KR20230153868A - Representative keyword extraction method for online conference summary - Google Patents

Abstract

A method for automatically extracting related words based on the meeting contents of a non-face-to-face meeting is provided. A representative keyword extraction method for non-face-to-face meeting summary according to a preferred embodiment of the present invention is to extract topics included in the meeting conversation by extracting related words according to the frequency of use together within the document in the meeting text, and extracting air information (Co Topic extraction step using -occurrence); And it includes a statistical topic modeling step based on LDA (Latent Dirichlet Allocation) that selects representative keywords by analyzing the correlation between each extracted topic and word.

Description

Representative keyword extraction method for online conference summary}

The present invention relates to a method of extracting representative keywords for summarizing non-face-to-face meetings, and more specifically, to a method of automatically extracting representative keywords based on the meeting contents of a non-face-to-face meeting.

Since the pandemic, non-face-to-face video conferencing has been rapidly increasing, and various video conferencing solutions are emerging. Some solutions allow you to edit recorded content from video conferences to create multimedia meeting minutes containing video, audio, and text.

Looking at the process of creating meeting minutes through a conventional video conference system, video and audio from the start of the meeting to the end of the meeting are recorded and recorded, the data necessary to create the meeting minutes are collected, and the collected data is stored by type. The stored data is collected based on time, and the user creates meeting minutes by editing remarks, keywords, etc. from the collected data.

Meanwhile, in addition to these basic functions, there have been attempts to add various functions related to meeting minutes. For example, Registered Patent No. 10-2061291 proposes a configuration that identifies meeting participants and records their remarks by analyzing the standard voice data of participants and the voice volume of sound signals. Registered Patent No. 10-1290439 extracts important sentences from the meeting minutes and determines the relationships between these sentences to generate a tree-structured summary of the minutes.

The automatic meeting minutes summary generation service has great utility in various fields. Because it allows for effective communication and is also convenient, it is a service that was previously provided by human resources. However, with the advent of AI such as deep learning, it has become a service that can be easily used without investing large resources.

Automatic summary of meeting minutes is performed after analyzing natural language, speech acts, and conversation types using multi-speaker voice recognition technology in a video conference environment. Automatic summarization of meeting minutes generally involves three steps: speech understanding, dialogue management, and summary generation. To create a summary, extracting optimal keywords according to the topic of the conversation must be preceded.

The present invention was made with these points in mind, and its purpose is to provide a method for extracting representative keywords for summarizing meeting minutes of non-face-to-face meetings.

A representative keyword extraction method for non-face-to-face meeting summary according to a preferred embodiment of the present invention is to extract topics included in the meeting conversation by extracting related words according to the frequency of use together within the document in the meeting text, and extracting air information (Co Topic extraction step using -occurrence); And it includes a statistical topic modeling step based on LDA (Latent Dirichlet Allocation) that selects representative keywords by analyzing the correlation between each extracted topic and word.

The topic extraction step using the air information includes examining the frequency with which words are used together in the same document (hereinafter referred to as "air frequency"), and determining word pairs whose air frequency is greater than a predetermined threshold. It may include a step of registering as a related word, and a topic extraction step of extracting topics in the conference conversation from the related word information.

In one embodiment, in the topic extraction step, representative words may be selected for each document, and representative words that have been selected as representative words a predetermined number of times or more in all conversations may be selected as topics. In another embodiment, in the topic extraction step, a word pair whose air frequency satisfies a predetermined condition may be selected as a topic.

In one embodiment, the LDA-based statistical topic modeling step includes calculating the degree of association between keywords and clustering keywords that satisfy predetermined conditions, calculating the influence of each keyword on each topic, and calculating the degree of influence of each keyword within each topic. It includes a step of selecting the keyword with the highest influence as the representative topic keyword of the topic, and a representative keyword selection step of selecting a representative keyword that represents the entire document from among the selected representative topic keywords.

The degree of correlation between the keywords can be defined as the number of documents in which two keywords appear simultaneously in all documents. Clustering keywords that satisfy a predetermined condition may mean clustering keywords with a degree of relevance greater than or equal to a predetermined standard.

The degree of influence of a keyword on a topic can be defined as the probability of becoming a topic when the keyword is present in a sentence.

In one embodiment, in the representative keyword selection step, the keyword with the highest influence among representative topic keywords is selected as the representative keyword.

