Intelligent answering method and device for live comment

Abstract

The present invention provides a method and device for intelligent answering of live broadcast comments, which extracts keywords from live broadcast comments posted by users and determines the user's intention based on the keywords, obtains multiple reply data matching the user's intention from a reply database, processes the multiple reply data based on a language generation model, generates intelligent replies to the live broadcast comments, updates the reply database based on the user's live broadcast comments and feedback on the intelligent replies, and implements intelligent answering to viewers in the live broadcast room, which not only improves the audience's experience and satisfaction, but also reduces the host's workload and optimizes the live broadcast content and interaction method.

Description

Technical Field

[0001] The present invention relates to the field of language processing technology, and in particular to an intelligent answering method and device for live broadcast comments. Background Art

[0002] Live streaming e-commerce has achieved remarkable development in recent years and has become an important force in driving the growth of online consumption.

[0003] The live streaming e-commerce industry is undergoing an all-round evolution from transactions to content, from buying and selling to experience, and from competing for sales to building brands. The industry has actively explored quality drive, content innovation, diversified development, and the cultivation of emerging talents to meet the challenges of market competition and meet consumer expectations.

[0004] When there are a large number of people in the live broadcast room, the host will have to answer the audience's questions in a timely manner, which will put him / her under a high workload. At the same time, the host may not be able to quickly reply to every audience's question and comment, resulting in the audience not being able to obtain product information in a timely manner, reducing audience satisfaction. In addition, if the live broadcast is long, the host may temporarily leave the camera, resulting in a cold scene, which is not conducive to the live broadcast performance.

[0005] Therefore, how to provide accurate and intelligent responses in the live broadcast room is an issue that needs to be paid attention to in the development of live e-commerce. Summary of the invention

[0006] The present invention provides an intelligent answering method and device for live broadcast comments, which are used to solve the problems raised in the background technology.

[0007] An intelligent answering method for live broadcast comments, comprising:

[0008] S1: Extract keywords from live broadcast comments posted by users and determine the user's intention based on the keywords;

[0009] S2: Obtain multiple reply data matching the user's intention from the reply database;

[0010] S3: Process multiple reply data based on the language generation model to generate intelligent replies to live comments;

[0011] S4: Based on the user's live broadcast comments and feedback on the smart reply, the reply database is updated.

[0012] Preferably, in S1, keywords are extracted from the live broadcast comments posted by the user, and the user's intention is determined based on the keywords, including:

[0013] Perform word segmentation operations on live broadcast comments posted by users to obtain multiple word segmentation features. Based on the part of speech of the word segmentation features, multiple word segmentation features are screened to obtain target word segmentation features.

[0014] Inputting the target word segmentation features into the intention recognition model, and outputting the user's initial intention;

[0015] Based on the historical intent information, the initial intent is optimized to determine the user's intent.

[0016] Preferably, optimizing the initial intent based on the historical intent information to determine the user's intent includes:

[0017] Matching the initial intention with the historical intention information to obtain multiple historical intentions with a matching degree greater than a preset matching degree;

[0018] Obtain the prediction accuracy of multiple historical intents for historical users and the occurrence frequency of multiple historical intents;

[0019] Determine a first weight based on the matching degree between the initial intent and the historical intent, determine a second weight based on the prediction accuracy of the historical intent for the historical user, determine a third weight based on the occurrence frequency of the historical intent, and determine a comprehensive weighted value of the historical intent based on the first weight, the second weight and the third weight;

[0020] The target historical intention with the largest comprehensive weighted value is selected, and the initial intention is extended based on the target historical intention to determine the user's intention.

[0021] Preferably, in S2, obtaining a plurality of reply data matching the user's intention from the reply database includes:

[0022] Determine product features and product information needs based on user intent;

[0023] Acquire product response data related to product features from a response database;

[0024] A plurality of response data matching the product information requirement is obtained from the product response data.

