Talk skill mining method and device, computer equipment and storage medium

Through semantic recognition and variant prediction of dialogue data, the problem of relying on manual writing in the existing technology of speech mining is solved, and the dynamic expansion of speech library and content quality assurance is realized, which improves the adaptability and richness of speech library. .

CN119025656APending Publication Date: 2024-11-26CHINA PING AN PROPERTY INSURANCE CO LTD

Patent Information

Application Number
CN202411192056.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-27
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

The existing speech mining methods rely on manual writing, resulting in high cost, low efficiency, and the speech library cannot continue to evolve and optimize, and cannot quickly adapt to the needs of different scenarios.

Method used

By semantic recognition of historical dialogue data, comparing it with preset speech topics, non-existent themes are predicted and evaluated. If the score is higher than the threshold, it is stored as a new speech topic. Otherwise, it is deleted to ensure that the speech is Dynamic scaling and content quality of the library.

Benefits of technology

It improves the richness and structure of the speech library, ensures the quality and relevance of the new speech, and realizes the continuous optimization and adaptability of the speech library.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119025656A_ABST
    Figure CN119025656A_ABST
Patent Text Reader

Abstract

The invention discloses a verbal skill mining method and device, computer equipment and a storage medium, and belongs to the technical field of artificial intelligence. Semantic recognition is performed on the historical dialogue data, the recognition result is compared with the preset verbal skill themes in the verbal skill library, and if the matched themes are found, the dialogue data are stored in the corresponding themes; if there is no matched topic, a new verbal variant is generated using a pre-trained variant model. And the variants are graded by using a pre-trained evaluation model, if the score is higher than a threshold value, the variants are stored in the verbal skill library as new verbal skill topics, and otherwise, the variants are deleted. The invention further relates to the technical field of block chains, and the dialogue data can be stored on nodes of the block chain. According to the method, the verbal skill library can be effectively archived and expanded, and the quality and correlation of the newly added verbal skill can be ensured.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application belongs to the field of artificial intelligence technology, and specifically relates to a speech mining method, device, computer equipment and storage medium. Background Art

[0002] Speech mining refers to the process of extracting, analyzing and organizing effective communication strategies and expressions from large amounts of conversation or text data. It aims to discover and optimize effective speech used in specific scenarios (such as sales, customer service, etc.) to improve communication efficiency and effectiveness.

[0003] At present, in order to improve the work efficiency of the agents and to quickly deal with customer objections, agents usually conduct real-time detection of customer conversations or text data during telephone operations, and provide key points of the corresponding speech to complete speech mining. However, the current speech mining is mainly summarized by agents who listen to recordings or read text data and manually write speech mining templates. This method requires a lot of manpower costs and cannot be quickly replicated. In addition, the key points of speech mining templates are usually relatively simple, resulting in relatively solidified speech mining results, which cannot achieve the continuous evolution and optimization of the existing speech library. Summary of the invention

[0004] The purpose of the embodiments of the present application is to propose a method, device, computer equipment and storage medium for speech mining, which are intended to effectively archive the speech and expand the speech library while ensuring the quality and relevance of the newly added speech.

[0005] In order to solve the above technical problems, the present application embodiment provides a method for mining speech, which adopts the following technical solution:

[0006] A method for mining speech techniques, comprising:

[0007] Perform semantic recognition on the historical conversation data to obtain semantic recognition results of the historical conversation data;

[0008] Compare the semantic recognition results with the preset speech topics in the speech library;

[0009] If there is a target speech topic that matches the semantic recognition result, the historical conversation data is stored in the storage space corresponding to the target speech topic in the speech library;

[0010] If there is no target speech topic that matches the semantic recognition result, the speech variants are predicted for the historical conversation data to obtain the variant speech;

[0011] Evaluate the variant speech and output the variant speech score;

[0012] Determine whether the score of the variant speech is greater than the preset score threshold;

[0013] If the variant speech score is greater than or equal to the score threshold, the variant speech topic is extracted and the variant speech topic is stored as a new speech topic in the speech library;

[0014] If the score of the variant is less than the score threshold, the variant is deleted.

[0015] Furthermore, semantic recognition is performed on the historical conversation data to obtain semantic recognition results of the historical conversation data, which specifically include:

[0016] Obtain the speech scores corresponding to the historical conversation data and obtain the historical speech scores;

[0017] Filter historical conversation data based on historical speech scores and score thresholds to obtain target conversation data;

[0018] Perform semantic recognition on the target conversation data to obtain semantic recognition results.

[0019] Furthermore, semantic recognition is performed on the target conversation data to obtain semantic recognition results, which specifically include:

[0020] Perform word segmentation on the target conversation data to obtain the word segmentation result;

[0021] Perform part-of-speech tagging on the word segmentation results to obtain the part-of-speech tagging results;

[0022] Perform named entity recognition on the word segmentation results to obtain named entity recognition results;

[0023] Performing semantic analysis on the word segmentation result to obtain a first semantic analysis result;

[0024] Construct a knowledge graph of the target conversation data based on the results of part-of-speech tagging, named entity recognition, and the first semantic analysis;

[0025] Perform semantic inference on the knowledge graph of the target conversation data to obtain semantic recognition results.

[0026] Furthermore, the historical conversation data is used to predict the speech variants to obtain the speech variants, including:

[0027] Pre-train the preset conditional random field model to obtain a speech variant model;

[0028] Use the speech variant model to extract features from historical conversation data and obtain conversation feature vectors;

[0029] The dialogue feature vector is sequence labeled to obtain a sequence of dialogue variant labels;

[0030] Generate speech variants for the speech variant label sequence to obtain a candidate set of speech variants;

[0031] The speech in the candidate set of variant speech is screened to obtain the variant speech.

[0032] Furthermore, the speech in the candidate set of variant speech is screened to obtain the variant speech, which specifically includes:

[0033] Sort and score the speech in the candidate set of variant speech to obtain the speech ranking score;

[0034] Normalize the ranking scores of the dialogues to obtain the probability distribution of variant dialogues;

[0035] Decode the probability distribution of variant speech to obtain the maximum probability variant speech sequence;

[0036] The variant speech is determined based on the maximum probability variant speech sequence.

[0037] Furthermore, the variant speech is evaluated and the variant speech score is output, which specifically includes:

[0038] Perform word segmentation on the variant speech to obtain the variant speech segmentation;

[0039] Filter emotional keywords from the variant word segmentation, and calculate the emotional score of the variant word based on the emotional keywords;

[0040] Performing logic analysis on variant speech segmentation words respectively, and calculating the logic score of the variant speech based on the results of the logic analysis of the variant speech;

[0041] The variant speech score is calculated based on the sentiment score and the speech logic score.

