognizer 111

112:Note Named Entity Recognizer

12: Similar sentence generation module

13: Sentence intent selection module

14: Voice-activated sentence construction module

2: Speech-to-text module

3: Parameters are repeatedly handed over to the module

4: Answer sentence generation module

5: Service API

6: Text-to-speech module

301-305: Process

401-406: Process

S501-S503: Steps

FIG. 1 is a schematic structural diagram of a semantic understanding system for processing rich text according to the present invention.

FIG. 2 is a schematic structural diagram of an application embodiment of the semantic understanding system for processing rich text of the present invention.

FIG. 3 is a flow chart of the primary selection of potential named entities by the potential named entity recognition module in the present invention.

FIG. 4 is a flow chart of generating similar sentences by the similar sentence generating module and finding similar sentences by the sentence intent selection module in the present invention.

FIG. 5 is a step diagram of the semantic understanding method for processing rich text according to the present invention.

The following describes the technical content of the present invention through specific embodiments, and those skilled in the art can easily understand the advantages and effects of the present invention from the content disclosed in this specification. However, the present invention can also be implemented or applied in other different specific implementation forms.

FIG. 1 is a schematic structural diagram of a semantic understanding system for processing rich text according to the present invention. As shown in the figure, the semantic understanding system 1 for processing rich text includes at least a latent named entity recognition module 11 , a similar sentence generation module 12 and a sentence intent selection module 13 .

The potential named entity recognition module 11 has a plurality of different types of named entity recognizers and receives converted text sentences from the speech input by the user, and each named entity recognizer is used to calculate the longest length between the text sentences and the pre-stored vocabulary. The sequence of common sub-symbols is scored and the vocabulary with the highest score and meeting the predetermined first threshold is selected as a potential named entity. In other words, the potential named entity recognition module 11 integrates several named entity (NE) recognition methods to initially select potential NEs, wherein the potential named entity recognition module 11 is preset with a plurality of different types of named entity recognition device, when the input speech input by the user is converted into a text sentence, it will be sent to the potential named entity recognition module 11 for analysis, and each named entity recognizer is used to calculate the difference between the received text sentence and the pre-stored vocabulary The longest common sub-symbol sequence is then compared with the longest common sub-symbol sequence for scoring, and the vocabulary with the highest score and meeting the predetermined first threshold is selected as a potential named entity.

In one embodiment, the named entity recognizer includes a character named entity recognizer and a musical note named entity recognizer, wherein the character named entity recognizer uses a Chinese character or an English word as a symbol unit, and the symbol of the text sentence Calculate the matching rate with the symbols of the vocabulary to obtain the longest common subsymbol sequence, and the phonetic note named entity recognizer uses a Chinese character phonetic symbol without tone or an English word stem as the symbol unit, and the character The matching rate calculation is performed between the symbols of the sentence and the symbols of the vocabulary to obtain the longest common sub-symbol sequence. In other words, the character NE recognition method and the musical note NE recognition method performed by the character named entity (NE) recognizer and the musical note named entity (NE) recognizer are respectively used for recognition in response to vocabulary missing characters and homophonic typos. The symbol unit of the NE recognition method is a Chinese character or an English word, and the symbol unit of the phonetic NE recognition method is a phonetic or pinyin symbol without tone, and the English word is a word stem. Therefore, after the text sentences and pre-stored vocabulary are converted into symbols, the primary selection of potential NEs can be carried out under the same potential NE selection framework.

In one embodiment, the latent named entity recognition module 11 finds the longest common subsymbol sequence between the text sentence and the vocabulary in a sliding window manner, and the scoring method is to compare the symbols of the text sentence and the vocabulary Symbols, adding points (such as one point) when the symbols are correct and deducting points (such as one point) when the symbols are wrong, so that the vocabulary with the highest score and meeting the predetermined first threshold is used as the potential named entity. In short, according to the symbol matching rate between the text sentence and the pre-stored vocabulary, K is the better (Top K) vocabulary is selected. The value of K can be set according to actual needs. Then, for each vocabulary, find out the The longest common sub-symbol sequence of the pre-stored vocabulary and text sentences, the length of the longest common sub-symbol sequence is less than or equal to the length of the vocabulary, and finally, the vocabulary with the highest score and greater than or equal to the predetermined first threshold value is selected from the Top K vocabulary as a potential naming The entity, wherein the predetermined threshold can be adjusted according to the scoring result of the longest common sub-symbol sequence to improve performance.

In addition, the scoring method of the longest common sub-symbol sequence is, for example, the scoring method of Longest Common Subsequence (LCS), including bonus points for correct symbols (such as one point), and incorrect symbols, such as more/less/replacement of one symbol, then deducting points (such as one points), the LCS score not only depends on the edit distance, but also refers to the correct extra points, and gives priority to picking out longer words to avoid picking out the best words locally.