[0025] Preferably, in S3, multiple reply data are processed based on the language generation model to generate intelligent replies to the live broadcast comments, including:

[0026] Acquire the data volume of the multiple reply data, and divide the multiple reply data based on the reply characteristics to obtain multiple reply data groups;

[0027] Determine the template length based on the data volume, obtain a reply template group of the template length from a template database, determine the template structure based on the reply feature of each reply data group, and select a target reply template that meets the template structure from the reply template group;

[0028] Match the reply data with the target reply template, and configure the reply data at the corresponding position in the target reply template according to the matching result to obtain the initial reply statement;

[0029] Taking reply politeness and reply enthusiasm as evaluation indicators, the initial artificial intelligence model is trained based on historical reply data to obtain a language generation model;

[0030] Inputting the initial reply sentence into the language generation model for language generation and optimization to obtain an optimized reply sentence;

[0031] Verifying the fluency and coherence of the optimized reply sentence to determine whether it meets the fluency and coherence requirements;

[0032] If yes, use the optimized reply statement as an intelligent reply to the live comment;

[0033] Otherwise, based on fluency and coherence as the second evaluation indicator, the parameters of the language generation model are adjusted to obtain a new language generation model, until the latest optimized reply sentence generated by the new language generation model meets the requirements of fluency and coherence, and the latest optimized reply sentence is used as an intelligent reply to the live broadcast comments.

[0034] Preferably, the step of inputting the initial reply sentence into the language generation model for language generation and optimization to obtain an optimized reply sentence includes:

[0035] Inputting the initial reply sentence into the language generation model to modify the reply vector order and insert politeness expressions and modal particles to obtain an intermediate reply sentence;

[0036] The intermediate reply sentence is cyclically input into the language generation model until the evaluation index is met, and an optimized reply sentence is output.

[0037] Preferably, in S4, based on the user's live broadcast comments and feedback on the smart reply, the reply database is updated, including:

[0038] Based on the similarity between the user's live broadcast comments and historical live broadcast comments, the live broadcast comments are divided into common comments and novel comments, a first update value is set for the common comments, and a second update value is set for the novel comments;

[0039] Based on the feedback of the smart reply to the ordinary comment, obtain the ordinary feedback result, determine the feedback evaluation value of the ordinary feedback result, determine the first update ratio of the smart reply to the ordinary comment based on the feedback evaluation value and the first update value, and update the smart reply to the ordinary comment according to the first update ratio.

[0040] Based on the feedback of the intelligent reply to the novel comment, a novel feedback result is obtained, a novel feedback evaluation value of the novel feedback result is determined, and whether the novel feedback evaluation value is greater than a preset evaluation value;

[0041] If so, determining a second update ratio of the smart reply to the novel comment based on the novel feedback evaluation value and the second update value, and updating the smart reply to the novel comment to the reply database according to the second update ratio;

[0042] Otherwise, manual reply data and manual reply feedback to the novel comments are collected based on a large database, and a selection ratio of the manual reply data is determined based on the degree of praise for the reply feedback. Target manual reply data is obtained from the manual reply data according to the selection ratio, and a target intelligent reply with a preset update ratio is selected from the intelligent replies to the novel comments. The reply database is updated based on the target manual reply data and the target intelligent reply.

[0043] A response device for an intelligent response method of live commentary, comprising:

[0044] The intention determination module is used to extract keywords from the live broadcast comments posted by users and determine the user's intention based on the keywords;

[0045] A reply determination module, used to obtain a plurality of reply data matching the user's intention from a reply database;

[0046] A response determination module is used to process multiple response data based on a language generation model to generate intelligent responses to live broadcast comments;

[0047] The data updating module is used to update the reply database based on the user's live broadcast comments and feedback on the smart reply.

[0048] Preferably, the intention determination module includes:

[0049] A word segmentation determination unit is used to perform a word segmentation operation on the live broadcast comments posted by the user to obtain multiple word segmentation features, and to screen the multiple word segmentation features based on the parts of speech of the word segmentation features to obtain a target word segmentation feature;

[0050] A model recognition unit, used to input the target word segmentation features into an intention recognition model and output the user's initial intention;

[0051] The intention optimization unit is used to optimize the initial intention based on the historical intention information to determine the user's intention.