[0042] Furthermore, emotional keywords are selected from the variant speech segmentation words, and the emotional scores of the variant speech are calculated based on the emotional keywords, specifically including:

[0043] Use a pre-trained sentiment analysis model to determine the emotional tendency of variant speech based on emotional keywords;

[0044] Calculate the sentiment scores of variant words based on sentiment tendency using decision tree algorithm;

[0045] The speech logic analysis includes lexical analysis, grammatical analysis and semantic analysis. The speech logic analysis is performed on the variant speech word segmentation, and the speech logic score of the variant speech is calculated based on the variant speech logic analysis results, which includes:

[0046] Perform lexical analysis on the variant speech segmentation words to obtain lexical analysis results;

[0047] Perform grammatical analysis on the variant speech segmentation words to obtain grammatical analysis results;

[0048] Perform semantic analysis on the variant word segmentation to obtain a second semantic analysis result;

[0049] The speech logic score of the variant speech is calculated based on the lexical analysis result, the grammatical analysis result and the second semantic analysis result.

[0050] In order to solve the above technical problems, the embodiment of the present application further provides a speech mining device, which adopts the following technical solution:

[0051] A speech mining device, comprising:

[0052] A semantic recognition module is used to perform semantic recognition on historical conversation data to obtain semantic recognition results of the historical conversation data;

[0053] The topic comparison module is used to compare the semantic recognition results with the preset speech topics in the speech library;

[0054] A speech storage module is used to store historical conversation data in a storage space corresponding to the target speech topic in the speech library when there is a target speech topic that matches the semantic recognition result;

[0055] The speech variant module is used to predict speech variants on historical conversation data to obtain speech variants when there is no target speech topic that matches the semantic recognition result;

[0056] The speech scoring module is used to evaluate the speech variants and output the speech variant scores;

[0057] A scoring comparison module is used to determine whether the score of the variant speech is greater than a preset scoring threshold;

[0058] A topic storage module is used to extract the variant speech topic when the variant speech score is greater than or equal to the score threshold, and store the variant speech topic as a new speech topic in the speech library;

[0059] The speech deletion module is used to delete the variant speech when the score of the variant speech is less than the score threshold.

[0060] In order to solve the above technical problems, the embodiment of the present application further provides a computer device, which adopts the following technical solution:

[0061] A computer device includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the processor executes the computer-readable instructions, the steps of the speech mining method as described in any one of the above items are implemented.

[0062] In order to solve the above technical problems, the embodiment of the present application further provides a computer-readable storage medium, which adopts the following technical solution:

[0063] A computer-readable storage medium having computer-readable instructions stored thereon, wherein the computer-readable instructions, when executed by a processor, implement the steps of the speech mining method as described in any one of the above.

[0064] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0065] The present application discloses a speech mining method, device, computer equipment and storage medium, which belongs to the field of artificial intelligence technology. The present application performs semantic recognition on historical conversation data, compares the recognition results with the preset speech topics in the speech library, and stores the conversation data in the corresponding topic if a matching topic is found; if there is no matching topic, a pre-trained variant model is used to generate a new speech variant. And the variant is scored using a pre-trained evaluation model. If the score is higher than a threshold, the variant is stored in the speech library as a new speech topic, otherwise the variant is deleted. The present application converts unstructured conversation data into structured information through semantic recognition technology, and then compares the recognition results with the preset speech topic to ensure that the conversation data can be effectively archived. If there is no matching topic, a pre-trained variant model is used to generate a new speech, and the quality is evaluated by the evaluation model. The variant with a score higher than the threshold will be stored in the speech library as a new topic, ensuring the dynamic expansion and content quality of the speech library, which not only improves the richness and structure of the speech library, but also ensures the quality and relevance of the newly added speech. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the scheme in the present application, a brief introduction is given below to the drawings required for use in the description of the embodiments of the present application. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0067] Figure 1 An exemplary system architecture diagram is shown in which the present application can be applied;

[0068] Figure 2 A flowchart of an embodiment of a method for mining speech according to the present application is shown;

[0069] Figure 3 A schematic structural diagram of an embodiment of a speech mining device according to the present application is shown;

[0070] Figure 4 A schematic structural diagram of an embodiment of a computer device according to the present application is shown. DETAILED DESCRIPTION

[0071] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of the present application; the terms used in the specification of the application herein are only for the purpose of describing specific embodiments and are not intended to limit the present application; the terms "including" and "having" and any variations thereof in the specification and claims of the present application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of the present application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0072] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0073] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings.

[0074] like Figure 1 As shown, the system architecture 100 may include terminal devices 101, 102, 103, a network 104 and a server 105. The network 104 is used to provide a medium for communication links between the terminal devices 101, 102, 103 and the server 105. The network 104 may include various connection types, such as wired, wireless communication links or optical fiber cables, etc.

[0075] Users can use terminal devices 101, 102, 103 to interact with server 105 through network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, 103, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0076] Terminal devices 101, 102, 103 can be various electronic devices with display screens and supporting web browsing, including but not limited to smart phones, tablet computers, e-book readers, MP3 players (Moving Picture Experts Group Audio Layer III), MP4 (Moving Picture Experts Group Audio Layer IV) players, laptop computers and desktop computers, etc.

[0077] Server 105 can be a server that provides various services, such as a background server that provides support for pages displayed on terminal devices 101, 102, and 103. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content distribution networks (CDNs), and big data and artificial intelligence platforms.

[0078] It should be noted that the speech mining method provided in the embodiment of the present application is generally executed by a server, and accordingly, the speech mining device is generally set in the server.

[0079] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0080] Continue to refer Figure 2 , showing a flow chart of an embodiment of the speech mining method according to the present application. The embodiment of the present application can acquire and process relevant data based on artificial intelligence technology. Among them, artificial intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0081] The basic technologies of artificial intelligence generally include sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics and other technologies. Artificial intelligence software technology mainly includes computer vision technology, robotics technology, biometrics technology, speech processing technology, natural language processing technology, and machine learning / deep learning. The speech mining method includes the following steps:

[0082] S201, performing semantic recognition on the historical conversation data to obtain a semantic recognition result of the historical conversation data.

[0083] Specifically, natural language processing technology is used to perform semantic recognition on historical conversation data to obtain semantic recognition results of historical conversation data. Natural language processing technology is used to convert the conversation content into semantic information that can be understood by computers. Natural language processing technology involves technologies such as word segmentation, part-of-speech tagging, and named entity recognition. The purpose is to extract key information from the conversation and convert unstructured conversation data into structured semantic information, which has a direct impact on the accuracy and efficiency of subsequent steps.