In another embodiment, after finding out the potential named entity, the potential named entity recognition module 11 uses the masking of the common subsequence in the text sentence, and then uses the same procedure to find other potential named entities in the text sentence until Until no other potential named entities are found. In short, after a potential named entity is found, if there are other potential named entities, the next potential named entity search can be performed after masking the common subsequence in the text sentence, for example, the symbol ★ can mask the text sentence In order to search for other potential named entities again, if there is no potential named entity, the exploration results of named entities can be aggregated.

Therefore, the potential named entity recognition module 11 integrates several named entity (NE) recognition methods to initially select potential NEs. Among them, the NE recognition method includes character and musical note NE recognition methods to deal with missing words and typos in words and sentences. It should be noted that the latent named entity recognition module 11 can also integrate conventional NE recognition methods to cope with other situations, and is not limited to the types of recognizers mentioned above. In addition, the latent named entity recognition module 11 collects all kinds of NE preliminary selection results, including NE vocabulary, start and end positions, for the similar sentence generation module 12 to construct similar sentences.

The similar sentence generation module 12 is used for non-repetitive combination of vocabulary corresponding to the potential named entity to generate similar sentences and score them, so that the similar sentence with the highest score and meeting the second threshold value is selected as an approximate statement. In short, the purpose of the similar sentence generation module 12 is to combine the vocabulary of potential named entities to generate similar sentences. Among them, non-overlapping words can be combined according to the sentence patterns predetermined by the system and the order of appearance of potential named entities in text sentences to generate similar sentences. statement, after that, calculate and score the longest common subsymbol sequence between each similar sentence and the original text sentence, and select the similar sentence with the highest score and meet the second threshold value as an approximate sentence, where the second threshold value is also The scoring result of the longest common sub-symbol sequence can be adjusted to improve the performance.

In one embodiment, the similar sentence generation module 12 is based on the sentence patterns supported by the system and the order of appearance of the potential named entity, combining non-overlapping words to generate the similar sentence, and calculates the difference between each similar sentence and the text sentence The longest common sub-symbol sequence, compare the symbols of the text sentence with the symbols of the similar sentence and score, so as to add points (such as one point) when the symbols are correct and deduct points (such as one point) when the symbols are wrong, so as to maximize the score And a similar sentence that meets the second threshold is used as the approximate sentence.

The sentence intent selection module 13 is used for providing parameters for triggering corresponding services, generating corresponding reply messages or generating interactive messages according to the quantity of the approximate sentences. In short, the purpose of the sentence intent selection module 13 is to select similar sentences generated by the similar sentence generation module 12, and estimate the sentence intent, Finally, the named entity combination is selected. In other words, when the number of similar sentences is zero, the system will reply that there is no corresponding service. When the number of similar sentences is unique, the intent of the similar sentence and the named entities that make up the similar sentence will be used as the path and delivery of the trigger service. Parameters, when the number of similar sentences is more than one, the named entity of the similar sentence is formed, and the option template is inserted to obtain several question options, and multiple-choice questions are formed to ask the user's intention.

In addition, the semantic understanding system 1 for processing rich text further includes a voice-activated sentence construction module 14, which is used to list important vocabulary combinations and ignore unimportant words in advance from multiple training sentences, so as to construct the pre-stored vocabulary . In short, a sentence pattern is composed of more than one vocabulary, and the order of the words does not affect the uniqueness of the sentence pattern. When constructing a voice-controlled sentence pattern in natural language, the present invention can use the voice-controlled sentence pattern construction module 14 to sort out the training sentences Important words are selected and unimportant words are ignored, and finally become the words stored in the system for subsequent analysis and comparison by the latent named entity recognition module 11 .

FIG. 2 is a schematic structural diagram of an application embodiment of the semantic understanding system for processing rich text of the present invention. As shown in the figure, the latent named entity recognition module 11, the similar sentence generation module 12 and the sentence intent selection module 13 of the semantic understanding system 1 for processing rich text are the same as those in FIG. Text-rich semantic understanding system 1 connected speech-to-text module 2, parameter repeating module 3, answer sentence generation module 4, service API 5, and text-to-speech module 6.

Semantic understanding system 1 for processing rich text is mainly used to process rich text and lighten it for semantic understanding, which includes a potential named entity recognition module 11 used to integrate several named entity recognition methods for primary selection of potential named entities. A similar sentence generation module 12 for combining the vocabulary of potential named entities to generate similar sentences and a sentence intent evaluation module 13 for selecting similar sentences. In actual operation, the general voice-activated service generates text sentences through the external speech-to-text module 2 as the input of the semantic understanding system 1 for processing rich text, and the latent named entity recognition module 11 distributes text sentences to the character named entity recognizer 111 , Musical Note Named Entity Recognizer 112 or other NE recognition modules, collect various recognition results For the similar sentence generation module 12, the similar sentence generation module 12 combines potential named entities according to the sentence pattern, generates similar sentences for the sentence intent selection module 13 to select similar sentences, and selects the named entity combination after estimating the sentence intent.