[0052] Preferably, the reply determination module includes:

[0053] A demand determination unit, used to determine product features and product information requirements based on the user's intention;

[0054] A first matching unit, used for obtaining product response data related to product features in a response database;

[0055] The second matching unit is used to obtain multiple reply data matching the commodity information requirement from the commodity reply data.

[0056] Compared with the prior art, the present invention has achieved the following beneficial effects:

[0057] By performing word segmentation operations on the live comments posted by users, multiple word segmentation features are obtained. Based on the parts of speech of the word segmentation features, the multiple word segmentation features are screened to obtain target word segmentation features. The target word segmentation features are input into the intent recognition model, and the user's initial intent is output. Based on historical intent information, the initial intent is optimized to determine the user's intent, providing a basis for designing intelligent replies to live comments.

[0058] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures specifically pointed out in this application document.

[0059] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0061] Figure 1 This is a flow chart of an intelligent response method for live commentary in an embodiment of the present invention;

[0062] Figure 2 A flowchart for determining user intent in an embodiment of the present invention;

[0063] Figure 3 The present invention is a structural diagram of an intelligent response device for live commentary in an embodiment of the present invention. DETAILED DESCRIPTION

[0064] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0065] Embodiment 1:

[0066] The embodiment of the present invention provides an intelligent response method for live commentary, such as Figure 1 As shown, including:

[0067] S1: Extract keywords from live broadcast comments posted by users and determine the user's intention based on the keywords;

[0068] S2: Obtain multiple reply data matching the user's intention from the reply database;

[0069] S3: Process multiple reply data based on the language generation model to generate intelligent replies to live comments;

[0070] S4: Based on the user's live broadcast comments and feedback on the smart reply, the reply database is updated.

[0071] In this embodiment, the reply database is a reply data segment composed of commodity information.

[0072] In this embodiment, the plurality of reply data includes information about the product, etc., and needs to be further improved to become a complete reply.

[0073] In this embodiment, the update of the reply database is the update of the product information introduction, or the language art, such as the update of the network hot meme.

[0074] The beneficial effects of the above design scheme are: by extracting keywords from the live comments posted by users, and determining the user's intention based on the keywords, multiple reply data matching the user's intention are obtained from the reply database, and the multiple reply data are processed based on the language generation model to generate intelligent replies to the live comments. Based on the user's live comments and feedback on the intelligent replies, the reply database is updated to achieve intelligent responses to the audience in the live broadcast room, which not only improves the audience's experience and satisfaction, but also reduces the host's workload and optimizes the live content and interaction methods.

[0075] Embodiment 2:

[0076] Based on Example 1, the embodiment of the present invention provides an intelligent response method for live broadcast comments. In S1, keywords are extracted from the live broadcast comments posted by the user, and the user's intention is determined based on the keywords, including:

[0077] Perform word segmentation operations on live broadcast comments posted by users to obtain multiple word segmentation features. Based on the part of speech of the word segmentation features, multiple word segmentation features are screened to obtain target word segmentation features.

[0078] Inputting the target word segmentation features into the intention recognition model, and outputting the user's initial intention;

[0079] Based on the historical intent information, the initial intent is optimized to determine the user's intent.

[0080] The beneficial effects of the above design scheme are: by performing word segmentation operations on the live broadcast comments posted by users, multiple word segmentation features are obtained, and based on the parts of speech of the word segmentation features, the multiple word segmentation features are screened to obtain target word segmentation features, and the target word segmentation features are input into the intention recognition model, and the user's initial intention is output. Based on historical intention information, the initial intention is optimized to determine the user's intention, thereby providing a basis for designing intelligent replies to live broadcast comments.

[0081] Embodiment 3:

[0082] Based on Example 2, the present invention provides an intelligent response method for live commentary, which is characterized in that: Figure 2 As shown, the initial intent is optimized based on the historical intent information to determine the user's intent, including:

[0083] Matching the initial intention with the historical intention information to obtain multiple historical intentions with a matching degree greater than a preset matching degree;

[0084] Obtain the prediction accuracy of multiple historical intents for historical users and the occurrence frequency of multiple historical intents;

[0085] Determine a first weight based on the matching degree between the initial intent and the historical intent, determine a second weight based on the prediction accuracy of the historical intent for the historical user, determine a third weight based on the occurrence frequency of the historical intent, and determine a comprehensive weighted value of the historical intent based on the first weight, the second weight and the third weight;

[0086] The target historical intention with the largest comprehensive weighted value is selected, and the initial intention is extended based on the target historical intention to determine the user's intention.

