CN118838993A - Method for constructing keyword library and related products thereof - Google Patents

Abstract

The present invention discloses a method for constructing a keyword library and related products. The method includes: obtaining text data input by a user in a dialogue system and/or historical log data of user operations; mining keywords in the text data input by the user at a preset frequency of use to obtain high-frequency keywords; mining keywords in the historical log data by a predetermined strategy to obtain extended keywords; analyzing the high-frequency keywords and/or the extended keywords to obtain candidate keywords; and reviewing the candidate keywords so that the keywords that have passed the review are used to construct the keyword library. Through the solution provided by the present invention, high-quality keywords can be accumulated quickly, which improves the efficiency and accuracy of constructing the keyword library in the dialogue system.

Description

A method for building a keyword library and related products

Technical Field

The present invention generally relates to the field of natural language processing technology. More specifically, the present invention relates to a method for building a keyword library, an electronic device and a computer-readable storage medium.

Background Art

With the development of natural language processing technology, keyword extraction technology, as one of the text mining technologies, has been widely used in many fields such as artificial intelligence and semantic recognition. The existing keyword library construction technology is mainly used in search engine optimization, advertising, content recommendation and other fields. Its main purpose is to analyze a large amount of text data and extract representative and relevant keywords in order to better understand and index the text content, thereby improving the accuracy and efficiency of the search. The existing technology works well in building the initial keyword library, but as time goes by and the domain knowledge changes, the keyword library may become outdated. At this time, the keyword library needs to be updated or expanded.

The existing technology usually includes the following steps when constructing an initial keyword library:

Data preprocessing: Clean, segment, and remove stop words on the original text data for subsequent processing.

Feature extraction: Extract keywords from preprocessed text data. Commonly used methods include TF-IDF, TextRank, LDA, etc.

Keyword screening and optimization: According to the set thresholds or rules, representative and relevant keywords are screened out from the extracted keywords and optimized to improve the accuracy and efficiency of the search.

However, the above technical solutions still have some problems, such as the general word segmenter often does not perform well on specific business data, cannot distinguish between user input text and knowledge base text, and traditional word segmentation has no special processing strategy for compound phrases such as noun phrases and verb phrases. In addition, the existing technology usually relies on statistical methods (such as TF-IDF) or graph-based methods (such as TextRank) to extract keywords. These methods may work well when processing large-scale data, but may not be accurate enough in understanding the nuances of specific fields or conversation contexts. Statistical methods such as TF-IDF often mistakenly identify common words as keywords, resulting in the keyword library being filled with non-key information. At the same time, word segmentation cannot directly solve the problem of abbreviations and noun aliases. These problems will further lead to the keyword library being not rich enough and cannot be actually used in downstream tasks. And there are also problems with key phrases composed of multiple words that are difficult to handle.

Therefore, existing keyword libraries often have low keyword accuracy and coverage, and the construction process requires a lot of manual intervention.

In view of this, there is an urgent need to provide a method for constructing a keyword library so as to improve the efficiency and quality of keyword library construction.

Summary of the invention

In order to at least solve one or more of the technical problems mentioned above, the present invention proposes solutions for constructing a keyword library in multiple aspects.

In a first aspect, the present invention provides a method for constructing a keyword library, comprising: obtaining text data input by a user in a dialogue system and/or historical log data of user operations; mining keywords in the text data input by the user at a preset usage frequency to obtain high-frequency keywords; mining keywords in the historical log data through a predetermined strategy to obtain extended keywords; analyzing the high-frequency keywords and/or the extended keywords to obtain candidate keywords; and reviewing the candidate keywords so that the reviewed keywords can be used to construct the keyword library.

In some embodiments, analyzing the high-frequency keywords and/or the extended keywords to obtain candidate keywords includes: acquiring knowledge base text data in the dialogue system; using the knowledge base text data to perform model training to obtain a text analysis model; and using the text analysis model to analyze the high-frequency keywords and/or the extended keywords to obtain candidate keywords.

In some embodiments, analyzing the high-frequency keywords and/or the extended keywords using the text analysis model includes: performing relationship analysis on the high-frequency keywords and/or the extended keywords through part-of-speech joint character similarity rules and vector similarity rules to filter out keyword pairs.

In some embodiments, mining keywords in the text data input by the user with a preset usage frequency to obtain high-frequency keywords includes: performing a keyword usage frequency analysis on the text data input by the user; and based on the results of the keyword usage frequency analysis, screening out keywords whose usage frequency within a predetermined time window is greater than a preset usage frequency to obtain high-frequency keywords.