In one embodiment, before the topic extraction step using the air information, a step of extracting words required for keyword analysis by performing speech act and intention analysis on the uttered sentence is further provided.

According to the present invention, keywords for meeting minutes summary can be efficiently extracted by extracting topics included in meeting conversations according to the frequency of use together in sentences, calculating the influence of keywords on each topic, and selecting representative keywords. .

Figure 1 is a network diagram illustrating a meeting minutes summary environment of a non-face-to-face meeting to which the present invention can be applied.
Figure 2 is a flowchart showing a general meeting minutes summary generation process.
Figure 3 is a flowchart showing the operation flow for performing the representative keyword extraction method for summarizing a non-face-to-face video conference of the present invention.
Figure 4 is a flowchart showing the topic extraction process using co-occurrence.
Figure 5 is a diagram showing an example of extracting words from a conference speech.
Figure 6 is a diagram to explain the process of extracting related words using co-occurrence.
Figure 7 is a flowchart showing the LDA-based statistical topic modeling process.
Figure 8 is an explanatory diagram to explain the LDA-based statistical topic modeling process.

In the present invention, in order to extract video conference keywords, related words are extracted according to the frequency with which they are used together in one document, and topics included in the conference conversation are extracted. This method can be performed utilizing co-occurence. Next, the influence of keywords on each topic is calculated and representative keywords are extracted. Here, a statistical topic modeling technique based on LDA (Latent Dirichlet Allocation) can be applied.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the drawings.

Figure 1 is a network diagram illustrating a meeting minutes summary environment of a non-face-to-face meeting to which the present invention can be applied.

One of the meeting participants connects to the non-face-to-face meeting server 20 and opens a meeting. In the present invention, the type of meeting can be set when establishing a meeting. When the meeting initiator sends a link to access the established meeting to the meeting participants, the meeting participants use the terminal devices 30a, .., 30n such as personal computers, tablet PCs, and smartphones to connect to the meeting through the link. Connect to the face-to-face conference server 20.

When a meeting starts, the meeting minutes recording device 10 records the meeting through video recording or audio recording, and converts the meeting audio into text using the STT (Speech To Text) function. The meeting minutes writing device 10 analyzes the converted text and generates a summary of the meeting minutes.

As shown in FIG. 2, the process of automatically generating a meeting summary may go through three stages: a conversation analysis stage (S10), a related information structuring stage (S20), and a summary generation stage (S30). In the conversation analysis step (S10), the user's utterances in the meeting text are processed into natural language to convert the user's intention into a semantic structure that can be understood by the machine, and in the related information structuring step (S20), related words are extracted based on the semantic structure of the user's utterances. After extracting the optimal keywords, a summary of the meeting minutes is automatically generated based on them in the summary generation step (S30).

The present invention relates to a method of extracting related words and optimal keywords from a conference conversation in the related information structuring step (S20), including a topic extraction step (S100) using co-occurrence as shown in FIG. It includes a statistical topic modeling step (S200) based on LDA (Latent Dirichlet Allocation). In the following description, “topic” and “subject” are used interchangeably.

Meanwhile, before performing steps S100 and S200, a process of extracting words required for keyword analysis can be performed by performing speech act and intention analysis on the uttered sentence. Speech act and intention analysis analyzes speech act information such as questions, statements, and requests through speech act analysis learned based on the intention and speech act knowledge base using morphemes and entity name identification information in sentences. This serves as an important clue to the meeting summary that determines the meeting components. Intention analysis clearly identifies the utterance intention combining the object and action based on the entity name representing the target information and the result of the speech act analysis representing the action. As a method for analyzing speech acts and intentions, for example, methods described in Patent Registration No. 10-2285232, Publication Patent No. 10-2019-0061706, etc. can be used, and the present invention is not limited to a method for analyzing specific speech acts and intentions.

Figure 5 shows an example of extracting words necessary for keyword analysis by performing intention and speech act analysis on each sentence uttered during a meeting. Figure 5 shows an example of extracting words by sentence, but it can also be configured to extract words by multiple sentences uttered by one person at once, that is, by paragraph. Hereinafter, the sentences or paragraphs that have been extracted in this way will be collectively referred to as documents.

In the topic extraction step (S100) using co-occurrence, the topics of conversation are extracted by analyzing words included in the meeting text. The topic extraction step (S100) using co-occurrence includes calculating the frequency of words that appear together as shown in FIG. 4 (S110), registering related words according to the calculation results (S120), It includes a step of extracting a topic from related word information (S130).