[0087] In this embodiment, extending the initial intent can enrich the information obtained by the user. For example, if the user wants to know the size of a piece of clothing, while replying with the size of the clothing, the height and weight of the person corresponding to the size range can be determined at the same time, bringing a good experience to the user.

[0088] The beneficial effect of the above design scheme is: by optimizing the initial intent based on the historical intent information, the user's intent is determined, providing a basis for designing intelligent replies to live broadcast comments.

[0089] Embodiment 4:

[0090] Based on Example 1, the embodiment of the present invention provides an intelligent response method for live commentary. In S2, multiple reply data matching the user's intention are obtained from the reply database, including:

[0091] Determine product features and product information needs based on user intent;

[0092] Acquire product response data related to product features from a response database;

[0093] A plurality of response data matching the product information requirement is obtained from the product response data.

[0094] The beneficial effects of the above design scheme are: by determining product characteristics and product information requirements based on user intentions; obtaining product response data related to product characteristics from the response database; obtaining multiple response data matching product information requirements from the product response data, accurate information provision for user needs is achieved, providing an accurate information basis for the generation of further intelligent responses.

[0095] Embodiment 5:

[0096] Based on Example 1, the embodiment of the present invention provides an intelligent response method for live broadcast comments. In S3, multiple reply data are processed based on a language generation model to generate an intelligent reply to the live broadcast comments, including:

[0097] Acquire the data volume of the multiple reply data, and divide the multiple reply data based on the reply characteristics to obtain multiple reply data groups;

[0098] Determine the template length based on the data volume, obtain a reply template group of the template length from a template database, determine the template structure based on the reply feature of each reply data group, and select a target reply template that meets the template structure from the reply template group;

[0099] Match the reply data with the target reply template, and configure the reply data at the corresponding position in the target reply template according to the matching result to obtain the initial reply statement;

[0100] Taking reply politeness and reply enthusiasm as evaluation indicators, the initial artificial intelligence model is trained based on historical reply data to obtain a language generation model;

[0101] Inputting the initial reply sentence into the language generation model for language generation and optimization to obtain an optimized reply sentence;

[0102] Verifying the fluency and coherence of the optimized reply sentence to determine whether it meets the fluency and coherence requirements;

[0103] If yes, use the optimized reply statement as an intelligent reply to the live comment;

[0104] Otherwise, based on fluency and coherence as the second evaluation indicator, the parameters of the language generation model are adjusted to obtain a new language generation model, until the latest optimized reply sentence generated by the new language generation model meets the requirements of fluency and coherence, and the latest optimized reply sentence is used as an intelligent reply to the live broadcast comments.

[0105] The beneficial effects of the above design scheme are: first, by matching the reply data with the target reply template, configuring the reply data at the corresponding position in the target reply template according to the matching result, obtaining the initial reply sentence, and realizing the acquisition of the intelligent response sentence, and then based on the language generation model, optimizing the sentence from four factors: reply politeness and reply enthusiasm, fluency and coherence, to ensure the accuracy of the information obtained in the intelligent reply to the live broadcast comment, and to obtain information more intuitively and satisfactorily, so as to realize the intelligent response to the audience in the live broadcast room, which not only improves the audience experience and satisfaction, but also reduces the workload of the anchor, and optimizes the live broadcast content and interaction methods.

[0106] Embodiment 6:

[0107] Based on Example 5, the embodiment of the present invention provides an intelligent response method for live commentary, wherein the initial reply sentence is input into the language generation model for language generation and optimization to obtain an optimized reply sentence, including:

[0108] Inputting the initial reply sentence into the language generation model to modify the reply vector order and insert politeness expressions and modal particles to obtain an intermediate reply sentence;

[0109] The intermediate reply sentence is cyclically input into the language generation model until the evaluation index is met, and an optimized reply sentence is output.