[0084] For example, in a property insurance sales conversation scenario, suppose a customer and a sales representative have a conversation about: "I recently bought a new car and would like to know more about auto insurance." First, the conversation is processed using semantic recognition technology. Specifically, the conversation text is segmented into words such as "I," "recently," "bought," "a," "new car," "want to know," "auto insurance," and "relevant information." Next, these words are tagged with parts of speech to identify which are nouns, verbs, adjectives, etc. In addition, named entity recognition is also required, such as identifying "new car" as a key entity. Through these processing steps, the system can extract the key information in the conversation, that is, the customer has bought a new car and is interested in auto insurance. These semantic recognition results will serve as input for subsequent steps to prepare for comparison with preset speech topics. In this way, semantic recognition can convert unstructured conversation text into structured semantic information, allowing the system to more accurately understand customer needs and process them accordingly.

[0085] S202, comparing the semantic recognition result with the preset speech topics in the speech library.

[0086] Specifically, by pre-defining the speech library, the semantic recognition results are compared with the preset speech topics in the speech library. The system needs to have a pre-defined speech library, which contains various preset speech topics. The comparison process involves semantic similarity calculation to determine which preset topic the recognition result best matches. This comparison can use methods such as cosine similarity and Euclidean distance. By comparing effectively classified and archived conversation data, the practicality and query efficiency of the speech library are improved, ensuring the full utilization of the existing speech library.

[0087] For example, in the above case, through semantic analysis, the semantic analysis result is obtained that the customer has purchased a new car and is interested in auto insurance. Next, the semantic recognition result is compared with the preset speech topics in the speech library. Assume that the speech library contains several preset topics, such as "Auto Insurance Introduction", "Auto Insurance Quote", "Auto Insurance Claims Process", etc. By calculating the similarity between the semantic recognition results and these preset topics, such as through cosine similarity or Euclidean distance, assuming that the calculation results show that the topic "Auto Insurance Introduction" has the highest similarity with the recognition result, the system will determine that the content of the customer's conversation is most consistent with the topic "Auto Insurance Introduction".

[0088] S203: If there is a target speech topic that matches the semantic recognition result, the historical conversation data is stored in the storage space corresponding to the target speech topic in the speech library.

[0089] Specifically, by comparing the semantic recognition results with the preset speech topics in the speech library, if there is a target speech topic that matches the semantic recognition results, the historical conversation data is directly stored in the storage space corresponding to the target speech topic in the speech library, mainly for data classification and storage, and the processed conversation data is archived by topic for subsequent use and query. Storing the matched historical conversation data in the storage space corresponding to the target speech topic in the speech library can improve the storage and query efficiency of the system, while ensuring the richness and structure of the content of the speech library.

[0090] S204: If there is no target speech topic that matches the semantic recognition result, speech variants are predicted for the historical conversation data to obtain variant speech.

[0091] Specifically, by comparing the semantic recognition results with the preset speech topics in the speech library, if there is no target speech topic that matches the semantic recognition results, the system needs to generate new speech variants through a pre-trained speech variant model. This speech variant model is a text generation model, such as the GPT model, which can generate multiple potential speech variants based on the input conversation data. Speech variant prediction for historical conversation data that cannot be matched enables the system to have the ability to dynamically generate and expand the speech library, improving the system's adaptability to new speech topics.

[0092] For example, in the above case, a pre-trained speech variant model is used to predict speech variants based on historical conversation data. Specifically, the pre-trained speech variant model is used to generate multiple potential speech variants based on the customer's original conversation data. For example, the following variant speech examples are generated as follows:

[0093] "Your newly purchased vehicle requires comprehensive auto insurance protection. Here are several recommended auto insurance options."

[0094] "For your new car, we have several car insurance packages to choose from, detailed below."

[0095] "To get the best protection for your new car, here are the auto insurance options we offer."

[0096] These variants are generated by predicting the variants of the speech based on historical conversation data, aiming to meet customers' consultation needs on auto insurance. In this way, the system can dynamically generate new variants of the speech, enrich the content of the speech library, and ensure that it can adapt to various customer consultation scenarios.

[0097] S205, evaluating the variant speech and outputting the variant speech score.

[0098] Specifically, the pre-trained evaluation model is used to evaluate the variant speech, and the variant speech score is output to determine whether it can be saved to the speech library for subsequent use. The quality of the generated speech variants is evaluated and the variants are scored using the pre-trained speech evaluation model. The speech evaluation model can evaluate the variant speech based on some specific evaluation indicators, such as emotional tendency, speech fluency, relevance, information content, etc., and output the variant speech score. The evaluation of the variant speech is quality controlled through the pre-trained model to ensure that the newly added speech topics meet the quality standards, which provides a guarantee for the quality of the speech library.

[0099] S206, determining whether the score of the variant speech is greater than a preset score threshold;

[0100] S207, if the variant speech score is greater than or equal to the score threshold, extract the variant speech topic, and store the variant speech topic as a new speech topic in the speech library;

[0101] S208: If the score of the variant speech is less than the score threshold, the variant speech is deleted.

[0102] Specifically, determine whether the variant speech score is greater than the preset scoring threshold; if the variant speech score is greater than or equal to the scoring threshold, extract the variant speech topic and store the variant speech topic as a new speech topic in the speech library; if the variant speech score is less than the scoring threshold, delete the variant speech. These steps determine the storage of variant speech. Variants with scores higher than the threshold will be added to the speech library as new speech topics, while those with low scores will be discarded to ensure the content quality in the speech library. The introduction of the scoring mechanism not only screens out high-quality variant speech topics, but also prevents low-quality speech from entering the speech library, maintaining the high standards and practicality of the speech library.

[0103] For example, in the above case, these scores are compared with the preset score threshold. Assuming the threshold is 75 points, it means that the variant speech with a score greater than or equal to 75 points will be considered qualified, and the variant speech with a score greater than or equal to 75 points will be extracted as a new speech topic and stored in the speech library, while the variant speech with a score lower than 75 points will be deleted.

[0104] In the above embodiment, the present application converts unstructured conversation data into structured information through semantic recognition technology, and then compares the recognition results with the preset speech topics to ensure that the conversation data can be effectively archived. If there is no matching topic, a pre-trained variant model is used to generate new speech, and its quality is evaluated through the evaluation model. Variants with scores higher than the threshold will be stored in the speech library as new topics to ensure the dynamic expansion and content quality of the speech library, which not only improves the richness and structured nature of the speech library, but also ensures the quality and relevance of the newly added speech.

