Session recommendation method and system and storage medium

By constructing a multi-level session recommendation dataset and using neural classifiers and Transformers encoders, the problem of ignoring the differences in user interaction intentions and communication styles in traditional methods is solved, and high-accuracy recommendations of interactive learning sessions are achieved.

CN119003749AActive Publication Date: 2024-11-22GUANGZHOU UNIVERSITY
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202410943776.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-15
Publication Date
2024-11-22
Estimated Expiration
2044-07-15

AI Technical Summary

Technical Problem

The traditional interactive learning session recommendation method ignores user interaction intention modeling and does not fully consider the differences in user communication styles, resulting in insufficient accuracy and accuracy of recommendation results.

Method used

By constructing session recommendation data sets, including course recommendation data sets, session context interaction intention sequence recommendation data sets, session context source/response recommendation data sets and user history session recommendation data sets, using technologies such as neural classifiers and Transformers encoders. , generate multi-level recommendation results to improve recommendation accuracy.

Benefits of technology

It realizes high-accuracy recommendations for the development direction and content of interactive learning sessions, improving the effectiveness and personalization of recommendation results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119003749A_ABST
    Figure CN119003749A_ABST
Patent Text Reader

Abstract

The invention discloses a session recommendation method and system and a storage medium, and the method comprises the steps: constructing session recommendation data sets according to a corpus, including a course recommendation data set, a session context interaction intention sequence recommendation data set, a session context source language / response pair recommendation data set and a user historical session recommendation data set; generating a first recommendation result; according to the first recommendation result and the session context interaction intention sequence recommendation data set, generating a second recommendation result; generating a third recommendation result according to the second recommendation result and the session context source language / response pair recommendation data set; generating a fourth recommendation result according to the third recommendation result and the user historical session recommendation data set; and generating a session recommendation report according to the first recommendation result, the second recommendation result, the third recommendation result and the fourth recommendation result. The method is high in accuracy, can effectively recommend the development direction and content of the interactive learning session, and can be widely applied to the technical field of computers.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a conversation recommendation method, system and storage medium. Background Art

[0002] Conducting interactive learning session recommendations plays a key role in optimizing learning experience, promoting knowledge sharing, and improving teaching effectiveness. Through session direction and content recommendations, it helps guide interactive learning discussions and exchanges towards higher quality and more constructive directions, thereby improving learners' participation and learning effectiveness.

[0003] At present, there are the following problems in interactive learning conversation recommendation: First, traditional methods mainly focus on conversation semantic similarity, while ignoring the modeling of users' interactive intentions, resulting in incomplete and inaccurate analysis of the direction of conversation development. Second, traditional methods usually use generalized models to recommend conversation content, without fully considering the style differences between different users. Due to differences in individual communication styles, it is necessary to identify individual communication style characteristics to improve conversation and interaction effects and efficiency. Third, traditional methods mostly focus on the examination of a single factor, such as semantic similarity, and rarely take multiple factors into consideration. There is a lack of comprehensive and in-depth analysis, which often makes the recommendation results lack precision and accuracy. Summary of the invention

[0004] The main purpose of the embodiments of the present invention is to provide a highly accurate conversation recommendation method, system and storage medium, which can effectively recommend the development direction and content of interactive learning conversations.

[0005] To achieve the above object, an embodiment of the present invention provides a conversation recommendation method, comprising the following steps:

[0006] Constructing a conversation recommendation dataset based on the corpus; wherein the conversation recommendation dataset includes a course recommendation dataset, a conversation context interaction intention sequence recommendation dataset, a conversation context source language / response pair recommendation dataset, and a user history conversation recommendation dataset;

[0007] Generate a first recommendation result according to the course recommendation dataset;

[0008] Generate a second recommendation result according to the first recommendation result and the session context interaction intention sequence recommendation data set;

[0009] generating a third recommendation result according to the second recommendation result and the conversation context source language / response pair recommendation data set;

[0010] Generate a fourth recommendation result according to the third recommendation result and the user historical session recommendation data set;

[0011] A session recommendation report is generated according to the first recommendation result, the second recommendation result, the third recommendation result, and the fourth recommendation result.

[0012] In some embodiments, constructing a conversation recommendation dataset based on a corpus includes:

[0013] Constructing the course recommendation dataset according to the course meta information and the course description and catalog content document information;

[0014] According to the conversation context sequence, construct the conversation context interaction intention sequence recommendation dataset;

[0015] According to the conversation context source / response pair data, a conversation context source / response pair recommendation dataset is constructed; wherein the conversation context source / response pair recommendation dataset includes a conversation context source / response pair interaction intention representation vector and a conversation context source / response pair self-attention / cross-attention matching matrix; wherein the conversation context source / response pair interaction intention representation vector is obtained by characterizing and classifying the conversation context source / response pair interaction intention through a neural classifier; wherein the conversation context source / response pair self-attention / cross-attention matching matrix is ​​obtained by encoding and characterizing the semantic dependency information in the conversation context source / response pair data through a Transformers encoder;

[0016] A user history conversation recommendation data set is constructed according to the user history conversation text data; wherein the user history conversation recommendation data set includes a personalized language statistical feature data set and a user conversation language feature data set based on a psychology dictionary.

[0017] In some embodiments, the step of constructing a conversation recommendation dataset based on a corpus further includes the step of constructing a tokenized conversation text dataset based on preprocessing of the user's historical conversation text data, which is specifically:

[0018] The tokenized conversation texts are integrated and summarized to obtain a tokenized conversation text dataset, which is used as input data in the process of extracting conversation language statistical feature data; wherein the tokenized conversation texts are composed of different conversation text elements, each conversation text element is separated by a space, and each conversation text element is a complete token;

[0019] Constructing a personalized language statistical feature dataset based on the tokenized conversation text dataset; specifically: extracting 6 personalized language statistical features of the user from the tokenized conversation text dataset, and forming the personalized language statistical feature dataset by integrating and aggregating; constructing a user conversation language feature dataset based on a psychology dictionary based on a psychology dictionary by using a linguistic survey and word counting tool or a Wenxin Chinese psychological analysis system; extracting the user conversation language features based on the psychology dictionary from the tokenized conversation text, and forming the user conversation language feature dataset based on the psychology dictionary by integrating and aggregating;

[0020] Among them, the psychological dictionary is intended to capture the social and psychological states of users.