[0110] The beneficial effect of the above design scheme is: by inputting the initial reply sentence into the language generation model to correct the reply vector sequence and insert polite words and modal particles, an intermediate reply sentence is obtained;

[0111] The intermediate reply sentence is cyclically input into the language generation model until the evaluation index is met, and an optimized reply sentence is output to ensure that the obtained optimized reply sentence not only answers the user's question but also demonstrates a warm and polite intelligent communication atmosphere, bringing a good experience to the user.

[0112] Embodiment 7:

[0113] Based on Example 1, the embodiment of the present invention provides an intelligent response method for live broadcast comments. In S4, based on the user's live broadcast comments and feedback on the intelligent reply, the reply database is updated, including:

[0114] Based on the similarity between the user's live broadcast comments and historical live broadcast comments, the live broadcast comments are divided into common comments and novel comments, a first update value is set for the common comments, and a second update value is set for the novel comments;

[0115] Based on the feedback of the smart reply to the ordinary comment, obtain the ordinary feedback result, determine the feedback evaluation value of the ordinary feedback result, determine the first update ratio of the smart reply to the ordinary comment based on the feedback evaluation value and the first update value, and update the smart reply to the ordinary comment according to the first update ratio.

[0116] Based on the feedback of the intelligent reply to the novel comment, a novel feedback result is obtained, a novel feedback evaluation value of the novel feedback result is determined, and whether the novel feedback evaluation value is greater than a preset evaluation value;

[0117] If so, determining a second update ratio of the smart reply to the novel comment based on the novel feedback evaluation value and the second update value, and updating the smart reply to the novel comment to the reply database according to the second update ratio;

[0118] Otherwise, manual reply data and manual reply feedback to the novel comments are collected based on a large database, and a selection ratio of the manual reply data is determined based on the degree of praise for the reply feedback. Target manual reply data is obtained from the manual reply data according to the selection ratio, and a target intelligent reply with a preset update ratio is selected from the intelligent replies to the novel comments. The reply database is updated based on the target manual reply data and the target intelligent reply.

[0119] In this embodiment, if the similarity is greater than a threshold, it is determined to be a novel comment; otherwise, it is determined to be a common comment.

[0120] In this embodiment, the first update value is smaller than the second update value.

[0121] In this embodiment, the preset update ratio is smaller than the second update ratio.

[0122] The beneficial effect of the above design scheme is: by considering the novelty of the live broadcast comments and feedback on the intelligent replies based on the users, reasonable database updates are carried out, ensuring that the updated reply database keeps pace with the trend of the times and the development of information, and providing an accurate and rich database for intelligent answers.

[0123] Embodiment 8:

[0124] Based on Example 1, the present invention provides a response device for an intelligent response method for live commentary, such as Figure 3 As shown, including:

[0125] The intention determination module is used to extract keywords from the live broadcast comments posted by users and determine the user's intention based on the keywords;

[0126] A reply determination module, used to obtain a plurality of reply data matching the user's intention from a reply database;

[0127] A response determination module is used to process multiple response data based on a language generation model to generate intelligent responses to live broadcast comments;

[0128] The data updating module is used to update the reply database based on the user's live broadcast comments and feedback on the smart reply.

[0129] In this embodiment, the reply database is a reply data segment composed of commodity information.

[0130] In this embodiment, the plurality of reply data includes information about the product, etc., and needs to be further improved to become a complete reply.

[0131] In this embodiment, the update of the reply database is the update of the product information introduction, or the language art, such as the update of the network hot meme.

[0132] The beneficial effects of the above design scheme are: by extracting keywords from the live comments posted by users, and determining the user's intention based on the keywords, multiple reply data matching the user's intention are obtained from the reply database, and the multiple reply data are processed based on the language generation model to generate intelligent replies to the live comments. Based on the user's live comments and feedback on the intelligent replies, the reply database is updated to achieve intelligent responses to the audience in the live broadcast room, which not only improves the audience's experience and satisfaction, but also reduces the host's workload and optimizes the live content and interaction methods.