In some embodiments, analyzing the high-frequency keywords to obtain candidate keywords includes: searching the high-frequency keywords in the knowledge base in the dialogue system; in response to searching the high-frequency keywords in the knowledge base, marking them as known keywords; in response to not searching the high-frequency keywords in the knowledge base, marking them as new keywords; and scoring the new keywords using the distribution frequency and/or category of similar keywords in the knowledge base that are similar to the new keywords for use in analyzing the high-frequency keywords.

In some embodiments, the historical log data of user operations includes one or more of the following: search data of the user in the dialogue system; interaction data generated by the interaction between the user and the question-and-answer assistant in the dialogue system; and content data generated by the user clicking on a content title in the dialogue system.

In some embodiments, mining the keywords in the historical log data through a predetermined strategy to obtain extended keywords includes: applying an edit distance algorithm and/or a longest common subsequence algorithm to the historical log data to obtain the extended keywords; including: sorting the search data in the historical log data by the frequency of use of the search keywords, and selecting a preset number of search keywords in sequence; and calculating the longest continuous common subsequence of the selected search keywords to obtain the extended keywords in the search data.

In some embodiments, the review of the candidate keywords and using the reviewed keywords to construct a keyword library includes: retaining the candidate keywords that have passed the review as valid keywords; performing part-of-speech analysis based on the valid keywords to split the keyword phrases included therein into entities and attributes; generating new keyword phrases by randomly replacing the entities or attributes of the keyword phrases with similar words; using the text analysis model to analyze and evaluate the new keyword phrases and retain reasonable keyword phrases; and using the valid keywords and the reasonable keyword phrases to construct a keyword library.

In a second aspect, the present invention provides an electronic device for constructing a keyword library, comprising: a processor; and a memory on which a computer program for constructing a keyword library is stored. When the computer program is executed by the processor, the method for constructing a keyword library as described in any embodiment of the first aspect is implemented.

In a third aspect, the present invention provides a computer-readable storage medium, wherein a computer-readable program for constructing a keyword library is stored on the storage medium. When the computer-readable program is executed by one or more processors, the method for constructing a keyword library as described in any embodiment of the first aspect is implemented.

Through the method for constructing a keyword library as provided above, the embodiment of the present invention extracts high-frequency keywords and extended keywords from the text data input by the user in the dialogue system and/or the historical log data of the user's operation, and obtains candidate keywords for constructing the keyword library by analyzing the high-frequency keywords and/or extended keywords. Finally, after reviewing the candidate keywords, the keywords for constructing the keyword library are obtained. The solution provided by the present invention can quickly accumulate high-quality keywords, thereby improving the efficiency and accuracy of constructing the keyword library in the dialogue system.

Furthermore, in some embodiments, by using the text data in the knowledge base to train the text analysis model, the data deficiency of the pre-trained model is compensated. The present invention can effectively analyze the nested relationship between keywords and the parts of speech in phrases, and identify topics and subtopics, which will help to further optimize the structure and content of the keyword library, thereby improving the operational efficiency of the question-answering system and the knowledge base.

Therefore, the solution provided by the present invention first determines the keywords through automated mining and then auditing, which can not only quickly accumulate high-quality keywords, but also deeply mine low-frequency long-tail keywords, thereby improving the efficiency and accuracy of building the keyword library in the dialogue system. At the same time, by combining the operation of public search engines and proprietary knowledge bases (such as adding modifications to update knowledge), newly emerging keywords can be captured in a timely manner, so that the keyword library can also be kept updated synchronously. Therefore, the solution of the present invention can enable the keyword library to continue to maintain timeliness and coverage, thereby achieving both the improvement of the operational efficiency of the question-and-answer system and the knowledge base, and the rapid and accurate organization of keywords, which directly helps the analysis of user question tags and the improvement of user satisfaction.

BRIEF DESCRIPTION OF THE DRAWINGS

By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

FIG1 shows an exemplary flow chart of a method for constructing a keyword library according to an embodiment of the present invention;

FIG. 2 shows an exemplary flow chart of keyword analysis according to an embodiment of the present invention; and

FIG. 3 shows an exemplary structural diagram of an electronic device according to an embodiment of the present invention.

DETAILED DESCRIPTION

The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

It should be understood that when the terms "first", "second", "third" and "fourth" are used in the claims, description and drawings of the present invention, they are only used to distinguish different objects, rather than to describe a specific order. The terms "include" and "comprise" used in the description and claims of the present invention indicate the presence of the described features, wholes, steps, operations, elements and/or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and/or their collections.