[0021] In some embodiments, generating a first recommendation result according to the course recommendation dataset comprises the following steps:

[0022] Acquire a first recommendation list from the course recommendation dataset according to a first preset rule;

[0023] Constructing a feature vector of key knowledge points of course description and catalog content documents according to the first recommendation list;

[0024] Calculate the course similarity based on the feature vectors of the course description and the key knowledge points of the catalog content document; wherein the course similarity is the similarity between the target feature vector of the target course description and the key knowledge points of the catalog content document and the candidate feature vector of the key knowledge points of the course description and content document of any candidate course set in the first recommendation list;

[0025] Sorting the candidate courses in the first recommendation list according to the course similarity scores, and obtaining the candidate courses whose similarity scores are greater than a user-preset threshold as the course set of the first recommendation result;

[0026] The candidate courses in the course set of the first recommendation result are sorted according to the similarity scores to generate the first recommendation result.

[0027] In some embodiments, generating a second recommendation result according to the first recommendation result and the session context interaction intention sequence recommendation dataset comprises the following steps:

[0028] Taking the session context sequence associated with the learning user related to the candidate courses in the course set of the first recommendation result as the second recommendation list;

[0029] The conversion frequency from one interactive intention to another interactive intention in the conversation context sequence is calculated according to the adjacency matrix, and the residual coefficient of the interactive intention conversion in the conversation context sequence is calculated; wherein the residual coefficient is obtained by obtaining the difference between the residual count and the expected count of the interactive intention conversion in the conversation context sequence, and then obtaining the residual coefficient according to the quotient of the difference and a preset error estimate;

[0030] According to the comparison between the residual value coefficient and the significance level threshold, identifying the interaction intention conversion mode significantly associated with the target conversation source language;

[0031] Acquire, from the second recommendation list, a conversation context sequence whose residual value coefficient of the interaction intention conversion mode is greater than a significance level threshold;

[0032] The conversation context sequences in the second recommendation list are sorted from high to low according to the significance level of the interaction intention conversion pattern to generate a second recommendation result.

[0033] In some embodiments, the conversation context source / response pair recommendation dataset includes conversation context source / response pairs, self-attention / cross-attention matching matrix, and conversation context source / response pair interaction intention representation vector; generating a third recommendation result according to the second recommendation result and the conversation context source / response pair recommendation dataset includes the following steps:

[0034] taking the conversation context source word / response pairs associated in the conversation context sequence of the second recommendation result as a third recommendation list;

[0035] Using an attention mechanism based on intent perception, weighted matching is performed on the conversation context source / response pair and the self-attention / cross-attention matching matrix;

[0036] Extracting the final representation through a convolutional neural network to generate a ranking score for the source / response pair in the conversation context;

[0037] Filtering out the conversation context source word / response pair whose conversation context source word / response pair ranking score is greater than the conversation context source word / response pair ranking score threshold from the second recommendation results as a third recommendation list set;

[0038] The conversation context source word / response pairs associated in the conversation context sequence in the second recommendation result are sorted according to the sorting scores of the conversation context source word / response pairs in the third recommendation list set to generate a third recommendation result.

[0039] In some embodiments, the user history conversation recommendation dataset includes a personalized language statistical feature dataset and a user conversation language feature dataset based on a psychological dictionary; generating a fourth recommendation result according to the third recommendation result and the user history conversation recommendation dataset includes the following steps:

[0040] A fourth recommendation list is formed based on the user history conversation recommendation data set involved in the conversation context source language / response pairs screened out according to the third recommendation result;

[0041] Constructing a personalized conversation language feature sequence according to the personalized language statistical feature data set and the user conversation language feature data set based on the psychology dictionary; wherein the personalized conversation language feature sequence includes the personalized language statistical feature vector and the user conversation language feature vector based on the psychology dictionary;

[0042] Calculate a first similarity based on the personalized language statistical feature vector; wherein the first similarity is the cosine similarity between the personalized language statistical feature vector of the target user and the personalized language statistical feature vector of any user in the fourth recommendation list associated user set; the fourth recommendation list associated user set is a set of users other than the target user in the fourth recommendation list associated users;

[0043] Calculate a second similarity based on the user conversation language feature vector based on the psychology dictionary; wherein the second similarity is a cosine similarity between the user conversation language feature vector based on the psychology dictionary and the user conversation language feature vector based on the psychology dictionary of any user in the user set associated with the fourth recommendation list;

[0044] Calculating a comprehensive similarity according to the first similarity and the second similarity;

[0045] The user whose comprehensive similarity is greater than the comprehensive similarity threshold is obtained from the user set associated with the fourth recommendation list, and the response recommendation set of the conversation context source / response pair recommendation data set associated with the user is taken as the fourth recommendation result.

[0046] Another aspect of the embodiment of the present invention further provides a conversation recommendation system, including:

[0047] The first module is used to construct a conversation recommendation dataset based on the corpus; wherein the conversation recommendation dataset includes a course recommendation dataset, a conversation context interaction intention sequence recommendation dataset, a conversation context source language / response pair recommendation dataset and a user history conversation recommendation dataset;

[0048] A second module is used to generate a first recommendation result according to the course recommendation dataset;

[0049] A third module is used to generate a second recommendation result based on the first recommendation result and the session context interaction intention sequence recommendation data set;

[0050] A fourth module is used to generate a third recommendation result according to the second recommendation result and the conversation context source language / response pair recommendation data set;

[0051] A fifth module, configured to generate a fourth recommendation result based on the third recommendation result and the user historical session recommendation data set;

[0052] The sixth module is used to generate a session recommendation report according to the first recommendation result, the second recommendation result, the third recommendation result and the fourth recommendation result.

[0053] Another aspect of an embodiment of the present invention further provides an electronic device, including a processor and a memory;

[0054] The memory is used to store programs;

[0055] The processor executes the program to implement the method described above.

[0056] Another aspect of the embodiments of the present invention further provides a computer-readable storage medium, wherein the storage medium stores a program, and the program is executed by a processor to implement the method described above.