[0105] Furthermore, semantic recognition is performed on the historical conversation data to obtain semantic recognition results of the historical conversation data, which specifically include:

[0106] Obtain the speech scores corresponding to the historical conversation data and obtain the historical speech scores;

[0107] Filter historical conversation data based on historical speech scores and score thresholds to obtain target conversation data;

[0108] Perform semantic recognition on the target conversation data to obtain semantic recognition results.

[0109] In this embodiment, the historical conversation data is processed in multiple steps to extract semantic information of the historical conversation data. First, the speech scores corresponding to the historical conversation data are obtained to evaluate their quality and effectiveness. Then, based on these scores and the preset score threshold, the historical conversation data is screened to retain only high-quality target conversation data, thereby ensuring the basic data quality for subsequent processing. Then, semantic recognition is performed on these target conversation data to convert them into structured semantic information.

[0110] In the above embodiment, multi-level screening and processing are used to ensure that the semantic recognition results finally extracted are accurate and useful, providing a reliable data foundation for further speech mining and application.

[0111] Furthermore, semantic recognition is performed on the target conversation data to obtain semantic recognition results, which specifically include:

[0112] Perform word segmentation on the target conversation data to obtain the word segmentation result;

[0113] Perform part-of-speech tagging on the word segmentation results to obtain the part-of-speech tagging results;

[0114] Perform named entity recognition on the word segmentation results to obtain named entity recognition results;

[0115] Performing semantic analysis on the word segmentation result to obtain a first semantic analysis result;

[0116] Construct a knowledge graph of the target conversation data based on the results of part-of-speech tagging, named entity recognition, and the first semantic analysis;

[0117] Perform semantic inference on the knowledge graph of the target conversation data to obtain semantic recognition results.

[0118] In this embodiment, first, the target conversation data is segmented to decompose the text into independent words. Next, the conditional random field (CRF) algorithm is used to perform part-of-speech tagging on the word segmentation results to identify the part of speech of each word. Then, the hidden Markov model (HMM) is used to perform named entity recognition on the word segmentation results to identify key entities such as names and place names. After that, the maximum entropy model (MaxEnt) is used to perform semantic analysis on the word segmentation results to extract the semantic relations in the text. Based on the part-of-speech tagging results, named entity recognition results and semantic analysis results, a knowledge graph of the target conversation data is constructed using a knowledge graph embedding model, and this information is integrated into a structured knowledge representation. Finally, semantic inference is performed on the knowledge graph to mine implicit semantic relations and information to obtain the final semantic recognition results.

[0119] The CRF algorithm is used to perform part-of-speech tagging on the target dialogue data. CRF is a sequence tagging model that can effectively consider contextual information, process dependencies between words, accurately identify the part-of-speech of words, and maximize the joint probability of input data and label sequences by constructing conditional probability distribution, thereby achieving accurate part-of-speech tagging.

[0120] HMM is a probabilistic model that infers the implicit state sequence (namely named entities) by observing the sequence (namely the word segmentation result). By using the state transition probability and observation probability, it can effectively identify key entities in the text, such as names of people, places, organizations, etc., and improve the structured degree of dialogue data.

[0121] The maximum entropy model is a classification method that avoids overfitting by maximizing the entropy value of the model. It is suitable for multi-category classification problems. In semantic analysis, MaxEnt can process complex features and contextual information, extract semantic relationships in conversation data, such as actions, intentions, etc., and enhance text understanding.

[0122] The knowledge graph embedding model embeds the nodes and relationships in the knowledge graph into a low-dimensional vector space, retaining semantic information and structural features. By integrating part-of-speech tagging, named entity recognition, and semantic analysis results, the embedding model creates a structured knowledge graph to facilitate further semantic inference.

[0123] The semantic inference analysis of the knowledge graph is used in the solution to infer the knowledge graph and mine implicit semantic relationships and information. Through inference, it can reveal relationships and patterns that are not explicitly expressed in the conversation data, and enhance the ability to understand and interpret semantic recognition results.

[0124] For example, in the above-mentioned property insurance sales dialogue scenario, first, the historical dialogue data is segmented to decompose the dialogue content into meaningful word units. For example, the customer asks about the details of vehicle insurance, and the question is segmented, such as keywords such as "vehicle" and "insurance". Next, the conditional random field (CRF) algorithm is used for part-of-speech tagging to determine the grammatical role of each word in the sentence in order to understand the structure and intention of the sentence. In this example, part-of-speech tagging can help the system identify that "vehicle" is a noun and "insurance" is a noun, and infer their functions in the sentence. Then, the hidden Markov model (HMM) is used for named entity recognition to identify specific named entities, such as specific car models or insurance company names, to further refine the understanding of the dialogue content. For example, a specific vehicle brand is identified. Next, the maximum entropy model (MaxEnt) is used for semantic analysis to deeply understand the meaning and context of the sentence to ensure accurate responses to customer questions, so as to help the system understand whether the customer is asking about a specific type of car insurance or insurance policy.

[0125] In the above embodiment, complex conversation data can be converted into structured, semantically rich knowledge representation, improving the ability to understand and process conversation data, providing a basis for subsequent conversation topic comparison, and ensuring the accuracy and comprehensiveness of semantic recognition results through a multi-level, multi-model combination approach.

[0126] Furthermore, the historical conversation data is used to predict the speech variants to obtain the speech variants, including:

[0127] Pre-train the preset conditional random field model to obtain a speech variant model;

[0128] Use the speech variant model to extract features from historical conversation data and obtain conversation feature vectors;

[0129] The dialogue feature vector is sequence labeled to obtain a sequence of dialogue variant labels;

[0130] Generate speech variants for the speech variant label sequence to obtain a candidate set of speech variants;

[0131] The speech in the candidate set of variant speech is screened to obtain the variant speech.

[0132] In this embodiment, first, a preset conditional random field (CRF) model is pre-trained to obtain a model specifically used for speech variant generation. The pre-training process involves using a large amount of conversation data to adjust the model parameters so that it can effectively extract and understand the features in the conversation. Next, the speech variant model is used to extract features from the historical conversation data to obtain conversation feature vectors. These feature vectors represent the core information and structure of the conversation data. Then, these feature vectors are sequence labeled to generate a speech variant label sequence to identify the key elements and structural patterns in the conversation. Next, a speech variant candidate set is generated based on these label sequences to form multiple potential speech variants. Finally, the variant speech candidate set is screened to select the best and most relevant variant speech.