It should also be understood that the terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the claims, the singular forms of "a", "an" and "the" are intended to include the plural forms unless the context clearly indicates otherwise. It should also be further understood that the term "and/or" used in the specification of the present invention and the claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes these combinations.

As used in this specification and claims, the term "if" may be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" may be interpreted as meaning "upon determination" or "in response to determining" or "upon detection of [described condition or event]" or "in response to detecting [described condition or event]," depending on the context.

In order to facilitate a full understanding of the technical solution of the present invention, some common terms of the question answering (QA) system involved therein from data preprocessing to system evaluation, as well as specific technologies and operating principles are introduced below.

Information Retrieval (IR): refers to the process of retrieving and extracting relevant information from large-scale unstructured data sets. In question-answering systems, information retrieval is often used to find text snippets that may be relevant to a question from a potentially large number of documents.

Knowledge Base (KB): is a database containing structured information, usually composed of knowledge. Each piece of knowledge consists of fields such as title, content, category, content type, etc. The knowledge base is used to store knowledge information that may be needed by the question-answering system.

Natural Language Processing (NLP): refers to the intersection of computer science, artificial intelligence and linguistics, which aims to enable computers to understand, interpret and generate human language.

Named Entity Recognition (NER): It is a task in natural language processing that is used to identify entities with specific meanings in text, such as names, places, organizations, time, etc.

Semantic Parsing: It is the process of converting natural language text into a machine-understandable logical form (such as semantic representation).

Question-Answer Pairs: It is the basic data unit in the question-answering system, consisting of a question and its corresponding correct answer.

Accuracy: It is an indicator to measure the performance of the question-answering system, indicating the proportion of questions answered correctly by the system.

Recall: It is an indicator to measure the performance of the question-answering system, indicating the proportion of questions correctly answered by the system to all possible correctly answered questions.

F1 score: It is the harmonic mean of precision and recall, and is used to comprehensively evaluate the performance of the question-answering system.

Clustering Algorithm: It is an algorithm in the field of data mining and machine learning. It is used to group objects in a data set so that objects in the same group have a high similarity, while objects in different groups have a low similarity. The goal of a clustering algorithm is to discover the inherent structure in the data in order to better understand and interpret the data. Common clustering algorithms include K-means, hierarchical clustering, and DBSCAN.

K-means (K-means clustering): is a simple and widely used clustering algorithm that aims to divide n data points into k clusters so that the distance between each point and the center (i.e., the mean) of its cluster is minimized. The K-means algorithm optimizes cluster formation by iteratively adjusting the position of cluster centers and redistributing data points to the nearest cluster center. This algorithm is suitable for solving various clustering problems, such as market segmentation, social network analysis, and image segmentation.

BERT (Bidirectional Encoder Representations from Transformers): is a pre-trained language representation model based on the Transformer architecture. It learns rich language features and knowledge by pre-training on a large amount of text data. The BERT model can understand the contextual meaning of words, thus achieving significant performance improvements in a variety of natural language processing tasks, such as text classification, named entity recognition, and question-answering systems.

Pre-training: refers to the process of training a model on a large-scale dataset before a specific task. The purpose of pre-training is to enable the model to learn general language knowledge or feature representations, which can be reused in subsequent downstream tasks. Pre-trained models are usually trained through unsupervised learning, for example, using tasks such as masked language models (MLM) or next sentence prediction (NSP).

MLM (Masked Language Model): is a pre-training task used to train a language model to understand the context of words. In the MLM task, some words in the model's input text are randomly masked (for example, replaced with a special [MASK] tag), and the model needs to predict these masked words. In this way, the model learns the meaning and usage of words in a specific context, thereby enhancing its ability to understand language. BERT is a typical example of pre-training through the MLM task.

The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

FIG. 1 shows an exemplary flow chart of a method 100 for constructing a keyword library according to an embodiment of the present invention.

As shown in FIG. 1 , at step S101, the text data input by the user in the dialogue system and/or the historical log data of the user's operation are obtained. In one embodiment scenario, the dialogue system provides a method for the user to input text data such as keywords, phrases, or sentences, so that the user can interact with the dialogue system through the text data, such as information retrieval. In addition, the user's operation in the dialogue system will also be saved as a historical record in the log data by the dialogue system, so that the user's usage habits or historical dialogue results can be quickly responded to in the next use. By analyzing the above text data or historical log data, the interests of the current system users can be quickly grasped, so as to timely improve the knowledge base and retrieval algorithm in the dialogue system. This will improve the user's usage efficiency and accuracy, and improve the user's experience.