[0133] Embodiment 9:

[0134] Based on Example 8, the embodiment of the present invention provides an intelligent response device for live commentary, wherein the intention determination module includes:

[0135] A word segmentation determination unit is used to perform a word segmentation operation on the live broadcast comments posted by the user to obtain multiple word segmentation features, and to screen the multiple word segmentation features based on the parts of speech of the word segmentation features to obtain a target word segmentation feature;

[0136] A model recognition unit, used to input the target word segmentation features into an intention recognition model and output the user's initial intention;

[0137] The intention optimization unit is used to optimize the initial intention based on the historical intention information to determine the user's intention.

[0138] The beneficial effects of the above design scheme are: by performing word segmentation operations on the live broadcast comments posted by users, multiple word segmentation features are obtained, and based on the parts of speech of the word segmentation features, the multiple word segmentation features are screened to obtain target word segmentation features, and the target word segmentation features are input into the intention recognition model, and the user's initial intention is output. Based on historical intention information, the initial intention is optimized to determine the user's intention, thereby providing a basis for designing intelligent replies to live broadcast comments.

[0139] Embodiment 10:

[0140] Based on Example 8, the embodiment of the present invention provides an intelligent response device for live commenting, wherein the reply determination module includes:

[0141] A demand determination unit, used to determine product features and product information requirements based on the user's intention;

[0142] A first matching unit, used for obtaining product response data related to product features in a response database;

[0143] The second matching unit is used to obtain multiple reply data matching the commodity information requirement from the commodity reply data.

[0144] The beneficial effects of the above design scheme are: by determining product characteristics and product information requirements based on user intentions; obtaining product response data related to product characteristics from the response database; obtaining multiple response data matching product information requirements from the product response data, accurate information provision for user needs is achieved, providing an accurate information basis for the generation of further intelligent responses.

[0145] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention belong to the scope of this application document and its equivalent technology, the present invention is also intended to include these changes and variations.

Claims

1. An intelligent answering method for live commentary, characterized in that: include: S1: Extract keywords from live broadcast comments posted by users and determine the user's intention based on the keywords; S2: Obtain multiple reply data matching the user's intention from the reply database; S3: Process multiple reply data based on the language generation model to generate intelligent replies to live comments; S4: Based on the user's live broadcast comments and feedback on the smart reply, the reply database is updated; In S3, multiple reply data are processed based on the language generation model to generate intelligent replies to live comments, including: Acquire the data volume of the multiple reply data, and divide the multiple reply data based on the reply characteristics to obtain multiple reply data groups; Determine the template length based on the data volume, obtain a reply template group of the template length from a template database, determine the template structure based on the reply feature of each reply data group, and select a target reply template that meets the template structure from the reply template group; Match the reply data with the target reply template, and configure the reply data at the corresponding position in the target reply template according to the matching result to obtain the initial reply statement; Taking reply politeness and reply enthusiasm as evaluation indicators, the initial artificial intelligence model is trained based on historical reply data to obtain a language generation model; Inputting the initial reply sentence into the language generation model for language generation and optimization to obtain an optimized reply sentence; Verifying the fluency and coherence of the optimized reply sentence to determine whether it meets the fluency and coherence requirements; If yes, use the optimized reply statement as an intelligent reply to the live comment; Otherwise, based on fluency and coherence as the second evaluation indicator, the parameters of the language generation model are adjusted to obtain a new language generation model, until the latest optimized reply sentence generated by the new language generation model meets the requirements of fluency and coherence, and the latest optimized reply sentence is used as an intelligent reply to the live broadcast comments.

2. According to claim 1, the intelligent answering method for live commentary is characterized in that: In S1, keywords are extracted from the live broadcast comments posted by the user, and the user's intention is determined based on the keywords, including: Perform word segmentation operations on live broadcast comments posted by users to obtain multiple word segmentation features. Based on the part of speech of the word segmentation features, multiple word segmentation features are screened to obtain target word segmentation features. Inputting the target word segmentation features into the intention recognition model, and outputting the user's initial intention; Based on the historical intent information, the initial intent is optimized to determine the user's intent.