[0133] In the above embodiments, the present application can generate diverse and high-quality speech variants through a pre-trained speech variant model to meet the needs of different dialogue scenarios, enhance the adaptability and practicality of the speech library, ensure the diversity and relevance of the generated speech, and at the same time ensure the quality of the speech through the screening step.

[0134] Furthermore, the speech in the candidate set of variant speech is screened to obtain the variant speech, which specifically includes:

[0135] Sort and score the speech in the candidate set of variant speech to obtain the speech ranking score;

[0136] Normalize the ranking scores of the dialogues to obtain the probability distribution of variant dialogues;

[0137] Decode the probability distribution of variant speech to obtain the maximum probability variant speech sequence;

[0138] The variant speech is determined based on the maximum probability variant speech sequence.

[0139] In this embodiment, variant speech is screened and optimized through multiple steps to determine the final variant speech. First, the bidirectional long short-term memory network (Bi LSTM) is used to sort and score the speech in the candidate set of variant speech. Bi LSTM can process and capture the contextual information in the speech, thereby assigning a ranking score to each variant speech. Then, these ranking scores are normalized using the softmax normalization function and converted into the probability distribution of variant speech, ensuring that the sum of the scores is 1, so that the score of each variant speech can be interpreted as a probability. Subsequently, the probability distribution of the variant speech is decoded by the Viterbi algorithm to find the variant speech sequence with the maximum probability. The Viterbi algorithm is a dynamic programming algorithm that can efficiently find the sequence with the highest probability. Finally, the final variant speech is determined based on the maximum probability variant speech sequence.

[0140] In the above embodiment, through the steps of sorting, normalization, decoding and selection, it is ensured that the selected variant speech not only meets the semantic and contextual requirements, but also has the highest quality and relevance. By combining deep learning and optimization algorithms, the accuracy and effectiveness of speech screening are ensured, and the overall quality and practicality of the speech library are improved.

[0141] Furthermore, the variant speech is evaluated and the variant speech score is output, which specifically includes:

[0142] Perform word segmentation on the variant speech to obtain the variant speech segmentation;

[0143] Filter emotional keywords from the variant word segmentation, and calculate the emotional score of the variant word based on the emotional keywords;

[0144] Performing logic analysis on variant speech segmentation words respectively, and calculating the logic score of the variant speech based on the results of the logic analysis of the variant speech;

[0145] The variant speech score is calculated based on the sentiment score and the speech logic score.

[0146] In this embodiment, first, the variant speech is segmented to decompose the speech text into independent words. Then, emotional keywords are selected from these segmented words, and the emotional components in the speech are identified using a predefined emotional dictionary or emotional analysis model, and the emotional score of the variant speech is calculated based on these emotional keywords to evaluate the emotional expression quality of the speech. Then, the segmentation results are subjected to speech logic analysis, which involves analyzing the structure, coherence and logical relationship of the speech to ensure that the speech is logically coherent and organized. Based on these analysis results, the system calculates the speech logic score of the variant speech to evaluate the logic and consistency of the speech. Finally, the emotional score and the speech logic score are combined, and the final score of the variant speech is calculated through a certain weighting or combination method.

[0147] In a specific embodiment of the present application, the method for calculating the variant speech score based on the sentiment score and the speech logic score can be implemented by weighted averaging. Assume that the weights of the sentiment score and the speech logic score are 0.4 and 0.6 respectively. First, the variant speech is segmented, the sentiment keywords are screened out and the sentiment score is calculated, for example, the sentiment score is 80 points. Then, the dialogue is logically analyzed and the speech logic score is calculated, for example, the logic score is 90 points. Next, the final score of the variant speech is calculated based on the weight: variant speech score = (emotion score × weight) + (speech logic score × weight), that is, 80 points × 0.4 + 90 points × 0.6 = 32 points + 54 points = 86 points. Through this weighted average method, the quality of both emotions and logic is comprehensively considered to obtain the final score of the variant speech.

[0148] In the above embodiment, the final score of the variant speech reflects the overall quality of the variant speech in terms of emotion and logic. By evaluating the variant speech from multiple aspects, it is ensured that the selected variant speech not only has good emotional expression, but also has logical coherence and rationality, thereby improving the overall quality of the speech library and user satisfaction.

[0149] Furthermore, emotional keywords are selected from the variant speech segmentation words, and the emotional scores of the variant speech are calculated based on the emotional keywords, specifically including:

[0150] Use a pre-trained sentiment analysis model to determine the emotional tendency of variant speech based on emotional keywords;

[0151] Calculate the sentiment scores of variant words based on sentiment tendency using decision tree algorithm;

[0152] In this embodiment, the sentiment score of the variant speech is calculated by using a pre-trained sentiment analysis model and a decision tree algorithm. First, the system performs word segmentation on the variant speech and screens out sentiment keywords. Then, the pre-trained sentiment analysis model is used to determine the sentiment tendencies of these sentiment keywords, such as positive, negative, or neutral. Next, based on these sentiment tendencies, a decision tree algorithm is used to synthesize the sentiment tendencies of each sentiment keyword and calculate the overall sentiment score of the variant speech. The decision tree algorithm can effectively evaluate the emotional expression quality of the variant speech through hierarchical decision-making and conditional judgment.

[0153] When using the decision tree algorithm to calculate the overall sentiment score of the variant speech, the sentiment tendency of each sentiment keyword is first classified, such as positive, negative, and neutral. Each node of the decision tree represents an sentiment keyword and its corresponding sentiment tendency, and each leaf node represents a final sentiment score. By traversing the decision tree, starting from the root node, conditional judgments are made based on the sentiment keyword tendency, and the corresponding branches are gradually selected downward until the leaf node is reached to obtain the final sentiment score. The decision tree determines the optimal split point and path based on the training data, so that the comprehensively calculated sentiment score can accurately reflect the sentiment tendency of the entire speech.

[0154] In the above embodiment, it is ensured that the emotion score accurately reflects the emotional characteristics of the speech, providing a reliable basis for the subsequent comprehensive score.

[0155] Furthermore, the speech logic analysis includes lexical analysis, grammatical analysis and semantic analysis. The speech logic analysis is performed on the variant speech segmentation words respectively, and the speech logic score of the variant speech is calculated based on the variant speech logic analysis results, which specifically includes:

[0156] Perform lexical analysis on the variant speech segmentation words to obtain lexical analysis results;

[0157] Perform grammatical analysis on the variant speech segmentation words to obtain grammatical analysis results;

[0158] Perform semantic analysis on the variant word segmentation to obtain a second semantic analysis result;

[0159] The speech logic score of the variant speech is calculated based on the lexical analysis result, the grammatical analysis result and the second semantic analysis result.