In order to better construct a keyword library, the above-obtained text data or historical log data can be used to mine keywords in a targeted manner. For example, text data can be used to mine frequently used keywords, and historical log data can be used to mine extended keywords.

After completing the data acquisition, in step S102, the keywords in the text data input by the user are mined with the preset usage frequency to obtain high-frequency keywords. In one embodiment, the keyword usage frequency analysis is first performed on the text data input by the user. Secondly, according to the results of the keyword usage frequency analysis, the keywords whose usage frequency is greater than the preset usage frequency within the predetermined time window are screened out to obtain high-frequency keywords. For example, the usage frequency can be preset to 3, and 2 to 3 months can be preset as the effective time window. Then, the keyword frequency analysis is performed on the text data through the predetermined effective time window.

Furthermore, the data in the business vocabulary or knowledge base in the dialogue system can be used to process the obtained high-frequency keywords and make corresponding annotations. For example, if the high-frequency keyword already exists in the business vocabulary or knowledge base, it is marked as an existing keyword; otherwise, if it does not exist, it is marked as a new keyword. In the actual technical processing process, there are many technical means to achieve specific annotation forms. For example, annotation is performed through keyword attributes or by setting identifiers, and this application does not impose too many restrictions on this aspect.

Furthermore, in one embodiment, the present invention provides a comprehensive scoring mechanism, which scores keywords based on their distribution frequency and categories in the knowledge base, and retains the results with the highest scores. The purpose of scoring is to find keywords that are valuable to the knowledge base, and comprehensively consider the keyword frequency and scarcity in the knowledge base documents.

The scoring formula can be: comprehensive score = tf * S; (1)

In the above formula (1), tf represents the word frequency score. The word frequency is calculated using the frequency of user input. S represents the importance of the category in which the user's question belongs.

Regarding the calculation method of word frequency score in formula (1), in actual implementation scenarios, there are three situations when inputting question text. The first is that the question input by the user is in the existing business vocabulary of the dialogue system. The second is that the question input by the user is not in the existing business vocabulary, but has appeared in the title and content of the knowledge base. The third is that the complete question in the question input by the user does not appear in the business vocabulary, but after the question is segmented, each individual word in it has appeared in the business vocabulary.

For the first and second cases, the frequency of the current input question in all questions in the business vocabulary or knowledge base can be directly calculated. For the third case, the current input question is segmented, and then the frequency of each word in all questions is calculated and the minimum value is taken.

The specific calculation formula for the category importance in formula (1) is:

S＝Σ((P|Q)＊(N/T)); (2)

In the above formula, S represents the importance of the category; P represents the category of the user question; Q represents the user question; N represents the amount of knowledge under the category; and T represents the total amount of knowledge.

In one embodiment, the detailed calculation method of the category importance score is as follows: first, BERT is used to fine-tune the classification and train on the knowledge base and category data. Then a BERT text classification model is obtained. Finally, BERT is used to calculate the probability of each category of the current input. For example, 0.8 for veterinary medicine, 0.1 for equipment, and 0.1 for information technology. Then the probability of each category is multiplied by the proportion of the knowledge quantity of the category.

Further, in step S103, a technical solution for analyzing and processing the historical log data is provided, that is, mining the keywords in the historical log data through a predetermined strategy to obtain extended keywords. The historical log data of user operations includes one or more of the following: search data of the user in the dialogue system; interaction data generated by the interaction between the user and the question-answering assistant in the dialogue system; and content data generated by the user clicking on the content title in the dialogue system.

In one embodiment, the predetermined strategy includes applying the edit distance algorithm and/or the longest common subsequence algorithm to the historical log data to obtain the extended keywords. Among them, one strategy can be to use the click content title and the search word to perform common word analysis, and then perform the edit distance algorithm and/or the longest common subsequence algorithm based on the analyzed common words to obtain the keywords in the common words as extended keywords; another strategy can be to use the search data in the historical log data, sort the keywords in the search data by the frequency of use, and select a preset number of search keywords in order, for example, from large to small; and calculate the longest continuous common subsequence of the selected search keywords to obtain the extended keywords.

Similar to the process of scoring keywords described in step S102 above, the extended keywords in step S103 can also be scored using algorithms such as formulas (1) and (2) based on the distribution frequency of the extended keywords in the knowledge base and the categories to which they belong, and the results with the highest scores are retained.