3. The intelligent answering method for live commentary according to claim 2 is characterized in that: The optimizing the initial intent based on the historical intent information to determine the user's intent includes: Matching the initial intention with the historical intention information to obtain multiple historical intentions with a matching degree greater than a preset matching degree; Obtain the prediction accuracy of multiple historical intents for historical users and the occurrence frequency of multiple historical intents; Determine a first weight based on the matching degree between the initial intent and the historical intent, determine a second weight based on the prediction accuracy of the historical intent for the historical user, determine a third weight based on the occurrence frequency of the historical intent, and determine a comprehensive weighted value of the historical intent based on the first weight, the second weight and the third weight; The target historical intention with the largest comprehensive weighted value is selected, and the initial intention is extended based on the target historical intention to determine the user's intention.

4. The intelligent response method for live commentary according to claim 1, characterized in that: In S2, a plurality of reply data matching the user's intention is obtained from the reply database, including: Determine product features and product information needs based on user intent; Acquire product response data related to product features from a response database; A plurality of response data matching the product information requirement is obtained from the product response data.

5. The intelligent answering method for live commentary according to claim 1 is characterized in that: The step of inputting the initial reply sentence into the language generation model for language generation and optimization to obtain an optimized reply sentence includes: Inputting the initial reply sentence into the language generation model to modify the reply vector order and insert politeness expressions and modal particles to obtain an intermediate reply sentence; The intermediate reply sentence is cyclically input into the language generation model until the evaluation index is met, and an optimized reply sentence is output.

6. The intelligent response method for live commentary according to claim 1, characterized in that: In S4, based on the user's live broadcast comments and feedback on the smart reply, the reply database is updated, including: Based on the similarity between the user's live broadcast comments and historical live broadcast comments, the live broadcast comments are divided into common comments and novel comments, a first update value is set for the common comments, and a second update value is set for the novel comments; Based on the feedback of the smart reply to the ordinary comment, obtain the ordinary feedback result, determine the feedback evaluation value of the ordinary feedback result, determine the first update ratio of the smart reply to the ordinary comment based on the feedback evaluation value and the first update value, and update the smart reply to the ordinary comment according to the first update ratio. Based on the feedback of the intelligent reply to the novel comment, a novel feedback result is obtained, a novel feedback evaluation value of the novel feedback result is determined, and whether the novel feedback evaluation value is greater than a preset evaluation value; If so, determining a second update ratio of the smart reply to the novel comment based on the novel feedback evaluation value and the second update value, and updating the smart reply to the novel comment to the reply database according to the second update ratio; Otherwise, manual reply data and manual reply feedback to the novel comments are collected based on a large database, and a selection ratio of the manual reply data is determined based on the degree of praise for the reply feedback. Target manual reply data is obtained from the manual reply data according to the selection ratio, and a target intelligent reply with a preset update ratio is selected from the intelligent replies to the novel comments. The reply database is updated based on the target manual reply data and the target intelligent reply.

7. An intelligent response device for live commentary, used to implement the steps of the intelligent response method according to claim 1, characterized in that: include: The intention determination module is used to extract keywords from the live broadcast comments posted by users and determine the user's intention based on the keywords; A reply determination module, used to obtain a plurality of reply data matching the user's intention from a reply database; A response determination module is used to process multiple response data based on a language generation model to generate intelligent responses to live broadcast comments; The data updating module is used to update the reply database based on the user's live broadcast comments and feedback on the smart reply.

8. The intelligent response device for live commentary according to claim 7, characterized in that: The intention determination module includes: A word segmentation determination unit is used to perform a word segmentation operation on the live broadcast comments posted by the user to obtain multiple word segmentation features, and to screen the multiple word segmentation features based on the parts of speech of the word segmentation features to obtain a target word segmentation feature; A model recognition unit, used to input the target word segmentation features into an intention recognition model and output the user's initial intention; The intention optimization unit is used to optimize the initial intention based on the historical intention information to determine the user's intention.

9. The intelligent response device for live commentary according to claim 7, characterized in that: The reply determination module includes: A demand determination unit, used to determine product features and product information requirements based on the user's intention; A first matching unit, used for obtaining product response data related to product features in a response database; The second matching unit is used to obtain multiple reply data matching the commodity information requirement from the commodity reply data.

Images