[0057] The embodiment of the present invention also discloses a computer program product or a computer program, which includes a computer instruction stored in a computer-readable storage medium. A processor of a computer device can read the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the above method.

[0058] The embodiments of the present invention include at least the following beneficial effects: the present invention provides a conversation recommendation method, system and storage medium, the present invention constructs a conversation recommendation data set according to a corpus; generates a first recommendation result according to the course recommendation data set; generates a second recommendation result according to the first recommendation result and the conversation context interactive intention sequence recommendation data set; generates a third recommendation result according to the second recommendation result and the conversation context source language / response pair recommendation data set; generates a fourth recommendation result according to the third recommendation result and the user history conversation recommendation data set; generates a conversation recommendation report according to the first recommendation result, the second recommendation result, the third recommendation result and the fourth recommendation result. The present invention has high accuracy and can achieve effective recommendation of the development direction and content of interactive learning conversations. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1is a flow chart of the overall steps provided by an embodiment of the present invention;

[0060] Figure 2 A schematic diagram of a session context sequence provided by an embodiment of the present invention;

[0061] Figure 3 is a schematic diagram of the hardware structure of an electronic device provided by an embodiment of the present invention;

[0062] Figure 4 A schematic diagram of a course similarity EMR algorithm provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the embodiments of the present invention. They are only examples of devices and methods consistent with some aspects of the embodiments of the present invention as detailed in the attached claims.

[0064] It is understood that the terms "first", "second", etc. used in the present invention may be used herein to describe various concepts, but unless otherwise specified, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of the present invention, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the words "if" and "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0065] The terms "at least one", "multiple", "each", "any", etc. used in the present invention, at least one includes one, two or more, multiple includes two or more, each refers to each of the corresponding multiple, and any refers to any one of the multiple.

[0066] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which the present invention belongs. The terms used herein are only for the purpose of describing the embodiments of the present invention and are not intended to limit the present invention.

[0067] The conversation recommendation method, system and storage medium provided by the embodiments of the present invention relate to the field of computer technology. The conversation recommendation method provided by the embodiments of the present invention can be applied to a terminal, a server, or a software running in a terminal or a server. In some embodiments, the terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, and a car terminal, etc., but is not limited to this; the server side can be configured as an independent physical server, or a server cluster or distributed system composed of multiple physical servers, and can also be configured as 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, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network; the software can be an application that implements the conversation recommendation method, etc., but is not limited to the above forms.

[0068] The present invention 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 electronic devices, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc. The present invention 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 invention 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.

[0069] Traditional recommendation systems or recommendation models mostly ignore the diversity of conversation contexts and user personalized needs, and lack intelligent and personalized ways to recommend and guide the direction of interactive learning conversations. To this end, the present invention proposes a conversation recommendation method, system, device, computer device and storage medium, which constructs a conversation recommendation dataset based on a corpus, including a course recommendation dataset, a conversation context interactive intention sequence recommendation dataset, a conversation context source / response pair recommendation dataset and a user history conversation recommendation dataset, and generates a first recommendation result based on the course recommendation dataset, generates a second recommendation result based on the first recommendation result and the conversation context interactive intention sequence recommendation dataset, generates a third recommendation result based on the second recommendation result and the conversation context source / response pair recommendation dataset, generates a fourth recommendation result based on the third recommendation result and the user history conversation recommendation dataset, and generates a conversation recommendation report based on the first recommendation result, the second recommendation result, the third recommendation result and the fourth recommendation result, thereby improving the effectiveness, richness and personalization of the interactive learning conversation recommendation results.

[0070] Reference Figure 1 , Figure 1 The flowchart of the conversation recommendation method applied to the server provided in the embodiment of the present invention, the execution subject of the method can be any of the aforementioned computer devices (including servers or terminals). The method can include the following steps:

[0071] S11. Construct a conversation recommendation dataset based on the corpus; wherein the conversation recommendation dataset includes a course recommendation dataset, a conversation context interaction intention sequence recommendation dataset, a conversation context source language / response pair recommendation dataset, and a user history conversation recommendation dataset;

[0072] S12, generating a first recommendation result according to the course recommendation dataset;

[0073] S13, generating a second recommendation result according to the first recommendation result and the conversation context interaction intention sequence recommendation data set;

[0074] S14, generating a third recommendation result according to the second recommendation result and the conversation context source language / response pair recommendation data set;

[0075] S15, generating a fourth recommendation result according to the third recommendation result and the user historical session recommendation data set;

[0076] S16. Generate a conversation recommendation report according to the first recommendation result, the second recommendation result, the third recommendation result, and the fourth recommendation result.

[0077] Specifically, for the above step S11, the conversation recommendation dataset includes a course recommendation dataset, a conversation context interaction intention sequence recommendation dataset, a conversation context source language / response pair recommendation dataset, and a user history conversation recommendation dataset. In some embodiments, in step S11, a conversation recommendation dataset is constructed based on a corpus, including but not limited to steps S111, S112, S113, and S114.

[0078] Step S111: construct a course recommendation dataset based on course meta information, course description and catalog content document information;

[0079] Step S112: constructing a conversation context interaction intention sequence recommendation dataset according to the conversation context sequence;

[0080] Step S113: construct a conversation context source word / response pair recommendation data set according to the conversation context source word / response pair data;

[0081] Step S114: construct a user history conversation recommendation data set based on the user history conversation text data.

[0082] Step S111: construct a course recommendation dataset based on course meta information, course description and catalog content document information;

[0083] In an embodiment of the present invention, the course recommendation data set is composed of course meta information and course description and catalog content key knowledge point information in the course description and catalog content document information.

[0084] The course meta-information is used to describe the basic information of the course, and the course description and catalog content document information includes the course description and catalog content key knowledge point information.

[0085] In some embodiments, the course meta-information includes course name, subject category, applicable grade, instructor, course offering institution, and course evaluation.

[0086] In some embodiments, the course description and directory content document information consists of a user-preset key knowledge point word set and an automatically extracted key knowledge point word set, wherein the automatically extracted key knowledge point word set performs text preprocessing on the course description and directory content documents to obtain the course description and directory content document key knowledge point information.