In the step of calculating the frequency of words appearing together (S110), the frequency with which words are used together in the same document (hereinafter referred to as "air frequency") is investigated. For example, if there are 100 documents in the meeting text, and the words "marketing" and "public relations" appear together in 15 documents, the air frequency between the two words is 15%.

After examining the air frequency between words for all documents in this way, word pairs whose air frequency is more than a predetermined threshold are registered as associated words from the search results (S120). Referring to FIG. 6, since word A is used together with word C and word D in all documents 1 to 3, word A, word C, and word D are determined to be related words. For example, if the word “Black Friday” is often used together with the word “home appliances,” register “Black Friday” and “home appliances” as related words. More specifically, if the threshold is 5% and the air frequency of “Black Friday” and “home appliances” is 15%, the word pair “Black Friday” and “home appliances” is registered as an associated word.

In the present invention, it is assumed that words that appear together in a document more frequently than a predetermined number of times are grouped together to form one subject (topic). In step S130, topics in the conference conversation are extracted from the related word information.

In one embodiment, representative words may be selected for each sentence, and representative words that have been selected as representative words in the entire conversation a predetermined number of times or more may be selected as topics. In other words, assuming that each sentence has a topic, the most frequently occurring topics can be selected as the meeting topics. In another embodiment, a word pair whose air frequency satisfies a predetermined condition (for example, whether it exceeds a threshold) may be selected as a topic. In another embodiment, words whose frequency of occurrence satisfies a predetermined condition among a predetermined number of word pairs with the highest air frequency may be selected as topics.

In the LDA-based statistical topic modeling step (S200), LDA-based statistical topic modeling is performed according to the number of extracted topics. The LDA-based statistical topic modeling step (S200) selects representative keywords by analyzing the correlation between extracted topics and words. This process will be described with reference to FIGS. 7 and 8.

First, the degree of correlation between keywords appearing in the document is calculated and keywords that satisfy predetermined conditions are clustered (step S210). In one embodiment, the degree of association between keywords can be defined as the number of documents in which A and B appear simultaneously in all documents. In one embodiment, clustering keywords that satisfy a predetermined condition may mean clustering keywords whose relevance is greater than or equal to a predetermined standard. The example in Figure 8 shows a case where keywords are clustered around three topics.

Next, the influence of each keyword on the topic is calculated for the keywords belonging to each topic, and the keyword with the highest influence is selected as the "representative topic keyword" of the topic (step S220). The influence of a keyword on a topic can be defined as the probability of becoming a topic when the keyword is present in a sentence. In other words, for each topic, the influence of each keyword is calculated by probabilistically calculating what topic it will be when a certain word is present.

For example, if there are words "apple", "eating", and "cutting", apple number 2 in the entire document (2 sentences), assuming there are 2 topics "eating apple" and "cutting apple". , Assuming that 1 peeling and 1 eating appear, the contribution of “apple” and “peeling” to the topic “peeling apples” is measured as 50:50. If the sentence “peeling an apple deliciously” appears additionally and is included in the topic “peeling an apple,” the influence of “deliciously” on “peeling an apple” is 20%, and “apple” and “peeling” each have an influence of 20%. It becomes 40%. In this way, the influence of each keyword is calculated for all documents. Through this process, the most influential keywords for each topic can be ultimately selected. In other words, for any topic, which word has the highest probability is calculated. The keyword with the highest impact for each topic is selected as the representative topic keyword for that topic.

In Figure 8, keywords are clustered in one topic (81). In Figure 8, for topic 1, keyword A with the highest influence within the topic with an influence of 0.05 is the representative topic keyword, for topic 2, keyword B with an influence of 0.03 is the representative topic keyword, and for topic 3, keyword C with an influence of 0.04 is the representative topic keyword. (82) was selected.

Next, a “representative keyword” that represents all documents of the meeting is selected from the selected “representative topic keywords” (step S230). In one embodiment, the keyword with the highest influence among representative topic keywords is selected as the representative keyword. In the example of Figure 8, keyword A had the highest influence and was therefore selected as a representative keyword representing the meeting.

The “representative topic keywords” and “representative keywords” selected through this process can be registered as meta information of the document and used for searches and services. Additionally, “representative topic keywords” and “representative keywords” can be used to generate summary information. For example, summary information of a meeting can be created using a summary template corresponding to representative keywords.