After obtaining high-frequency keywords and extended keywords through the above steps S102 and S103, the process proceeds to step S104. This step obtains the knowledge base text data in the dialogue system; uses the knowledge base text data for model training to obtain a text analysis model; and uses the text analysis model to analyze high-frequency keywords and/or extended keywords to obtain candidate keywords. In one embodiment, the specific analysis process may include: obtaining a noun keyword set; using the noun keyword set to generate keyword pairs; and screening keyword pairs through part-of-speech joint character similarity rules and vector similarity rules.

In one embodiment, analyzing high-frequency keywords to obtain candidate keywords includes: searching for high-frequency keywords in a knowledge base in a dialogue system; in response to finding high-frequency keywords in the knowledge base, marking them as known keywords; in response to not finding high-frequency keywords in the knowledge base, marking them as new keywords; and scoring the new keywords using the distribution frequency and/or category of similar keywords in the knowledge base that are similar to the new keywords for use in analyzing the high-frequency keywords.

The specific operation method of step S104 is described below through a complete embodiment process:

First, an unsupervised data training model is constructed through the title and content of the knowledge base of the dialogue system. Specifically, the bidirectional encoding model (Bert) based on Transformer can be used for training.

In one embodiment, the Transformer-based bidirectional encoding model is incrementally unsupervised trained using the "masked language model". In this method, some words in the input text of the BERT model are randomly replaced with a special mask tag (e.g., [MASK]), and the model's task is to predict these masked words, thereby learning the deep semantic and syntactic information in the text. This pre-training task helps the model capture dependencies and contextual information between words, providing powerful semantic representation capabilities for subsequent downstream tasks.

The Transformer-based bidirectional encoding model is a pre-trained model that has a good understanding of open-domain texts, but is not good at learning business-domain texts. Using the knowledge base to build texts for enhanced training makes up for the shortcomings of the pre-trained model.

Secondly, conduct keyword relationship analysis through statistics and analysis.

In one embodiment, natural language processing technology, including part-of-speech analysis (such as the Jieba word segmentation tool), is first used to process the text, so as to count and identify the nested relationship between keywords and the parts of speech of elements in phrases.

Specifically, first obtain all noun keywords. Then sort them according to length, do Cartesian product on the keyword set, and generate n*(n-1)/2 pairs of keywords. Then, further filter the keyword pairs through part-of-speech joint character similarity rule and vector similarity rule.

In one embodiment scenario, keyword phrases are screened by part of speech, character similarity and point mutual information, and the following rules need to be met:

1. Non-verb or noun parts of speech cannot appear after the keyword phrase segmentation.

2. Character similarity uses Jaccard similarity, and phrase pairs with similarity greater than 0.3 are retained.

3. Phrase pairs with extremely high mutual information values.

Through the Transformer-based bidirectional encoder model in the above steps, the phrase text can be converted into a predetermined dimension, such as a 768-dimensional vector. Then, the cosine similarity is used to measure the similarity between the vectors corresponding to the two phrases. Finally, phrase pairs with a similarity greater than 0.85 are retained as candidate phrase pairs.

For candidate phrase pairs, shorter phrases are used as core words, and the phrases are grouped according to the shorter phrases. Each group can retain a set number of phrase pairs, such as retaining 20 phrase pairs.

Secondly, identify the sub-topics in the keywords and their phrases, and perform topic and sub-topic identification.

In one embodiment, for a keyword-corresponding phrase, the keyword (topic) and other components (subtopics) in the phrase are identified, such as a disease keyword and a subtopic constituting a subtopic phrase. The specific implementation may use the above-mentioned unsupervised data training model combined with the K-means clustering algorithm to identify and extract topics and subtopics.

In this process, first, the text is vectorized using an unsupervised data training model, and the output is a high-dimensional vector (such as the 768-dimensional vector mentioned above). Then, the K-means clustering algorithm is used to cluster these vectors to identify the subject entities and related sub-topic information in the text. After clustering, the center point of each cluster is obtained, and a predetermined number of keywords closest to the center point are recalled, such as 50 candidate keywords. Next, N-gram analysis is performed on each keyword, the results are merged, and the N-gram core words with a high probability of occurrence (such as more than 0.2) are counted. These core words can be regarded as potential subject or sub-topic keywords.

Finally, expand keywords based on the topic and sub-topic components.