[0160] In this embodiment, through lexical analysis, grammatical analysis and semantic analysis, a comprehensive logical analysis of the variant speech is performed, and its logical score is calculated. First, the variant speech segmentation is subjected to lexical analysis to check the part of speech and correctness of each word to ensure that the word usage complies with the language rules. Next, a grammatical analysis is performed to analyze the relationship between words and the structure of sentences to ensure that the sentence structure is correct and the logic is clear, using technologies such as dependency syntactic analysis or phrase structure tree. Then, a semantic analysis is performed on the variant speech to understand the semantic relationship and context in the sentence to ensure that the speech is semantically coherent and consistent. Semantic analysis can use methods such as hidden Markov model (HMM) or maximum entropy model (MaxEnt). Based on the lexical analysis results, grammatical analysis results and semantic analysis results, this information is combined to calculate the speech logic score of the variant speech.

[0161] In the above embodiment, the speech logic score reflects the logical integrity and coherence of the speech, ensuring that there is no ambiguity or logical error in the expression of the speech. Through multi-level analysis and comprehensive scoring, the logical quality of the variant speech can be effectively evaluated and the overall level of the speech library can be improved.

[0162] In this embodiment, the electronic device (eg Figure 1 The server shown in the figure) can receive instructions or obtain data through a wired connection or a wireless connection. It should be noted that the above wireless connection method may include but is not limited to 3G / 4G connection, WiFi connection, Bluetooth connection, WiMAX connection, Zigbee connection, UWB (ultra wideband) connection, and other wireless connection methods currently known or to be developed in the future.

[0163] It should be emphasized that in order to further ensure the privacy and security of the above conversation data, the above conversation data can also be stored in a blockchain node.

[0164] The blockchain referred to in this application is a new application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanism, encryption algorithm, etc. Blockchain is essentially a decentralized database, a string of data blocks generated by cryptographic methods. Each data block contains a batch of network transaction information, which is used to verify the validity of its information (anti-counterfeiting) and generate the next block. Blockchain can include the underlying blockchain platform, platform product service layer, and application service layer.

[0165] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing related hardware through computer-readable instructions, and the computer-readable instructions can be stored in a computer-readable storage medium. When the computer-readable instructions are executed, they can include the processes of the embodiments of the above-mentioned methods. Among them, the aforementioned storage medium can be a non-volatile storage medium such as a disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0166] It should be understood that, although the steps in the flowchart of the accompanying drawings are displayed in sequence as indicated by the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least a part of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

[0167] Further references Figure 3 , as a response to the above Figure 2 The present application provides an embodiment of a speech mining device, and the device embodiment is similar to Figure 2 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic devices.

[0168] like Figure 3 As shown, the speech mining device 300 described in this embodiment includes:

[0169] The semantic recognition module 301 is used to perform semantic recognition on the historical conversation data to obtain the semantic recognition result of the historical conversation data;

[0170] Theme comparison module 302 is used to compare the semantic recognition result with the speech theme preset in the speech library;

[0171] The speech storage module 303 is used to store the historical conversation data in the storage space corresponding to the target speech topic in the speech library when there is a target speech topic that matches the semantic recognition result;

[0172] A speech variant module 304 is used to predict speech variants on historical conversation data to obtain speech variants when there is no target speech topic that matches the semantic recognition result;

[0173] The speech scoring module 305 is used to evaluate the speech variants and output the speech variant scores;

[0174] The score comparison module 306 is used to determine whether the score of the variant speech is greater than a preset score threshold;

[0175] The theme storage module 307 is used to extract the variant speech theme when the variant speech score is greater than or equal to the score threshold, and store the variant speech theme as a new speech theme in the speech library;

[0176] The speech deletion module 308 is used to delete the variant speech when the score of the variant speech is less than the score threshold.

[0177] Furthermore, the semantic recognition module 301 specifically includes:

[0178] The first speech scoring unit is used to obtain speech scores corresponding to historical conversation data to obtain historical speech scores;

[0179] A data screening unit, used to screen the historical conversation data based on the historical speech scores and the score threshold to obtain the target conversation data;

[0180] The semantic recognition unit is used to perform semantic recognition on the target dialogue data to obtain a semantic recognition result.

[0181] Furthermore, the semantic recognition unit specifically includes:

[0182] The first word segmentation subunit is used to perform word segmentation processing on the target dialogue data to obtain a word segmentation result;

[0183] The part-of-speech tagging subunit is used to perform part-of-speech tagging on the word segmentation result to obtain a part-of-speech tagging result;

[0184] A named entity recognition subunit is used to perform named entity recognition on the word segmentation result to obtain a named entity recognition result;

[0185] A first semantic analysis subunit is used to perform semantic analysis on the word segmentation result to obtain a first semantic analysis result;

[0186] A knowledge graph construction subunit, used to construct a knowledge graph of the target conversation data based on the part-of-speech tagging result, the named entity recognition result, and the first semantic analysis result;

[0187] The semantic inference subunit is used to perform semantic inference on the knowledge graph of the target conversation data to obtain semantic recognition results.

[0188] Furthermore, the speech variation module 304 specifically includes:

[0189] A variant model training unit, used to pre-train a preset conditional random field model to obtain a speech variant model;

[0190] A feature extraction unit, used to extract features from historical conversation data using a speech variant model to obtain a conversation feature vector;

[0191] The sequence labeling unit is used to label the dialogue feature vectors and obtain the sequence of dialogue variant labels.

[0192] A speech variant generation unit, used for generating speech variants for a speech variant label sequence to obtain a candidate set of variant speeches;

[0193] The speech screening unit is used to screen the speech in the variant speech candidate set to obtain the variant speech.

[0194] Furthermore, the speech screening unit specifically includes:

[0195] The ranking and scoring subunit is used to rank and score the speech in the candidate set of variant speech to obtain the speech ranking score;

[0196] A normalization processing subunit is used to normalize the dialogue ranking scores to obtain the probability distribution of variant dialogues;

[0197] A probability decoding subunit, used to decode the probability distribution of variant speech and obtain the maximum probability variant speech sequence;

[0198] The speech screening subunit is used to determine the variant speech based on the maximum probability variant speech sequence.

[0199] Furthermore, the speech scoring module 305 specifically includes:

[0200] The second word segmentation sub-unit is used to perform word segmentation processing on the variant speech to obtain the variant speech segmentation;

[0201] The sentiment scoring unit is used to select sentiment keywords from the variant speech segmentation words and calculate the sentiment score of the variant speech based on the sentiment keywords;

[0202] A speech logic scoring unit is used to perform speech logic analysis on variant speech segment words respectively, and calculate the speech logic score of the variant speech based on the variant speech logic analysis result;

[0203] The speech scoring unit is used to calculate the variant speech scores based on the sentiment score and the speech logic score.