[0087] Among them, text preprocessing includes word segmentation, stop word removal, noise filtering and word ordering.

[0088] Specifically, in an embodiment of the present invention, the JE-analysis word segmentation tool is used to segment each course description and catalog content document information, and extract key knowledge point word sets to obtain a course description and catalog content document key knowledge point information set.

[0089] Step S112: constructing a conversation context interaction intention sequence recommendation dataset according to the conversation context sequence;

[0090] The conversation context sequence consists of multiple source and response pairs, each of which consists of a single utterance, and each single sentence consists of the utterance subject, utterance content, and its interactive intent category label. Each utterance content contains different interactive graphs. Considering this information can provide additional signals, which helps to improve the conversation recommendation with the right intention and reduce the conversation recommendation results with the wrong intention, thereby improving the accuracy and effectiveness of the conversation recommendation model.

[0091] In some embodiments, the subject of the speech is a teacher or a learner; the speech content is a speech text transcribed from a course teaching video file or a course teaching audio file; the interaction intention category label is generated by manually or machine-coding the speech content by the intention category.

[0092] In an embodiment of the present invention, an interaction intention coding system including 12 interaction intention categories is used to describe the characteristics of users in information seeking sessions. These interaction intention categories include labels related to questions, such as "initial question (QQ)", "question clarification (CQ)"; labels related to answers, such as "potential answer (PA)", "more details (FD)"; labels related to feedback, such as "positive feedback (PF)", "negative feedback (NF)"; labels related to greetings, such as "greetings / thanks (GG)", which cover most types of user interaction intentions in learning community interaction sessions.

[0093] Specifically, "Initial Question (QQ)" refers to the initial question of a conversation context sequence; "Repeat Question (RQ)" refers to asking a question that repeats the previous question; "Clarification Question (CQ)" refers to asking for clarification of a question; "More Details (FD)" refers to providing more details; "Follow-up Question (FQ)" refers to following up on a related question; "Information Provision (IR)" refers to asking for information; "Potential Answer (PA)" refers to potential solutions to a problem; "Positive Feedback (PF)" refers to positive feedback on effective solutions; "Negative Feedback (NF)" refers to negative feedback on useless solutions; "Greetings / Thanks (GG)" refers to greeting each other or expressing gratitude; "Useless Information (JK)" refers to useless information in the conversation; and "Other (O)" refers to utterances that cannot be classified.

[0094] Figure 2A sample conversation context sequence is shown to illustrate the interaction intention conversion pattern of users in the learning community, involving seven interaction intentions: "initial question (QQ)", "potential answer (PA)", "follow-up question (FQ)", "question clarification (CQ)", "more details (FD)", "positive feedback (PF)" and "greetings / thanks (GG)".

[0095] In this embodiment, the coding of the interaction intention category is performed by a coding staff who has undergone annotation training. The coding staff labels the conversation context sequence according to the encoding procedure based on the preset interaction intention coding system.

[0096] In some embodiments, when the amount of data is very large, an automatic encoding model can be trained based on part of the labeled data, and the resulting training model can be used to automatically label the remaining unencoded data.

[0097] After interaction intent encoding, any conversation context sequence can be represented as the following conversation context interaction intent sequence:

[0098] s i =[I1, ..., I i , ..., I I ],I i ∈d I

[0099] Among them I i is the interaction intent category label of the conversation context sequence.

[0100] S113, constructing a conversation context source word / response pair recommendation dataset according to the conversation context source word / response pair data in the conversation context sequence;

[0101] The conversation context source word / response pair data is each source word and response pair in the conversation context sequence.

[0102] Among them, the conversation context source / response pair recommendation dataset contains the conversation context source / response pair interaction intention representation vector and the conversation context source / response pair self-attention / cross-attention matching matrix.

[0103] In this embodiment, for a session context sequence data set consisting of multiple session context sequences in in is the content of the speech in the tth round of the i-th conversation. i and Y i Is the response candidate set Corresponding tags in express It's U iA true response. Otherwise Conversational context source / response pairs Corresponding to the interaction intent category label at the sequence level, where and are the interaction intent category labels of the source / response pair in the i-th conversation context.

[0104] In some embodiments, in step S113, for U i A response candidate given in and a source language The conversation context source / response pairs can be represented from different perspectives, including but not limited to the following two methods: one is to perform vector representation of the interaction intent of the conversation context source / response pairs through an intent classifier; the other is to perform semantic encoding representation of the conversation context source / response pairs through a Transformers encoder and generate a self-attention / cross-attention matching matrix of the conversation context source / response pairs through self-attention and cross-attention matching.

[0105] In this embodiment, a neural classifier is used to characterize and classify the interactive intent of the conversation context source language / response pair, and generate a representation vector of the interactive intent of the conversation context source language / response pair.

[0106] Specifically, given the conversation context source / response pair data, generate and Embedding vector sequence, applying convolutional filter of shape (f, d) to a window of f words to generate new feature c i , and generate a feature map c = {c1, c2, ..., c n--+1 These features are concatenated and flattened into an output tensor and projected into a fully connected layer (l t ,1). Among them, l t is the number of different interaction intent category labels.

[0107] In this embodiment, a Transformers encoder is used to encode and represent the semantic dependency information in the conversation context source / response pair data.

[0108] Specifically, the context representation of each position is first calculated through Transformers using a scaled dot product attention mechanism, which consists of a query vector and a set of key-value pairs. Secondly, a feedforward neural network with linear rectification as the activation function is added to the output attention (Q, K, V) and the layer-normalized sum of the query Q as a Transformer Encoder module for semantic feature extraction of the conversation context source / response pair to capture the dependency information within words in the same conversation context source / response pair and the association between words in two different conversation context source / response pairs. Finally, the extracted semantic encoding representation of the conversation context source / response pair is self-attention and cross-attention matched to extract the matching representation, and the conversation context source / response pair self-attention / cross-attention matching matrix is ​​generated.

[0109] Step S114: construct a user history conversation recommendation data set based on the user history conversation text data;

[0110] During conversational interactions, individual communication styles vary. For example, women tend to use language that builds connection and intimacy, while men tend to use language that establishes status and independence. Identifying individual communication styles and style characteristics, and designing ways to convey information based on situational requirements, can help improve the effectiveness and efficiency of communication and interaction.