Sentence intent selection module 13 generates parameters to trigger the corresponding service through the parameter repeating module 3 of the backend when there is only one similar sentence, or when there are zero or more similar sentences, passes the parameter of the backend The reply sentence generation module 4 generates the reply message or the interaction message. Specifically, after the semantic understanding system 1 for processing rich text selects the intent and combines it with NE, it can be used by the external parameter repeating module 3 to generate parameters to trigger the back-end service, that is, the service application programming interface (API) 5 executes the corresponding service, Alternatively, the reply sentence generation module 4 combines the selected NEs to generate a reply sentence, and then the text-to-speech module 6 emits a voice to interact with the user.

It should be noted that in this embodiment, modules such as speech-to-text module 2, parameter repeating module 3, reply sentence generation module 4, and text-to-speech module 6 are used in the semantic understanding system 1 for processing rich text. Externally, that is, the above-mentioned modules are installed in different devices, servers, and systems according to different purposes. After the semantic understanding system 1 that processes rich text has parsed the text and deduced the intention of the text, it is handed over to the Corresponding system for follow-up processing. However, in other application embodiments, the above-mentioned modules can also be integrated into the semantic understanding system 1 for processing rich text, and become an integrated system service from voice conversion, language analysis, and finally feedback to customers.

Before the operation of the semantic understanding system 1 for processing rich text, it includes the construction of pre-stored vocabulary in advance, and the vocabulary construction can be carried out by the voice-controlled sentence structure module (as shown in Figure 1) in the semantic understanding system 1 for processing rich text, because Voice control sentences are composed of several words. The order of words usually does not affect the semantics. For example, "play the story of the three little pigs" has the same meaning as "play the story of the three little pigs". Voice control sentences are always simple, and users even omit verbs Call the program name directly, such as the three little pigs, so the voice control sentence pattern The construction module only needs to list important vocabulary combinations, ignore unimportant words, and finally form a pre-stored vocabulary for comparison by the latent named entity recognition module 11 when searching for potential named entities. As shown in Table 1 below, it is explained that information such as supporting sentence patterns, intentions, and corresponding parameters can be deduced through the combination of important words.

Follow-up will further explain the process of performing the primary selection of potential named entities by the potential named entity recognition module, generating similar sentences by the similar sentence generation module, and finding similar sentences by the sentence intent selection module.

FIG. 3 is a flow chart of the primary selection of potential named entities by the potential named entity recognition module in the present invention, illustrating the process of the primary selection of potential NEs by the potential named entity recognition module integrating several named entity (NE) recognition methods.

In the process 301, Top K words are recommended according to the symbol matching rate between the text sentence and the vocabulary. This process is to carry out the symbol matching rate between the converted text sentences of the input voice and the pre-stored vocabulary, wherein, the symbol matching rate formula is as follows:

Among them, for the character named entity recognizer, the symbol unit is a Chinese character or an English word, and for the musical note named entity recognizer, the symbol unit is a Chinese character. Tones or an English word stem, whether it is a character named entity recognizer or a musical note named entity recognizer, the matching rate calculation formula is different from the conventional BM25 (Best Match25) calculation formula, and the frequency of reverse files is eliminated to avoid common words in the program name. After being sorted, a smaller range K can be used to calculate the longest common subsymbol sequence of sentences and vocabulary.

In the process 302, for each vocabulary, the longest common subsymbol sequence (LCS) between the vocabulary and the text sentence is found in a sliding window manner. Specifically, the length of the longest common sub-symbol sequence≦vocabulary length, wherein, the scoring method of the longest common sub-symbol sequence can be an extra point if the symbol is correct (such as one point), and a wrong symbol (more/less/replacement of one symbol) Points are deducted (such as one point), because LCS scoring not only depends on the edit distance, but also refers to the correct extra points, so longer vocabulary will be selected first, which can avoid taking out locally optimal vocabulary.

In the process 303, the vocabulary with the highest score and ≧α is selected from the K vocabulary as a potential NE. This process is to select the vocabulary with the highest score and meet the preset condition α as a potential named entity after lexical scoring, where α can adjust the threshold according to the scoring result of the longest common subsymbol sequence to improve the effect.

Next, it is judged whether there is a candidate named entity (NE), if there is a candidate NE, go to process 304 , if there is no candidate NE, go to process 305 .

In the process 304, symbols are used to mask the common subsequence (LCS) in the text sentence, and then other potential NEs are searched. When there are other candidate NEs, symbols (such as ★) can be used to cover the common subsequences in the text sentence, and then, other potential NEs in the text sentence can be found, that is, after using symbols to cover the common subsequences in the text sentence, and then return to The process 301 repeats the preceding process.

In the process 305, the NE exploration results are compiled. When there are no other candidate NEs, the NE exploration results can be compiled, which includes the start and end positions of the words in the sentence.

It should be noted that, in addition to the character NE recognition method carried out by the above-mentioned character named entity (NE) recognizer and the musical note NE recognition method carried out by the musical note named entity (NE) recognizer, it is also possible to expand the conventional part-of-speech sentence pattern (part- of-speech Pattern, POS Pattern), Markov model (Hidden Markov Model, HMM), conditional random field (Conditional Random Field, CRF), recurrent neural network (Recurrent neural network, RNN) or convolutional neural network ( Convolutional Neural Network, CNN) and other NE recognition methods to latent named entity recognition modules to cover different types of text sentences.