[0204] Furthermore, the sentiment scoring unit specifically includes:

[0205] An emotional tendency determination subunit, used to determine the emotional tendency of the variant speech based on emotional keywords using a pre-trained sentiment analysis model;

[0206] The sentiment scoring subunit is used to calculate the sentiment scores of the variant speech words based on the sentiment tendency by using a decision tree algorithm;

[0207] Furthermore, the speech logic analysis includes lexical analysis, grammatical analysis and semantic analysis. The speech logic scoring unit specifically includes:

[0208] The lexical analysis subunit is used to perform lexical analysis on the variant speech segmentation words to obtain the lexical analysis results;

[0209] The grammatical analysis subunit is used to perform grammatical analysis on the variant speech segmentation words to obtain grammatical analysis results;

[0210] The second semantic analysis subunit is used to perform semantic analysis on the variant speech segmentation to obtain a second semantic analysis result;

[0211] The speech logic scoring subunit is used to calculate the speech logic score of the variant speech based on the lexical analysis result, the grammatical analysis result and the second semantic analysis result.

[0212] In the above embodiment, the present application discloses a speech mining device, which belongs to the field of artificial intelligence technology. The present application performs semantic recognition on historical conversation data, compares the recognition results with the preset speech topics in the speech library, and if a matching topic is found, the conversation data is stored in the corresponding topic; if there is no matching topic, a pre-trained variant model is used to generate a new speech variant. And the variant is scored using a pre-trained evaluation model. If the score is higher than the threshold, the variant is stored in the speech library as a new speech topic, otherwise the variant is deleted. The present application converts unstructured conversation data into structured information through semantic recognition technology, and then compares the recognition results with the preset speech topics to ensure that the conversation data can be effectively archived. If there is no matching topic, a pre-trained variant model is used to generate a new speech, and the quality is evaluated by the evaluation model. The variant with a score higher than the threshold will be stored in the speech library as a new topic, ensuring the dynamic expansion and content quality of the speech library, which not only improves the richness and structure of the speech library, but also ensures the quality and relevance of the newly added speech.

[0213] To solve the above technical problems, the present application also provides a computer device. Figure 4 , Figure 4 This is a basic structural block diagram of the computer device in this embodiment.

[0214] The computer device 4 includes a memory 41, a processor 42, and a network interface 43 that are interconnected through a system bus. It should be noted that the figure only shows a computer device 4 with components 41-43, but it should be understood that it is not required to implement all the components shown, and more or fewer components can be implemented instead. Among them, those skilled in the art can understand that the computer device here is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to microprocessors, application specific integrated circuits (Application Specific Integrated Circuit, ASIC), field-programmable gate arrays (Field-Programmab l eGate Array, FPGA), digital processors (Digital Signal Processor, DSP), embedded devices, etc.

[0215] The computer device may be a computing device such as a desktop computer, a notebook, a PDA, a cloud server, etc. The computer device may interact with a user through a keyboard, a mouse, a remote controller, a touch pad, or a voice control device.

[0216] The memory 41 includes at least one type of readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the memory 41 can be an internal storage unit of the computer device 4, such as a hard disk or memory of the computer device 4. In other embodiments, the memory 41 can also be an external storage device of the computer device 4, such as a plug-in hard disk equipped on the computer device 4, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (FlashCard), etc. Of course, the memory 41 can also include both the internal storage unit of the computer device 4 and its external storage device. In this embodiment, the memory 41 is generally used to store the operating system and various application software installed on the computer device 4, such as computer-readable instructions of the speech mining method, etc. In addition, the memory 41 can also be used to temporarily store various types of data that have been output or are to be output.

[0217] The processor 42 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip in some embodiments. The processor 42 is generally used to control the overall operation of the computer device 4. In this embodiment, the processor 42 is used to run the computer-readable instructions stored in the memory 41 or process data, such as computer-readable instructions for running the speech mining method.

[0218] The network interface 43 may include a wireless network interface or a wired network interface. The network interface 43 is generally used to establish a communication connection between the computer device 4 and other electronic devices.

[0219] In the above embodiment, the present application discloses a computer device, which belongs to the field of artificial intelligence technology. The present application performs semantic recognition on historical conversation data, compares the recognition results with the preset speech topics in the speech library, and if a matching topic is found, the conversation data is stored in the corresponding topic; if there is no matching topic, a pre-trained variant model is used to generate a new speech variant. And the variant is scored using a pre-trained evaluation model. If the score is higher than a threshold, the variant is stored in the speech library as a new speech topic, otherwise the variant is deleted. The present application converts unstructured conversation data into structured information through semantic recognition technology, and then compares the recognition results with the preset speech topic to ensure that the conversation data can be effectively archived. If there is no matching topic, a pre-trained variant model is used to generate a new speech, and the quality is evaluated by the evaluation model. The variant with a score higher than the threshold will be stored in the speech library as a new topic, ensuring the dynamic expansion and content quality of the speech library, which not only improves the richness and structure of the speech library, but also ensures the quality and relevance of the newly added speech.

[0220] The present application also provides another embodiment, namely, providing a computer-readable storage medium, wherein the computer-readable storage medium stores computer-readable instructions, and the computer-readable instructions can be executed by at least one processor to enable the at least one processor to perform the steps of the above-mentioned speech mining method.

[0221] In the above embodiment, the present application discloses a computer-readable storage medium, which belongs to the field of artificial intelligence technology. The present application performs semantic recognition on historical conversation data, compares the recognition results with the preset speech topics in the speech library, and if a matching topic is found, stores the conversation data in the corresponding topic; if there is no matching topic, a pre-trained variant model is used to generate a new speech variant. And the variant is scored using a pre-trained evaluation model. If the score is higher than a threshold, the variant is stored in the speech library as a new speech topic, otherwise the variant is deleted. The present application converts unstructured conversation data into structured information through semantic recognition technology, and then compares the recognition results with the preset speech topics to ensure that the conversation data can be effectively archived. If there is no matching topic, a pre-trained variant model is used to generate a new speech, and the quality is evaluated by the evaluation model. The variant with a score higher than the threshold will be stored in the speech library as a new topic, ensuring the dynamic expansion and content quality of the speech library, which not only improves the richness and structure of the speech library, but also ensures the quality and relevance of the newly added speech.

[0222] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present application.