[0111] In this embodiment, a user history conversation recommendation data set is constructed based on the user history conversation text data, wherein the user history conversation recommendation data set includes a personalized language statistical feature data set and a user conversation language feature data set based on a psychology dictionary.

[0112] In some embodiments, in step S114, a user history conversation recommendation data set is constructed based on the user history conversation text data, including but not limited to steps S1141, S1142, and S1143:

[0113] Step S1141, constructing a tokenized conversation text data set based on the user's historical conversation text data preprocessing;

[0114] Step S1142: construct a personalized language statistical feature dataset based on the tokenized conversation text dataset;

[0115] Step S1143, constructing a user conversation language feature dataset based on the psychology dictionary;

[0116] Step S1141: construct a tokenized conversation text dataset based on the user's historical conversation text data preprocessing.

[0117] Specifically, a series of tokens are made to the user's historical conversation text data, which are separated by spaces and roughly correspond to "words". Considering that people with different personalities may have different habits of using punctuation marks, symbols, emoticons and capital letters, the original elements of the user's historical conversation text are retained as much as possible, and no words, letters or symbols are deleted, so as to extract relatively complete personalized language statistical features. Specifically, only necessary spaces are added between different conversation text elements (words, punctuation marks, emoticons, URLs, numbers, etc.), and unnecessary spaces in single conversation text elements (such as ∧∧, (*~*) and other emoticons) are deleted to ensure that each conversation text element can be regarded as a complete token, rather than a series of meaningless separators or punctuation marks. The tokenized conversation text output in this step is integrated and aggregated to form a tokenized conversation text dataset, which is used as the input data in the process of extracting user conversation language statistical feature data. Among them, the tokenized conversation text consists of different conversation text elements, each conversation text element is divided by a space, and each conversation text element is a complete token.

[0118] Step S1142: construct a personalized language statistical feature dataset based on the tokenized conversation text dataset;

[0119] Specifically, during interactive learning, users tend to use informal language, which may contain arbitrarily defined terms and punctuation marks, and may also use capital letters and emoticons to express their emotions. Statistical analysis of these special tokens in the tokenized conversational text dataset helps identify user conversational style.

[0120] In an embodiment of the present invention, six personalized language statistical features of the user are extracted from the tokenized conversation text data set, and a personalized language statistical feature data set is formed by integration and aggregation.

[0121] The six personalized language statistical features include: 1) the usage ratio of emoticons, 2) the usage rate of characters with no less than 3 consecutive repeated letters or symbols, 3) the usage ratio of uppercase letters, 4) the usage ratio of uppercase words, 5) the usage ratio of reply-type conversation items, and 6) the usage ratio of initiation-type conversation items. Among them, the usage ratio of emoticons = (the usage frequency of emoticons as conversation text elements / the total number of conversation text elements in the tokenized conversation text) × 100%; the usage rate of character strings of no less than 3 consecutive repeated letters or symbols = (the usage frequency of character strings of no less than 3 consecutive repeated letters or symbols as conversation text elements / the total number of conversation text elements in the tokenized conversation text) × 100%; the usage ratio of uppercase letters = (the frequency of uppercase letters / the total number of characters in the tokenized conversation text) × 100%; the usage ratio of uppercase words = (the usage frequency of uppercase words as conversation text elements / the total number of conversation text elements in the tokenized conversation text) × 100%; the ratio of reply-type conversation items = (the number of reply-type conversation items / the total number of user individual conversation items) × 100%; the ratio of initiated conversation items = (the number of initiated conversation items / the total number of conversation items) × 100%.

[0122] Step S1143, constructing a user conversation language feature dataset based on the psychology dictionary;

[0123] The Linguistic Inquiry and Word Count (LIWC-22) tool or the TextMind Chinese psychological analysis system is used to extract user conversation language features based on the psychological dictionary from the tokenized conversation text, and then a user conversation language feature dataset based on the psychological dictionary is formed through integration and summary.

[0124] Specifically, in one embodiment of the present invention, a popular English text analysis software widely used in psychology research, namely LIWC-22, is used to extract four major types of conversational language features based on a psychology dictionary from the tokenized conversational text of each user.

[0125] Alternatively, in another embodiment of the present invention, the TextMind Chinese psychological analysis system is used to extract four types of conversational language features based on the psychological dictionary from each user's tokenized conversation text. The word library, text and symbol processing methods of the TextMind Chinese psychological analysis system used are specifically for the simplified Chinese context, and its word library classification system is compatible with LIWC-22.

[0126] The psychological dictionaries are designed to capture the social and psychological states of users. Each dictionary contains a list of words, word stems, emoticons, and other specific verbal structures.

[0127] The four major categories of conversational language features include: 1) functional vocabulary, such as prepositions, conjunctions, pronouns, etc., which are used to understand the sentence structure, logical relations and language style of conversational texts; 2) content vocabulary, such as objects, actions, states, etc., which are used to understand the topics, emotional colors and description methods involved in conversational texts; 3) psychological process vocabulary, such as thinking, feeling, consciousness, etc., which are used to understand the emotional attitudes, thinking methods and emotional expressions contained in conversational texts; 4) social relationship vocabulary, such as friends, family, leaders, etc., which are used to understand the interpersonal relationship network and social background involved in conversational texts.

[0128] The use of the user conversation language features based on the psychology dictionary can reflect the user's psychological interests and personalized style types to understand the user's psychological state, including their emotions, thinking patterns, and social concerns. For example, if a person uses words such as "happy" and "exciting" more frequently, it indicates a positive emotional state. People who are more confident and have higher social status use the word "you" more frequently, but use the word "I" less frequently.

[0129] Regarding the above step S12, the embodiment of the present invention generates a first recommendation result based on the course recommendation data set, wherein the first recommendation result is a recommendation result of similar courses.

[0130] In some embodiments, in step S12, a first recommendation result is generated based on the course recommendation data set, including but not limited to step S121, step S122 and step S123.

[0131] Step S121: Obtain a first recommendation list from a course recommendation data set according to a first preset rule;

[0132] Step S122: construct a feature vector of key knowledge points of the course description and catalog content document according to the first recommendation list;

[0133] Step S123: Calculate the course similarity based on the course description and the feature vector of the key knowledge points in the catalog content document to generate a first recommendation result.