In one embodiment, the high-quality word set A found in the above steps is collected, audited, valid keywords are retained, and saved to the user dictionary. Then the user dictionary is reloaded and the text is subjected to part-of-speech analysis. Thereafter, these keyword phrases are split into entities and attributes through part-of-speech tagging. For example, for a keyword phrase such as "blue ear disease treatment", it can be split into "blue ear disease" (entity) + "treatment" (attribute). Next, after audit, a series of new keyword phrases can be expanded by randomly replacing entities or attributes in the phrase with similar words. Finally, the semantic rationality of these newly generated keyword phrases is evaluated using a large text model, and those keyword phrases that appear in internal or external knowledge bases or are semantically reasonable are retained, thereby obtaining a new batch of high-quality keyword candidate sets.

Returning to the process shown in FIG. 1 , at step S105, the candidate keywords obtained at step S104 are reviewed so that the keywords that have passed the review are used to construct the keyword library. This step also includes: based on the candidate keywords that have passed the review, they are retained as valid keywords; based on the valid keywords, part-of-speech analysis is performed to split the keyword phrases included therein into entities and attributes; then, new keyword phrases are generated by randomly replacing the entities or attributes of the keyword phrases with similar words; finally, the new keyword phrases are analyzed and evaluated using a text analysis model, and reasonable keyword phrases are retained; and the keyword library is constructed using valid keywords and reasonable keyword phrases.

In summary, based on the teachings and inspiration of the embodiments provided by the present invention, those skilled in the art can understand that the present invention mines keywords based on the input text of the user in the dialogue system. Including: for the text of high-frequency historical user input, combined with the knowledge base of the dialogue system, high-frequency keyword mining of the user input is carried out. For historical log data based on users, an expanded keyword extraction strategy is adopted. Thus, by distinguishing the mining strategy of high-frequency and low-frequency keywords, effective analysis of user input and accurate extraction of keywords are achieved. Among them, high-frequency keyword mining focuses on quickly locking key information, and its main purpose is to quickly accumulate high-quality keywords. Through the high frequency used in user interaction information, high-frequency keywords are efficiently locked, thereby improving the operational efficiency of the question-and-answer system and the knowledge base. Low-frequency keyword mining focuses on identifying and expanding long-tail keywords, and the combination of the two can more comprehensively build a keyword library.

In one embodiment, mining the keywords that are inputted infrequently by users can achieve heuristic mining and expansion based on the keyword phrases already in the knowledge base. Its main purpose is to mine low-frequency long-tail keywords. Specifically, more keywords can be fully identified by fully considering the character semantics and part-of-speech information and combining the accurate understanding ability of the large model.

Furthermore, the technical solution provided by the present invention can also perform keyword relationship analysis and topic identification. Through natural language processing technology and clustering algorithms, the present invention can effectively analyze the nested relationship between keywords and the parts of speech in phrases, and identify topics and subtopics, which helps to further optimize the structure and content of the keyword library.

Furthermore, in one embodiment, the present invention proposes a comprehensive scoring mechanism that combines the TF-IDF algorithm and the classification probability model, as well as rules such as character similarity and point mutual information, to evaluate and screen keywords, thereby ensuring the quality of the keyword library.

FIG2 shows an exemplary process of keyword analysis 200 according to an embodiment of the present invention. Keyword analysis is based on existing keyword phrases and performs heuristic discovery and expansion. Its main purpose is to mine low-frequency long-tail keywords, and by fully considering character semantics and part-of-speech information, combined with the accurate understanding ability of the large model, more keywords can be fully identified.

As shown in FIG. 2 , the process provides specific implementation steps for performing keyword analysis on high-frequency keywords and/or extended keywords.

Specifically, at step S201, the knowledge base text data in the dialogue system is obtained. In the existing dialogue system, the existing knowledge base text includes the business vocabulary and/or the question and answer text data in the knowledge base, etc. By reading the knowledge base text, the existing keywords in the dialogue system can be obtained.

After obtaining the knowledge base text data, the process proceeds to step S202, where the model is trained using the obtained knowledge base text data so as to obtain a text analysis model. The specific process of training the text analysis model is as follows:

First, we build an unsupervised data training model through the knowledge base title and content in the knowledge base text data. Specifically, we can use the Transformer-based bidirectional encoding model (Bert) for pre-training. This pre-training task helps the model capture the dependencies and contextual information between words, and provides powerful semantic representation capabilities for subsequent downstream tasks.

In one embodiment, the Transformer-based bidirectional encoding model is incrementally unsupervised trained using the "masked language model". In this method, some words in the input text of the BERT model are randomly replaced with a special mask token (e.g., [MASK]), and the model's task is to predict these masked words, thereby learning the deep semantic and syntactic information in the text.