Fig. 4 is a flow chart of generating similar sentences by the similar sentence generation module and finding similar sentences by the sentence intent evaluation module in the present invention, which illustrates the process of the similar sentence generation module combining potential NE vocabulary to generate similar sentences, and the sentence intent evaluation The process of selecting similar sentences by the module.

In the process 401, according to the sequence of appearance of the system supported sentence pattern and the potential NE of the text sentence, non-overlapping words are combined to generate a similar sentence. This process is that the similar sentence generation module combines the non-overlapping words according to the sentence pattern supported by the system and the sequence of potential NE appearance in the text sentence to generate similar sentences.

In the process 402, the longest common subsymbol sequence between each similar sentence and the original sentence is calculated. This process is to score each similar sentence, that is, to compare the symbols with the original text sentence. The LCS scoring method adopted is as mentioned above. If the symbol is correct, it will add points (such as one point), and more/less/ Substitution of a symbol and other symbol errors will result in a deduction of points (such as one point).

In the process 403, the similar sentence with the highest score and ≧β is selected as an approximate sentence. This process is to select after scoring each similar sentence in the process 402, select the similar sentence with the highest score and meet the preset condition β as an approximate sentence, where β can be adjusted according to the scoring result of the longest common subsymbol sequence to improve the effect.

Afterwards, there will be different results according to the number of approximate sentences. When the number of approximate sentences is 1, go to process 404. When the number of approximate sentences is 0, go to process 405. When the number of approximate sentences is multiple (≧2), Go to process 406 .

In the process 404, the intent of the approximate statement and the NEs that compose the approximate statement are used as parameters for triggering the path and delivery of the service. This process means that when there is only one similar sentence, the sentence intent selection module generates the path and delivery parameters for triggering the service according to the intent of the similar sentence and the named entities that make up the similar sentence. In addition, the named entities that make up the approximate sentence can be inserted into the reply sentence template to obtain the introduction sentence at the beginning.

In the process 405, the user is answered that the sentence cannot be understood. This process means that when the similar sentence does not exist, the sentence intent selection module generates corresponding feedback to answer that the user does not understand the sentence.

In the process 406, the NE forming the approximate sentence is inserted into the option template to obtain several question options, and a multiple-choice question is formed to ask the user's intention. This process means that when there is more than one similar sentence (greater than or equal to 2), the sentence intention selection module will put the named entities that make up the similar sentence into the option template, and then obtain several question options to form multiple choice questions, so as to ask the user's intention.

FIG. 5 is a step diagram of the semantic understanding method for processing rich text according to the present invention. The semantic understanding method for processing rich text of the present invention can be executed on computer equipment such as personal computers, servers or cloud devices, wherein the semantic understanding method for processing rich text of the present invention includes the following steps.

In step S501, the potential named entity recognition module receives the converted text sentence from the user's input voice, and uses different types of named entity recognizers to calculate the longest common subsymbol sequence between the text sentence and the pre-stored vocabulary Points, to select the vocabulary with the highest score and meeting the predetermined first threshold as a potential named entity. In this step, the potential named entity recognition module receives the converted text sentence from the user's input voice, and uses different types of named entity recognizers to analyze the text The longest common sub-symbol sequence of both the sentence and the pre-stored vocabulary is then scored for the symbol position, and the vocabulary with the highest score and meeting the predetermined first threshold is selected as a potential named entity.

In one embodiment, the named entity recognizer may be a character named entity recognizer or a musical note named entity recognizer. The character named entity recognizer uses a Chinese character or an English single character as a symbol unit, and calculates the matching rate between the symbol of the text sentence and the symbol of the vocabulary to obtain the longest common subsymbol sequence, while the musical note named entity recognizer uses A phonetic symbol of a Chinese character does not contain tones or an English single word stem as a symbol unit, so as to calculate the matching rate between the symbol of the word sentence and the symbol of the vocabulary, and obtain the longest common sub-symbol sequence. In other words, by transforming text sentences and pre-stored vocabulary into symbols, the primary selection of potential NEs can be performed under the same potential NE selection framework.

In another embodiment, the latent named entity recognition module uses a sliding window to find the longest common subsymbol sequence between the text sentence and the vocabulary, and by comparing the symbols of the text sentence with the symbols of the vocabulary, in Add points (such as one point) when the symbol is correct and deduct points (such as one point) when the symbol is wrong, and finally select the word with the highest score and meet the predetermined first threshold as the potential named entity.

In addition, after the potential named entity recognition module finds out the potential named entity, if there are other candidate NEs, it uses the method of masking the common subsequence in the text sentence, and then uses the same procedure to find other potential names in the text sentence entities until no other potential named entities are found.