Above, the present invention has been described with several examples. However, even though all the components constituting the embodiments of the present invention are described as being combined or operated in combination, the present invention is not necessarily limited to these embodiments. Within the scope of the purpose of the present invention, all of the components may be operated by selectively combining one or more of them. In addition, although all of the components may be implemented as a single independent hardware, a program module in which some or all of the components are selectively combined to perform some or all of the functions of one or more pieces of hardware. It may also be implemented as a computer program having. The codes and code segments that make up the computer program can be easily deduced by a person skilled in the art of the present invention. Such a computer program can be stored in a computer-readable storage medium (computer readable media) and read and executed by a computer or processor, thereby implementing embodiments of the present invention.

Terms such as “include,” “comprise,” or “have,” as used above, unless specifically stated to the contrary, mean that the corresponding component may be present, and do not exclude other components. It should be interpreted that it may further include other components.

The above description is merely an illustrative explanation of the technical idea of the present invention, and various modifications and variations will be possible to those skilled in the art without departing from the essential characteristics of the present invention. Accordingly, the embodiments disclosed in the present invention are not intended to limit the technical idea of the present invention, but are for illustrative purposes, and the scope of the technical idea of the present invention is not limited by these embodiments. The scope of protection of the present invention should be interpreted in accordance with the claims below, and all technical ideas within the equivalent scope should be construed as being included in the scope of rights of the present invention.

10 Meeting minutes recording device,
20 non-face-to-face conference servers,
30a, .., 30n terminal device.

Claims (11)

As a representative keyword extraction method for summarizing meeting minutes on a meeting minutes writing server,
A topic extraction step using co-occurrence, in which topics included in the conference conversation are extracted by extracting related words according to the frequency with which they are used together in the document of the conference text; and
Statistical topic modeling step based on LDA (Latent Dirichlet Allocation) to select representative keywords by analyzing the correlation between each extracted topic and word.
Representative keyword extraction method for non-face-to-face meeting summary including.
The method of claim 1, wherein the topic extraction step using the air information includes:
A step of examining the frequency with which words are used together in the same document (hereinafter referred to as “air frequency”);
registering a word pair whose air frequency is greater than or equal to a predetermined threshold as an associated word;
Topic extraction step of extracting topics in conference conversations from related word information
Representative keyword extraction method for non-face-to-face meeting summary, including.
According to paragraph 2,
In the topic extraction step, representative words are selected for each document and representative words that have been selected as representative words a predetermined number of times in all conversations are selected as topics.
Representative keyword extraction method for non-face-to-face meeting summary.
According to paragraph 2,
The topic extraction step is to select a word pair whose air frequency satisfies a predetermined condition as a topic,
Representative keyword extraction method for non-face-to-face meeting summary.
The method of claim 2, wherein the LDA-based statistical topic modeling step is,
A step of calculating the degree of correlation between keywords and clustering keywords that satisfy predetermined conditions;
Calculating the influence of each keyword on each topic and selecting the keyword with the highest influence within each topic as the representative topic keyword for the topic;
Representative keyword selection step of selecting representative keywords that represent the entire document from among the selected representative topic keywords.
Representative keyword extraction method for non-face-to-face meeting summary, including.
According to clause 5,
A representative keyword extraction method for summarizing non-face-to-face meetings in which the degree of correlation between the keywords is defined as the number of documents in which two keywords appear simultaneously in all documents.
According to clause 6,
Clustering keywords that satisfy a predetermined condition is a method of extracting representative keywords for summarizing non-face-to-face meetings, which involves clustering keywords with a degree of relevance higher than a predetermined standard.
The method of claim 5, wherein the degree of influence on the topic of the keyword is:
A representative keyword extraction method for summarizing non-face-to-face meetings, defined as the probability of becoming a topic when the keyword is present in a sentence.
The method of claim 5, wherein the representative keyword selection step is,
A representative keyword extraction method for summarizing non-face-to-face meetings, which involves selecting the keyword with the highest impact among representative topic keywords as the representative keyword.
According to any one of claims 1 to 9,
A representative keyword extraction method for summarizing a non-face-to-face meeting, further comprising the step of extracting words required for keyword analysis by performing speech act and intent analysis on the uttered sentence before the topic extraction step using the public information.
According to any one of claims 1 to 9,
The above document is a representative keyword extraction method for summarizing non-face-to-face meetings, which is one sentence from a meeting conversation.