[0134] Step S121: Obtain a first recommendation list from a course recommendation data set according to a first preset rule;

[0135] The first recommendation list is a first recommendation list of candidate courses.

[0136] The first preset rule includes a fuzzy matching condition and an exact matching rule for screening different attributes in the course meta-information of the course recommendation data set.

[0137] Specifically, in one embodiment of the present invention, the course name, lecturer, institution offering the course and other input items input by the user are obtained, and fuzzy matching is performed from the course recommendation data set to obtain a preliminary course recommendation list; based on the subject category, applicable grade, course evaluation and other options selected by the user, precise matching is performed from the preliminary course recommendation list to obtain a first recommendation list.

[0138] Alternatively, in another embodiment of the present invention, fuzzy matching is performed from the course recommendation data set based on the subject category, applicable grade, course evaluation and other options selected by the user to obtain a preliminary course recommendation list; and precise matching is performed from the preliminary course recommendation list based on input items such as the course name, instructor, and course offering institution entered by the user to obtain a first recommendation list.

[0139] Step S122: construct a feature vector of key knowledge points of the course description and catalog content document according to the first recommendation list;

[0140] Specifically, in an embodiment of the present invention, the total number of courses in the first recommendation list is n, and the total number of key knowledge points in the course description and catalog content document key knowledge point information set T is m.

[0141] Construct the key knowledge point feature vector of the course description and catalog content document as X = {x1, ..., x n}, where x i (1≤i≤n) represents the i-th course in the first recommendation list. Each course x i It can be represented by a feature vector of m key knowledge point dimensions, that is, x i =[k i,1 ,k i,2 ,…,k i,m ], where k i,j (1≤i≤n,1≤j≤m) represents the weight of the jth key knowledge point in the i-th course in the first recommendation list.

[0142] In the embodiment of the present invention, TF-IDF is used to measure the weight k, that is, k i,j =tfi i,j *idf j , where tfi i,j Indicates the frequency of occurrence of key knowledge point h in course i in the first recommendation list, idf j represents the inverse document frequency of key knowledge point j. Then the feature vector X of key knowledge points in course description and catalog content documents can be expressed as:

[0143]

[0144] Step S123, calculating the course similarity based on the course description and the feature vector of the key knowledge points of the catalog content document, and generating a first recommendation result;

[0145] The course similarity is the similarity between the target course description and the key knowledge point feature vector of the catalog content document and any course description and the key knowledge point feature vector of the catalog content document in the candidate course set in the first recommendation list. The candidate course set is a set of courses other than the target course in the first recommendation list. The candidate courses will be ranked by calculating their similarity with the target course. The target course is the course pointed to by the conversation sequence to be recommended.

[0146] In an embodiment of the present invention, the similarity between the target course and any course description of the candidate course set in the first recommendation list and the feature vector of the key knowledge points of the catalog content document is calculated based on the efficient manifold ranking algorithm (EMR).

[0147] Specifically, Figure 4 As shown, according to the feature vector X of the key knowledge points of the course description and catalog content document, the embodiment of the present invention first uses the k-means algorithm to calculate the d centers of X as a set of anchor points that share the same space with the data set, that is, The qth (1≤q≤n) course x q is the target course, and the rest are candidate courses. Let f:X→R represent the ranking function, which is for each course x i (1≤i≤n) is assigned a ranking value fi i . Define the initial vector y={y1,…,y n}T, where y q =1, its y i =0(1≤i≤n,i≠q).

[0148] Specifically, the EMR algorithm is as follows:

[0149] The first step is to calculate the weight matrix Z∈R d*n , where z k,i Represents data point x i With anchor point u k The weight between

[0150] In the embodiment of the present invention, Nadaraya Watson kernel regression and Epanechnikov quadratic kernel function are used to calculate z k,i .

[0151] Among them, Nadaraya Watson kernel regression is expressed as:

[0152]

[0153] The Epanechnikov quadratic kernel function is expressed as:

[0154]

[0155] Where λ(x i )=x i -u [s] |,u [s] is x i The sth nearest neighbor anchor point.

[0156] In the second step, an anchor graph is constructed by connecting each data point to the nearest s anchor points, and then a weight is assigned to each connection through Nadaraya Watson kernel regression and Epanechnikov quadratic kernel function calculation;

[0157] The third step is to construct a T Z defines the adjacency matrix W. If two data points x i and x j is related (w i ,j>0,1≤i,j≤n), then they share at least one common anchor point, otherwise w i,j =0;

[0158] Step 4: Through S=D -1 / 2 WD -1 / 2 W is symmetrically normalized, where D is a diagonal matrix corresponding to the following equation.

[0159]

[0160] where z i is the i-th column of Z,

[0161] Step 5, iterate the following equation until convergence, where α is a parameter in [0,1);

[0162] f(t+1)=αSf(t)+(1-α)y

[0163] Step 6: Let f i * Represents the sequence fi i (t). According to each candidate course x i The fraction f of (1≤i≤n) i * , generate the first recommendation result;

[0164] Specifically, obtaining candidate courses with similarity scores greater than a user-preset threshold from the candidate course set in the first recommendation list as the first recommendation result course set;

[0165] The candidate courses in the first recommendation result course set are sorted according to the similarity scores to generate the first recommendation result.

[0166] For the above step S13, in an embodiment of the present invention, the conversation context sequence associated with the learning user related to the candidate course in the first recommendation result course set is used as the second recommendation list. By analyzing the conversation context interaction intention sequence associated with the second recommendation list, the conversation context sequence with statistically significant interaction intention conversion pattern is identified and sorted to generate the second recommendation result.

[0167] In this embodiment, for the conversation context interaction intention sequence in the conversation context interaction intention sequence recommendation data set, the adjacency matrix is ​​used to analyze the conversion information between the conversation context interaction intentions. The conversion frequency from one interaction intention to another interaction intention in the conversation context interaction intention sequence is calculated according to the adjacency matrix, and the residual coefficient of the interaction intention conversion in the conversation context interaction intention sequence is calculated. Among them, the residual coefficient can be obtained by obtaining the difference between the residual count and the expected count of the interaction intention conversion in the conversation context interaction intention sequence, and then obtaining it according to the quotient of the difference and the preset error estimate.