In step S502, the similar sentence generation module combines the vocabulary corresponding to the potential named entity without repetition to generate similar sentences and perform scoring, so that the similar sentence with the highest score and meeting the second threshold value is selected as the Approximate sentences. This step is that the similar sentence generation module combines the non-overlapping vocabulary according to the sentence pattern predetermined by the system and the order of appearance of potential named entities in the text sentence, so as to generate similar sentences, and then calculates the longest common subsymbol between each similar sentence and the original text sentence sequence, and score each similar sentence, and select the one with the highest score and the second threshold similar sentences as approximate sentences. Among them, the scoring method includes comparing the symbols of the text sentence with the symbols of the similar sentence to score, adding points (such as one point) when the symbols are correct and deducting points (such as one point) when the symbols are wrong, so as to maximize the score and A similar sentence that meets the second threshold is used as the approximate sentence.

In step S503, the sentence intent selection module provides parameters for triggering corresponding services, generates corresponding reply messages, or generates interactive messages according to the number of similar sentences. In this step, the statement intent selection module gives corresponding feedback to the back-end system or platform based on the number of similar sentences generated in the previous step, including when there is only one similar sentence, the module is repeatedly passed to the back-end parameter to generate a trigger The parameters of the corresponding service, or when the approximate sentence is zero or more, the reply message or the interaction message is generated through the reply sentence generation module at the backend.

In other embodiments, before the system executes the operation of semantic understanding, the voice-controlled sentence construction module is used to list important vocabulary combinations and ignore unimportant words in advance through the voice-controlled sentence construction module to construct the pre-stored vocabulary . In other words, through the voice-activated sentence construction module, important vocabulary combinations are found from the training sentences and become pre-stored vocabulary for system analysis.

The present invention is described below with a specific embodiment. In this embodiment, natural language is used to construct the sentence patterns supported by voice-activated services, and two named entity (Named Entity, NE) recognition methods are used to recommend various potential NEs through a potential NE primary selection mechanism, and then combined according to sentence patterns Potential NEs generate similar sentences. Finally, through similar sentence selection formulas, sentence intentions are established, and NE combinations are selected for service trigger parameter fulfillment or reply sentence generation. The present invention integrates the two NE recognition results, and exerts synergistic effects covering missing characters and typos, so it can effectively improve the accuracy of NE recognition.

First, let the voice control service support the sentence patterns in Table 2 below, where {Story} contains stories such as "Three Little Pigs", {Song} contains songs such as "Spider", and {Listen} contains verbs such as "I want to listen".

When the word-missing sentence "I want to listen to three pigs" is input into the system, a character named entity recognizer is used to find potential NEs in the sentence. First, corresponding to the process 301, the Top K vocabulary is recommended according to the matching rate between the sentence and the vocabulary. In this embodiment, the resulting Top K vocabulary is "Three Little Pigs" and "To Listen", namely Top 2. Next, corresponding to the process 302, for each vocabulary, the longest common subsymbol sequence (LCS) between the vocabulary and the sentence is found in a sliding window manner (shifting one word by one word), wherein the LCS length≦vocabulary length.

The calculation process of the character LCS of the word "Three Little Pigs" is as follows. "I want to listen to three pigs" is the word-missing sentence provided initially, and the character comparison is performed with the vocabulary "three little pigs". The character numbers of "I want to listen to three pigs" are 0-5 in sequence.

Comparing the vocabulary "Three Little Pigs" with the text segment "I want to listen to three", the character LCS score = 1-3=-2, where 1 means that there is one character that is the same, and 3 means that there are three characters that are different, so The character LCS score is -2 after subtracting the two.

Comparing the vocabulary "Three Little Pigs" with the fragment of the text sentence "To listen to three", the character LCS score = 2-2=0, wherein, the front 2 means that there are two characters that are the same, and the back 2 means that there are two characters are different, so the character LCS score is 0 after the subtraction of the two.

Comparing the vocabulary "three little pigs" with the segment "listen to three pigs" in the text sentence, the character LCS score = 3-2=1, where 3 means that there are three characters that are the same, and 2 means that there are two characters that are different (there are text replacement), so the character LCS score is 1 after the subtraction of the two.

Compare the vocabulary "three little pigs" with the fragment "three pigs" of the text sentence, the character LCS score = 3-1=2, where 3 means that there are three characters that are the same, and 1 means that there is one character that is different (less characters) , so the character LCS score is 2 after subtracting the two.

To sum up, after the character comparison between "I want to listen to three pigs" and the vocabulary "three little pigs", the LCS position of the result process 302 is (3,6), and the score is 2, that is, starting from number 3 ( Where the word "three" is located), after number 6, it does not include.

In addition, the calculation process of the character LCS of another word "want to listen" is as follows. Similarly, "I want to listen to three pigs" is the word-missing sentence provided initially, and the character comparison is performed with the vocabulary "Yao Ting", and the character numbers of "I want to listen to three pigs" are 0-5 in sequence.

Comparing the vocabulary "want to listen" with the segment "I want" in the text sentence, the character LCS score = 1-1=0, where the preceding 1 indicates that there is a character that is the same, and the following 1 indicates that there is a character that is different, so the character The LCS score is 0 after subtracting the two.

Comparing the vocabulary "YaoListen" with the segment "YaoListen" of the text sentence, the character LCS score = 2-0 = 2, where 2 means that there are two characters that are the same, and the following 0 means whether there are different characters, so the character LCS score After subtracting the two, it is 2.