[0223] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments, in which tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.

[0224] Obviously, the embodiments described above are only some embodiments of the present application, rather than all embodiments. The preferred embodiments of the present application are given in the accompanying drawings, but they do not limit the patent scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application is described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions recorded in the aforementioned specific implementation methods, or to perform equivalent replacement of some of the technical features therein. Any equivalent structure made using the contents of the specification and drawings of this application, directly or indirectly used in other related technical fields, is similarly within the scope of patent protection of this application.

Claims

1. A method for mining speech, characterized in that: include: Performing semantic recognition on the historical conversation data to obtain a semantic recognition result of the historical conversation data; Compare the semantic recognition result with the speech topics preset in the speech library; If there is a target speech topic that matches the semantic recognition result, the historical conversation data is stored in the storage space corresponding to the target speech topic in the speech library; If there is no target speech topic that matches the semantic recognition result, performing speech variant prediction on the historical conversation data to obtain a variant speech; Performing speech evaluation on the variant speech and outputting a score for the variant speech; Determine whether the score of the variant speech is greater than a preset score threshold; If the variant speech score is greater than or equal to the score threshold, extracting the variant speech topic, and storing the variant speech topic as a new speech topic in the speech library; If the score of the variant speech is less than the score threshold, the variant speech is deleted.

2. The method for mining speech as claimed in claim 1, characterized in that: Performing semantic recognition on the historical conversation data to obtain a semantic recognition result of the historical conversation data specifically includes: Obtain the speech score corresponding to the historical conversation data to obtain the historical speech score; Filtering the historical conversation data based on the historical speech score and the score threshold to obtain target conversation data; Perform semantic recognition on the target conversation data to obtain the semantic recognition result.

3. The method for mining speech as claimed in claim 2, characterized in that: Performing semantic recognition on the target conversation data to obtain the semantic recognition result specifically includes: Performing word segmentation processing on the target dialogue data to obtain a word segmentation result; Performing part-of-speech tagging on the word segmentation result to obtain a part-of-speech tagging result; Performing named entity recognition on the word segmentation result to obtain a named entity recognition result; Performing semantic analysis on the word segmentation result to obtain a first semantic analysis result; Constructing a knowledge graph of the target conversation data based on the part-of-speech tagging result, the named entity recognition result, and the first semantic analysis result; Semantic inference is performed on the knowledge graph of the target conversation data to obtain the semantic recognition result.

4. The method for mining speech as claimed in claim 1, characterized in that: Predicting speech variants on the historical conversation data to obtain speech variants specifically includes: Pre-training a preset conditional random field model to obtain the speech variant model; Using a speech variant model to extract features from the historical conversation data to obtain a conversation feature vector; Performing sequence labeling on the speech feature vector to obtain a speech variant label sequence; Generating speech variants for the speech variant label sequence to obtain a candidate set of speech variants; The speech in the candidate set of variant speech is screened to obtain the variant speech.

5. The method for mining speech as claimed in claim 4, characterized in that: Screening the words in the candidate set of variant words to obtain the variant words specifically includes: Ranking and scoring the speech in the candidate set of variant speech to obtain a speech ranking score; Normalizing the ranking scores of the speech to obtain probability distribution of variant speech; Decoding the probability distribution of variant speech to obtain a maximum probability variant speech sequence; The variant speech is determined based on a maximum probability variant speech sequence.

6. The method for mining speech as claimed in claim 1, characterized in that: The variant speech is evaluated and a variant speech score is output, which specifically includes: Performing word segmentation processing on the variant speech to obtain the variant speech word segmentation; Filtering emotional keywords from the variant speech segmentation words, and calculating the emotional score of the variant speech based on the emotional keywords; Performing speech logic analysis on the variant speech segmentation words respectively, and calculating the speech logic score of the variant speech based on the variant speech logic analysis result; The variant speech score is calculated based on the sentiment score and the speech logic score.

7. The method for mining speech techniques as claimed in claim 6, characterized in that: Selecting emotional keywords from the variant speech segmentation words, and calculating the emotional score of the variant speech based on the emotional keywords, specifically includes: Using a pre-trained sentiment analysis model to determine the emotional tendency of the variant speech based on the emotional keywords; Calculate the emotion score of the variant speech by using a decision tree algorithm based on the emotion tendency; The speech logic analysis includes lexical analysis, grammatical analysis and semantic analysis. The speech logic analysis is performed on the variant speech segmentation words respectively, and the speech logic score of the variant speech is calculated based on the variant speech logic analysis results, which specifically includes: Performing lexical analysis on the variant speech segmentation to obtain a lexical analysis result; Performing grammatical analysis on the variant speech segmentation to obtain a grammatical analysis result; Performing semantic analysis on the variant speech segmentation to obtain a second semantic analysis result; A speech logic score of the variant speech is calculated based on the lexical analysis result, the grammatical analysis result and the second semantic analysis result.

8. A speech mining device, characterized in that: include: A semantic recognition module, used to perform semantic recognition on the historical conversation data to obtain a semantic recognition result of the historical conversation data; A topic comparison module is used to compare the semantic recognition result with the preset speech topics in the speech library; A speech storage module, used for storing the historical conversation data in a storage space corresponding to the target speech topic in the speech library when there is a target speech topic that matches the semantic recognition result; A speech variant module, used for predicting speech variants on the historical conversation data to obtain speech variants when there is no target speech topic matching the semantic recognition result; A speech scoring module, used for evaluating the speech variants and outputting the speech variant scores; A scoring comparison module is used to determine whether the score of the variant speech is greater than a preset scoring threshold; A theme storage module, used for extracting a variant speech theme when the variant speech score is greater than or equal to the score threshold, and storing the variant speech theme as a new speech theme in the speech library; The speech deletion module is used to delete the variant speech when the score of the variant speech is less than the score threshold.

9. A computer device, characterized in that: It includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the processor executes the computer-readable instructions, the steps of the speech mining method as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the speech mining method as described in any one of claims 1 to 7.

Citation Information

Patent Citations

  • Dialogue corpus generation method for improving quality and diversity of man-machine interaction dialogue corpora

    CN111026884A

  • Intelligent voice robot verbal skill forming method, device and system

    CN111177350A

  • Dialogue generation method and device based on mutual role perception, equipment and medium

    CN113239170A

  • Conversation quality evaluation method and system based on deep learning

    CN114154517A

  • Short text feature extension method based on topic association degree and keyword association

    CN114462392A

Cited By

  • Dialogue data processing method, dialogue method, dialogue device, dialogue equipment, medium and product

    CN121071095A