[0168] Specifically, in this embodiment, the adjacency matrix is ​​displayed as a square array, representing the conversation interaction intention category. The category in the first column is the source node, that is, the conversation context interaction intention type, and the category listed in the first row is the target node, that is, the conversation context interaction intention type. The target node category is usually regarded as a response to the source node category. First, this embodiment calculates the conversion frequency of each conversation interaction intention conversion mode, that is, the conversion frequency from one interaction intention type to another interaction intention type. The larger the number, the more likely a certain conversion is to occur. For example, the matrix value of the row category "initial question (QQ)" and the column category "question clarification (CQ)" is 30, which means that the conversion frequency from "initial question (QQ)" to "question clarification (CQ)" in the conversation context interaction intention sequence is 30. Secondly, the significance of the conversation interaction intention conversion mode is calculated, that is, the residual value coefficient, and the Z value is generated as the residual value coefficient by dividing the original residual by the standard error estimate. If the Z value is greater than a certain significance level threshold (usually 1.96), the association is considered significant, indicating that this type of conversion has outstanding importance in the interactive learning conversation context.

[0169] According to the comparison between the residual value coefficient Z value and the significance level threshold, the interactive intention conversion mode significantly associated with the target conversation source language is identified, and the conversation context sequence whose interactive intention conversion mode residual value coefficient is greater than the significance level threshold is obtained from the second recommendation list, wherein the significance level threshold is the interactive intention conversion significance level preset by the user. The target conversation source language is the terminal conversation of the conversation sequence to be recommended.

[0170] The conversation context sequences in the second recommendation list are sorted according to the significance of the interaction intention conversion pattern to generate a second recommendation result.

[0171] In some embodiments, in one embodiment of the present invention, the conversation context sequences in the second recommendation list are sorted from high to low according to the significance of the interaction intention conversion pattern to generate a second recommendation result.

[0172] For the above step S14, the embodiment of the present invention uses the conversation context source / response pairs associated in the conversation context sequence of the second recommendation result as the third recommendation list. The intention-aware attention mechanism is used to perform weighted matching on the conversation context source / response pair self-attention / cross-attention matching matrix, and the final representation is extracted through a convolutional neural network to generate a ranking score for the conversation context source / response pair. The third recommendation result is generated based on the ranking score of the conversation context source / response pair associated in the conversation context sequence of the second recommendation result.

[0173] The conversation context source / response pair recommendation dataset includes the conversation context source / response pair self-attention / cross-attention matching matrix and the conversation context source / response pair interaction intention representation vector.

[0174] Given a conversation context source / response pair self-attention / cross-attention matching matrix, first stack and aggregate it to form a 4D matching tensor: Among them, l c , l u , l r , L are the number of conversation rounds, the number of words in the conversation context source, the number of words in the response candidate, and the number of stacked layers in TransformerEncoder. t, p, q, l are the indices of the matching tensor in these four dimensions.

[0175] Next, an intent-aware attention mechanism is used to perform weighted matching of matching features through the source language / response in the conversation context, to achieve scaling of the 4D matching tensor and generate a weighted matching tensor, thereby learning different contextual changes in the conversation context and achieving weighted and aggregated matching features of different conversation rounds in a flexible manner.

[0176] Specifically, in this embodiment, it is assumed that Represents the interactive intent representation vector of the source / response pair in the conversation context. The Bilinear intent-aware attention mechanism is used to first calculate and The bilinear interaction between them is then normalized: in, is the bilinear interaction matrix. A t is the attention weight corresponding to the tth round of conversation in the conversation context. According to the attention weight of the source language in the conversation context, the 4D matching tensor is scaled to generate a weighted matching tensor:

[0177] Finally, a two-layer 3D convolutional neural network is used to extract the final representation on the matching tensor and generate the ranking score of the source / response pair in the conversation context. Where f(·) is the candidate response R about D and E i A sorted list of . is a session context sequence dataset consisting of multiple session context sequences, where in is the content of the speech in the tth round of the i-th conversation; R i and Y i Is the response candidate set Corresponding tags in express It's U i A true response, otherwise in is the interaction intent category label associated with the sequence level of the conversation context source / response pair, where and are the interaction intent category labels of the source / response pair in the i-th conversation context.

[0178] For a given target conversation source language associated with a conversation sequence to be recommended, according to the ranking score Filter out the conversation context source word / response pair whose conversation context source word / response pair ranking score is greater than the conversation context source word / response pair ranking score threshold from the second recommendation results as a third recommendation list set;

[0179] The target conversation source language is the terminal conversation of the conversation sequence to be recommended.

[0180] The ranking score threshold of the conversation context source word / response pair is a ranking score threshold of the conversation context source word / response pair preset by the user.

[0181] The conversation context source word / response pairs associated in the conversation context sequence in the second recommendation result are sorted according to the sorting scores of the conversation context source word / response pairs to generate a third recommendation result.

[0182] In some embodiments, in one embodiment of the present invention, the associated conversation context source word / response pairs in the second recommendation result conversation context sequence are sorted from high to low according to the ranking scores of the conversation context source word / response pairs to generate a third recommendation result.

[0183] Regarding the above step S15, the embodiment of the present invention forms a fourth recommendation list by filtering out the user history conversation recommendation data set involved in the conversation context source language / response pairs according to the third recommendation result.

[0184] The user history conversation recommendation data set includes a personalized language statistical feature data set and a user conversation language feature data set based on a psychological dictionary. The above step S15 includes but is not limited to step S151, step S152, step S153, step S154 and step S155.

[0185] Step S151, constructing a personalized conversation language feature sequence based on the user's historical conversation recommendation data set;

[0186] Step S152, calculating a first similarity according to the personalized language statistical feature vector;

[0187] Step S153, calculating a second similarity based on the user conversation language feature vector of the psychology dictionary;

[0188] Step S154, calculating a comprehensive similarity according to the first similarity and the second similarity;

[0189] Step S155: Generate a fourth recommendation result according to the comprehensive similarity.