Comparing the vocabulary "to listen" with the segment "listen three" of the text sentence, the character LCS score = 2-1=1, where 2 means that there are two characters that are the same, and the following 1 means that there is a character that is different, so the character LCS The score is 1 after subtracting the two.

Therefore, after the character comparison of "I want to listen to three pigs" and the word "Yao Ting", the final result of the process 302 is the LCS position (1,3), and the score is 2, that is, starting from number 1 ("Yao" word location), after number 3, it is not included.

Next, corresponding to the process 303, the vocabulary with the highest score and ≧1 is selected from the aforementioned Top 2 vocabulary as a potential NE. As a result, the potential vocabulary is found to be "three little pigs", the LCS is "three pigs" and the score is 2. After that, cover the LCS in the sentence with the symbol ★, and then search for other potential NEs of "I want to listen to ★★★".

The process 301 is executed again, and the Top K words are recommended according to the matching rate between the sentences and the words. As a result, the Top K vocabulary is "to be listened to". Then, the process 302 is executed, and for each vocabulary, the LCS between the vocabulary and the sentence segment is found in the form of a sliding window, wherein the sentence segment length≦vocabulary length.

The calculation process of the character LCS of the vocabulary "To Listen" is as follows. "I want to listen to ★★★" is a masked sentence, which is compared with the word "want to listen". The character numbers of "I want to listen to ★★★" are 0-5 in sequence.

Compare the vocabulary "Yaolisten" with the fragment "I want" of the text sentence, the character LCS score = 1-1=0; compare the vocabulary "Yaolisten" with the fragment "Yaolisten" of the text sentence, the character LCS score = 2-0=2; compare the vocabulary "want to listen" with the segment "listen★" of the text sentence, the character LCS score=2-1=1. Therefore, after the character comparison of "I want to listen ★★★" and the word "YaoListen", the LCS of the result flow 302 is "YaoListen", with a score of 2.

Again, the process 303 is executed, and the vocabulary with the highest score and ≧1 is selected from the Top 1 vocabulary as a potential NE. The potential vocabulary of the result flow 303 is "to listen", the LCS is "to listen" and the score is 2. Afterwards, execute the process 303, cover the LCS in the statement with the symbol ★, and then search for other potential NEs of "I★★★★★". As a result, there are no other potential NEs, enter the process 305, and compile the execution results of the character named entity recognizer as shown in the following table three.

In addition, the musical note named entity recognizer uses musical notes to find potential NEs in a sentence, where the musical note unit is a phonetic symbol of a Chinese character without tone or an English word stem. Therefore, in the present embodiment, the sentence "I want to listen to three pigs" and the vocabulary, the text is converted into phonetic notation, and the process of translating phonetic notation from phonetic notation is as shown in Table 4 below.

Corresponding to the process 301, Top K words are recommended according to the matching rate between sentences and words. As a result, the Top K words were "Three Little Pigs", "Spider" and "To Listen". Next, corresponding to the process 302 , for each vocabulary, the LCS of the vocabulary and the sentence is found out in a sliding window manner, wherein the length of the LCS≦the length of the vocabulary. The comparison process in flow 302 is as follows.

A B C D E F is a sentence note, wherein, the note code is D E G F vocabulary "three little pigs" and sentence fragment "three pigs" comparison, note LCS position (3,6), score 3-1=2; note code is E F The vocabulary "spider" and the sentence fragment "pig", the note LCS position (4,6), score 2-0=2; the note code is B C The vocabulary "want to listen" and the sentence fragment "want to listen", the note LCS position (1,3), fraction 2-0=2.

Next, corresponding to the process 303, the vocabulary with the highest score and ≧1 is selected from the 3 vocabulary notes as the potential NE. As a result, the potential vocabulary in the process 303 is "three little pigs", and the LCS is "three little pigs". Pig" with a score of 2. Afterwards, corresponding to the process 304, the LCS in the sentence is covered by the symbol ★, and then other potential NEs of "I want to listen to ★★★" are searched.

Corresponding to the process 301 again, Top K words are recommended according to the matching rate between sentences and words. As a result, the Top K vocabulary in the process 301 is "to listen to". Then, corresponding to the process 302, for each vocabulary, the LCS of the vocabulary and the sentence is found in the form of a sliding window, wherein the length of the LCS≦the length of the vocabulary.

Similarly, A B C D E F is a sentence note, wherein, the word "want to listen" whose note code is B C is compared with the sentence fragment "want to listen", the note LCS position is (1,3), and the score is 2-0=2.

The process 303 is executed again, and the result is that the potential word is "to listen", the LCS is "to listen" and the score is 2. Afterwards, the process 304 is executed, the LCS in the statement is covered by the symbol ★, and other potential NEs of "I★★★★★" are searched. As a result, there are no other potential NEs, so proceed to process 305, and compile the execution results of the note named entity recognizer as shown in Table 5 below.