The Transformer-based bidirectional encoding model is a pre-trained model that has a good understanding of open-domain texts, but is not good at learning business-domain texts. For example, professional terms or phrases in the aquaculture field have specific explanations, and using the knowledge base to build text for enhanced training can make up for the shortcomings of the pre-trained model.

Further, the process proceeds to step S203, where the high frequency keywords and/or the extended keywords are analyzed using the text analysis model to obtain candidate keywords. In this step, the keyword relationship analysis is mainly performed by statistical and analytical methods.

In one embodiment, natural language processing technology, including part-of-speech analysis (such as the Jieba word segmentation tool), can be used to process the text, so as to count and identify the nested relationship between keywords and the parts of speech of elements in a phrase.

Specifically, all noun keywords are first obtained; then they are sorted according to length, and the keyword set is Cartesian product to generate n*(n-1)/2 pairs of keywords; then, the keyword pairs are further screened through the part-of-speech joint character similarity rule and the vector similarity rule.

In one embodiment scenario, keyword phrases are screened by part of speech, character similarity and point mutual information, and the following rules need to be met:

1. Non-verb or noun parts of speech cannot appear after the keyword phrase segmentation.

2. Character similarity uses Jaccard similarity, and phrase pairs with similarity greater than 0.3 are retained.

3. Phrase pairs with extremely high mutual information values.

By using the Transformer-based bidirectional encoder model in step S202, the phrase text can be converted into a vector of a predetermined dimension, such as a vector of 768 dimensions. Then the cosine similarity is used to measure the similarity between the vectors corresponding to the two phrases. Finally, we retain the phrase pairs with a similarity greater than 0.85 as candidate phrase pairs.

For candidate phrase pairs, the shorter phrase is used as the core word, and then the phrases are grouped according to the shorter phrases, and 20 phrase pairs can be retained in each group.

Secondly, identify the sub-topics in the keywords and their phrases, and perform topic and sub-topic identification.

In one embodiment, for keyword-corresponding phrases, keywords (topics) and other components (subtopics) in the phrases are identified, such as disease keywords and subtopics forming subtopic phrases. The unsupervised data training model described above is combined with the K-means clustering algorithm to identify and extract topics and subtopics.

In an embodiment of the text analysis process, first, the text is vectorized using an unsupervised data training model to output a vector of, for example, 768 dimensions. Then, the vectors are clustered using a K-means clustering algorithm to identify the subject entities and related sub-topic information in the text. After clustering, the center point of each cluster is obtained, and the keywords closest to the center point are recalled in a predetermined number, such as 50 candidate keywords. Next, an N-gram analysis is performed on each keyword, the results are merged, and the N-gram core words with a large probability of occurrence (for example, more than 0.2) are counted. These core words can be regarded as potential subject or sub-topic keywords.

Finally, expand keywords based on the topic and sub-topic components.

In one embodiment, the high-quality word set A found in the above steps is first collected and audited. The audit here can be performed manually or automatically screened using an audit mechanism. For example, for the key word that should be "blue ear disease treatment", the audited word is "after blue ear disease treatment", then the word does not meet the retention standard and should be screened out. Then, for the words that have passed the audit, retain the valid keywords and save them in the user dictionary. Then, reload the user dictionary and perform part-of-speech analysis on the text. And through part-of-speech tagging, these keyword phrases are split into entities and attributes, such as "blue ear disease treatment" can be split into "blue ear disease" (entity) + "treatment" (attribute). Next, after auditing, a series of new keyword phrases are expanded by randomly replacing entities or attributes in phrases with similar words. Finally, the text big model is used to perform semantic rationality assessment on these newly generated keyword phrases, retaining those keyword phrases that appear in internal or external knowledge bases or are semantically reasonable, thereby obtaining a new batch of high-quality keyword candidate sets.

FIG3 schematically shows an exemplary structural diagram of an electronic device 300 according to an embodiment of the present invention. As shown in FIG3, the electronic device 300 of the present invention may include: a processor 301; and a memory 302, on which a computer program for building a keyword library is stored, and when the computer program is executed by the processor 301, the method described in the first aspect or any one of its embodiments is implemented.

Specifically, in the electronic device 300, according to different application scenarios, the processor 301 can adopt various types of chips, such as a central processing unit (CPU), other general-purpose microprocessors, or any conventional processors. In addition, these processors can be selected and configured as needed. For example, if high-speed computing and complex image processing are required, a high-performance GPU or other specialized processors can be selected.