After the character named entity recognizer and musical note named entity recognizer are executed, the exploration results (Table 3 and Table 5) are compiled, and the results shown in Table 5 are finally obtained, which will be used for similar sentence generation and similar sentence selection. Corresponding to the process 401, according to the sentence patterns supported by the system and the sequence of appearance of potential NEs, non-overlapping words are combined to generate similar sentences. Corresponding to process 402, calculate the longest common sub-symbol sequence between each similar sentence and the original sentence, and the scoring method is that if the symbol is correct, one point will be added, and if the symbol is wrong (more/less/replacement of a symbol), one point will be deducted.

For example, the characters "I want to listen to three pigs" and the musical notes "A B C D E F" in the original sentence, and the scoring process of each similar sentence are shown in Table 6 below.

Next, corresponding to process 403, select the approximate sentence with the highest score ≧ 0 from all similar sentences. As a result, the similar sentence "Listen to the Three Little Pigs" has the highest score and is recommended as the only similar sentence. From the sentence pattern comparison table 7 below, we can see , the intent of the statement is listenStory. After the vocabulary combination is selected, the process 404 fills in the NE vocabulary into the corresponding execution parameters, and transmits the back-end restful API to trigger the service together with the intent code, for example: /listenStory? Book=Three little pigs. In addition, by replacing the NE in the opening introduction template with the selected NE vocabulary, the opening introduction sentence can be obtained, which can be used to introduce the title of the program before streaming, "playing the story of the three little pigs for you".

If "The Three Little Pigs" is both {Story} and {Song}, the sentence "I want to listen to the Three Pigs" will have two similar sentences with the same score. The intention of "I want to listen to the Three Little Pigs" may be listenStory or listenSong, you need to confirm with the user at this time. Corresponding to process 406, the NE that forms the approximate sentence can be inserted into the option template to get several question options, and the multiple -choice questions are formed to ask the user's intention . Pig ” multiple-choice questions to clarify user intent.

In addition, the present invention also discloses a computer-readable medium, which is applied to a computing device or computer having a processor (for example, CPU, GPU, etc.) and/or memory, and stores instructions, and can be used by this computing device or The computer executes the computer-readable medium through the processor and/or memory, so as to execute the above-mentioned method and each step when executing the computer-readable medium.

The modules, units, devices, etc. of the present invention include a microprocessor and a memory, and algorithms, data, programs, etc. are stored in the memory or a chip, and the microprocessor can load data or algorithms or programs from the memory. Processing such as data analysis or calculation will not be repeated here. In other words, the semantic understanding system for processing rich text of the present invention can be executed on electronic equipment, such as a general computer, tablet or server, and performs analysis and calculation after receiving text sentences, so the semantic understanding system for processing rich text The program to be carried out can be designed and built on electronic devices with processors, memory and other components through software, so as to run on various electronic devices; in addition, various modules of the semantic understanding system for processing rich text or The units are composed of independent components, such as a calculator, a memory, a storage device, or a firmware with a processing unit, all of which can be used to realize the components of the present invention, and related components such as named entity recognizers can also be selected as software Presentation of program, hardware or firmware architecture.

In summary, the semantic understanding system and method for processing rich text of the present invention is used to convert content-rich text into lightweight semantics for easy analysis and understanding. Among them, the potential named entity recognition module integrates multiple NE recognition methods to find potential NE vocabulary, similar sentence generation module According to the sentence patterns supported by the system and the potential NE appearance order of the sentence, the combination of non-overlapping words generates similar sentences, and the sentence intention selection model The group selects the closest to the original sentence from the number of similar sentences, estimates the sentence intent, and selects the NE combination. In other words, for users who use voice-activated services without hesitation, some words are omitted because the program name is too long, or the homonym of the program name is often wrongly written, which affects the voice-activated recognition rate. This invention proposes a framework to construct voice-activated sentences Type, select several Named Entity (NE) recognition results, estimate the intent of the sentence, and use the synergy to deal with missing words and typos in the sentence, so as to improve the quality and accuracy of voice control services. Therefore, the present invention has the following effects.

First of all, it constructs a module framework and summarizes important vocabulary through voice-controlled sentence patterns, and ignores unimportant words, so it does not require a large number of example sentences and a large number of computing resources for training.

Secondly, the Named Entity (NE) selection framework integrates multiple NE recognition results to play a synergistic effect to improve the recognition rate of missing or misspelled words.

The above-mentioned embodiments are for illustrative purposes only, and are not intended to limit the present invention. Anyone skilled in the art can make modifications and changes to the above-mentioned embodiments without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention is defined by the scope of patent application attached to the present invention, as long as it does not affect the effect and implementation purpose of the present invention, it should be included in this disclosed technical content.