Based on the above, the present invention further discloses a computer-readable storage medium, which contains a computer-readable program for building a keyword library. When the computer-readable program is executed by one or more processors, the method described in the above multiple embodiments or implementation methods is implemented.

In some implementation scenarios, the computer-readable storage medium may be any suitable magnetic storage medium or magneto-optical storage medium. Any such computer storage medium may be part of a device or apparatus, or may be accessible or connectable to a device or apparatus. Any application or module described in the present invention may be implemented using computer-readable/executable instructions that may be stored or otherwise maintained by such a computer-readable medium.

Although multiple embodiments of the present invention have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Those skilled in the art may conceive of many changes, modifications, and alternatives without departing from the thought and spirit of the present invention. It should be understood that in the process of practicing the present invention, various alternatives to the embodiments of the present invention described herein may be adopted. The appended claims are intended to define the scope of protection of the present invention, and therefore cover equivalents or alternatives within the scope of these claims.

Claims (10)

1. A method for constructing a keyword library, comprising: Acquire text data input by the user in the dialogue system and/or historical log data of the user's operations; Mining keywords in the text data input by the user with a preset usage frequency to obtain high-frequency keywords; mining the keywords in the historical log data by a predetermined strategy to obtain extended keywords; Analyzing the high frequency keywords and/or the extended keywords to obtain candidate keywords; and The candidate keywords are reviewed so that the keywords that pass the review are used to construct a keyword library. 2. The method according to claim 1, characterized in that analyzing the high-frequency keywords and/or the extended keywords to obtain candidate keywords comprises: Acquire knowledge base text data in the dialogue system; Using the knowledge base text data to perform model training to obtain a text analysis model; and The high-frequency keywords and/or the extended keywords are analyzed using the text analysis model to obtain candidate keywords. 3. The method according to claim 2, characterized in that analyzing the high-frequency keywords and/or the extended keywords using the text analysis model comprises: Get the noun keyword set; Using the noun keyword set, generating keyword pairs; and The keyword pairs are screened by using part-of-speech combined character similarity rules and vector similarity rules. 4. The method according to claim 1, wherein mining keywords in the text data input by the user with a preset usage frequency to obtain high-frequency keywords comprises: Performing keyword usage frequency analysis on the text data input by the user; and According to the result of the keyword usage frequency analysis, the keywords whose usage frequency is greater than the preset usage frequency within the predetermined time window are screened out to obtain high-frequency keywords. 5. The method according to claim 1 or 4, characterized in that analyzing the high-frequency keywords to obtain candidate keywords includes: Searching the high-frequency keywords obtained in the knowledge base of the dialogue system; In response to searching the high-frequency keyword in the knowledge base, marking it as a known keyword; In response to not finding the high-frequency keyword in the knowledge base, marking it as a new keyword; and The new keyword is scored using the distribution frequency and/or category of similar keywords in the knowledge base that are similar to the new keyword, so as to be used for analyzing the high-frequency keywords. 6. The method according to claim 1, wherein the historical log data of user operations includes one or more of the following: Search data of the user in the dialog system; Interaction data generated by the interaction between the user and the question-answering assistant in the dialogue system; and Content data generated when a user clicks on a content title in the dialog system. 7. The method according to claim 6, characterized in that mining the keywords in the historical log data by a predetermined strategy to obtain extended keywords comprises: Applying an edit distance algorithm and/or a longest common subsequence algorithm to the historical log data to obtain the extended keyword. 8. The method according to claim 1, wherein the step of reviewing the candidate keywords and using the reviewed keywords to construct a keyword library comprises: In response to the candidate keywords that have passed the review, retain them as valid keywords; Performing part-of-speech analysis on the effective keywords to split the keyword phrases included therein into entities and attributes; Generating new keyword phrases by randomly replacing entities or attributes of the keyword phrases with similar words; and Analyze and evaluate the new keyword phrases using the text analysis model, and retain reasonable keyword phrases; The effective keywords and the reasonable keyword phrases are used to construct a keyword library. 9. An electronic device for building a keyword library, comprising: Processor; and A memory having a computer program for constructing a keyword library stored thereon, wherein when the computer program is executed by the processor, the method for constructing a keyword library according to any one of claims 1 to 8 is implemented. 10. A computer-readable storage medium, characterized in that a computer-readable program for constructing a keyword library is stored on the storage medium, and when the computer-readable program is executed by one or more processors, the method for constructing a keyword library according to any one of claims 1 to 8 is implemented.