1: Semantic understanding system for processing rich text

11: Potential Named Entity Recognition Module

12: Similar sentence generation module

13: Sentence intent selection module

14: Voice-activated sentence construction module

Claims (11)

A semantic understanding system for processing rich text, including: a latent named entity recognition module, which has a plurality of different types of named entity recognizers and receives speech-converted text sentences, wherein, the character named entity recognizer and the musical note named entity The recognizer calculates the matching rate between the symbols of the text sentence and the symbols of the pre-stored vocabulary, picks out the pre-stored words with a high matching rate, and then uses the length of the vocabulary as the range and slides the window to compare the words between the text sentence and the words with a high matching rate. Find the longest common sub-symbol sequence between the pre-stored vocabulary, and compare the symbols of the text sentence with the symbols of the pre-stored vocabulary to score, so as to add points when the symbols are correct and deduct points when the symbols are wrong, so as to select the highest score and meet the The pre-stored vocabulary of the first threshold is used as a potential named entity, and after a potential named entity is found, the common sub-symbol sequence in the word sentence is masked, and then other potential named entities in the word sentence are searched for using the same procedure until finding Until there are no other potential named entities; the similar sentence generation module is used to combine the pre-stored vocabulary corresponding to the potential named entities in a non-overlapping manner according to the sentence patterns supported by the system and the order of appearance of the potential named entities in the text sentence to generate Similar sentences, and then compare the similarities and differences between the symbols of each similar sentence and the text sentence, so as to add points when the symbols are correct and deduct points when the symbols are wrong, so as to select the highest score and meet the second threshold Similar sentences are used as similar sentences; and the sentence intent selection module is used to estimate the intent of the literal sentence based on the intent and quantity of the similar sentence, select the correct named entity, and generate the parameters of the service corresponding to the intent and the The reply message of the text statement. The semantic understanding system for processing rich text as described in Claim 1, wherein the character named entity recognizer uses a Chinese character or an English single character as a symbol unit. The semantic understanding system for processing rich text as described in Claim 1, wherein the phonetic note named entity recognizer uses a Chinese phonetic symbol without tone or an English word stem as a symbol unit. The semantic understanding system for processing rich text as described in Claim 1, wherein, when the sentence intent selection module has only one similar sentence, the parameters for triggering the corresponding service are generated by repeatedly submitting to the module through back-end parameters, When there are multiple similar sentences, use the answer sentence generation module at the back end to generate several question options to form multiple choice questions to ask the user's intent. When there are zero similar sentences, use the answer sentence generation module to generate statement to indicate that there is no corresponding service. The semantic understanding system for processing rich text as described in Claim 1 further includes a voice-controlled sentence structure module, which is used to list important vocabulary combinations and ignore unimportant words in advance from multiple training sentences, so as to construct The system supports sentence patterns and pre-stored vocabulary. A semantic understanding method for processing rich text, which is executed by a computer device, the method includes the following steps: a latent named entity recognition module receives a speech-converted text statement, and uses a character named entity recognizer and a musical note named entity recognizer Calculate the matching rate between the symbols of the text sentence and the symbols of the pre-stored vocabulary, pick out the pre-stored vocabulary with a high matching rate, and then use the length of the vocabulary as the range and slide the window to compare the words between the text sentence and the pre-stored vocabulary with a high matching rate Find the longest common sub-symbol sequence between them, and compare the symbols of the text sentence with the symbols of the pre-stored vocabulary to score, so as to add points when the symbols are correct and deduct points when the symbols are wrong, so as to select the highest score and meet the first The pre-stored vocabulary of the threshold value is used as a potential named entity, and after a potential named entity is found, the common sub-symbol sequence in the word sentence is masked, and then other potential named entities in the word sentence are searched for using the same procedure until none is found. other potential named entities; The similar sentence generation module combines the pre-stored vocabulary of the potential named entities in a non-overlapping manner according to the sentence patterns supported by the system and the order of appearance of the potential named entities in the text sentence to generate similar sentences, and then compares each of the similar sentences For the similarities and differences between the symbols and the text sentences, points are added when the symbols are correct and points are deducted when the symbols are wrong, so that similar sentences with the highest score and meeting the second threshold can be selected as similar sentences; The selection module estimates the intent of the text sentence based on the intent and quantity of the similar sentence, selects the correct named entity, and generates the parameters of the service corresponding to the intent and the reply message of the text sentence. The semantic understanding method for processing rich text as described in Claim 6, wherein the character named entity recognizer uses a Chinese character or an English single character as a symbol unit. The semantic understanding method for processing rich text as described in Claim 6, wherein the phonetic named entity recognizer uses a Chinese phonetic symbol without tones or an English word stem as a symbol unit. The semantic understanding method for processing rich text as described in Claim 6, wherein, when the sentence intent selection module has only one similar sentence, the parameters of the backend are repeatedly passed to the module to generate parameters that trigger the corresponding service, When there are multiple similar sentences, use the answer sentence generation module at the back end to generate several question options to form multiple choice questions to ask the user's intent. When there are zero similar sentences, use the answer sentence generation module to generate statement to indicate that there is no corresponding service. The semantic understanding method for processing rich text as described in Claim 6 further includes, before performing semantic understanding, using a voice-controlled sentence pattern construction module to list important word combinations and ignore unimportant words from multiple training sentences, so as to Construct the sentence patterns supported by the system and the vocabulary to be stored in advance. A computer-readable medium, used in a computing device or a computer, stores instructions to execute the semantic understanding method for processing rich text as described in any one of claims 6 to 10.

Images