[0190] For the above step S151, a personalized conversation language feature sequence is constructed according to the user history conversation recommendation data set, wherein the user history conversation recommendation data set includes a personalized language statistical feature data set and a user conversation language feature data set based on a psychology dictionary.

[0191] A personalized conversation language feature sequence is constructed according to personalized language statistical feature data and user conversation language feature data based on a psychological dictionary.

[0192] The personalized conversation language feature sequence includes a personalized language statistical feature vector and a user conversation language feature vector based on a psychological dictionary.

[0193] Specifically, the personalized language statistical feature data is one-hot encoded to form a personalized language statistical feature vector, and the user conversation language feature data based on the psychological dictionary is one-hot encoded to form a user conversation language feature vector P based on the psychological dictionary. l .

[0194] Generate a personalized conversation language feature sequence according to the personalized language statistical feature vector and the user conversation language feature vector based on the psychological dictionary;

[0195] Specifically, in an embodiment of the present invention, according to the personalized language statistical feature vector P in the fourth recommendation list si and the user conversation language feature vector P based on the psychological dictionary li , get the personalized conversation language feature sequence:

[0196] [(P s1 ,P l1 ),...,(P si ,P li ),...,(P sn ,P ln )]

[0197] The personalized conversation language feature entry is a tuple including a personalized language statistical feature vector and a user conversation language feature vector based on a psychology dictionary. Therefore, the fourth recommendation list can be expressed as:

[0198] C={(P s1 ,P l1 ),...,(P si ,P li ),...,(P sn ,P ln )},P si ∈d s ,P li ∈d l

[0199] Step S152, calculating a first similarity according to the personalized language statistical feature vector;

[0200] Among them, the first similarity is the personalized language statistical feature vector w' of the target user s The personalized language statistical feature vector w of any user in the user set associated with the fourth recommendation list s The cosine similarity between the fourth recommendation list and the target user. The fourth recommendation list associated user set is a set of other users in the fourth recommendation list associated users except the target user. The target user is the end session provider of the to-be-recommended session sequence.

[0201] First similarity:

[0202] s s = CosSimilarity(w′ s ,w s )

[0203] S153, calculating a second similarity according to a user conversation language feature vector based on a psychology dictionary;

[0204] The second similarity is the user conversation language feature vector w' based on the psychological dictionary lThe user conversation language feature vector w based on the psychological dictionary of any user in the user set associated with the fourth recommendation list l The cosine similarity between .

[0205] Second similarity:

[0206] s l = CosSimilarity(w′ l ,w l )

[0207] Step S154, calculating a comprehensive similarity according to the first similarity and the second similarity;

[0208] Specifically, in the embodiment of the present invention, the weight coefficient a of the personalized language statistical feature vector and the weight coefficient b of the user conversation language feature vector based on the psychology dictionary are used to calculate the comprehensive similarity according to the first similarity and the second similarity.

[0209] Comprehensive similarity:

[0210]

[0211] S155, generating a fourth recommendation result according to the comprehensive similarity;

[0212] A user whose comprehensive similarity is greater than a comprehensive similarity threshold is obtained from the associated user set in the fourth recommendation list, and a response recommendation set of the associated conversation context source / response pair recommendation data set is used as a fourth recommendation result; wherein the response recommendation is a recommended response to the target conversation source. The target conversation source is the terminal conversation of the conversation sequence to be recommended.

[0213] The comprehensive similarity threshold is a comprehensive similarity threshold preset by the user.

[0214] Specifically, the users in the fourth recommendation list associated user set are sorted according to the comprehensive similarity to generate a fourth recommendation result.

[0215] In some embodiments, in one embodiment of the present invention, the users in the fourth recommendation list associated user set are sorted from high to low according to the comprehensive similarity, and a candidate recommended user set is selected according to the comprehensive similarity threshold. The response recommendation set of the conversation context source / response pair recommendation data set associated with the candidate recommended user is used as the fourth recommendation result.

[0216] For the above step S16, in an embodiment of the present invention, the first recommendation result belongs to the recommendation result at the macro level, and the course with similar attributes and knowledge point sets is recommended according to the different attribute features in the course meta-information and the feature vectors of the key knowledge points of the course description and the catalog content document; the second recommendation result and the third recommendation result belong to the recommendation results at the meso level, wherein the second recommendation result recommends the conversation context sequence significantly associated with the conversation sequence interaction intention conversion mode associated with the target conversation source language according to the conversation context interaction intention conversion mode, and the recommendation result is the conversation context fragment; the third recommendation result adopts the intention-aware-based attention mechanism according to the conversation context source language / response pair interaction intention representation and semantic encoding representation, performs weighted matching on the conversation context source language / response pair self-attention / cross-attention matching matrix, and extracts the final representation through the convolutional neural network to generate the ranking score of the conversation context source language / response pair, and recommends the conversation context source language / response pair with a higher ranking score for the target conversation source language on the basis of the above; the fourth recommendation result belongs to the recommendation result at the micro level, and the response set associated with users with similar conversation styles is recommended according to the user personalized language statistical features in the user historical conversation data and the user conversation language features based on the psychological dictionary.

[0217] In the embodiment of the present invention, after the session recommendation report is generated in step S16, the process further includes: displaying the session recommendation report.

[0218] According to the conversation recommendation method described in steps S11-S16, the present invention constructs a conversation recommendation dataset including a course recommendation dataset, a conversation context interactive intention sequence recommendation dataset, a conversation context source language / response pair recommendation dataset and a user history conversation recommendation dataset based on a corpus, and generates a first recommendation result based on the course recommendation dataset, generates a second recommendation result based on the first recommendation result and the conversation context interactive intention sequence recommendation dataset, generates a third recommendation result based on the second recommendation result and the conversation context source language / response pair recommendation dataset, and generates a fourth recommendation result based on the third recommendation result and the user history conversation recommendation dataset, thereby improving the accuracy and personalization of interactive learning conversation recommendation results.

[0219] In summary, the personalized conversation recommendation model proposed in the present invention is aimed at the interactive learning environment, and combines multiple key elements: similarity of course knowledge points, similarity of conversation context interaction intentions, similarity of conversation context semantics, and similarity of user conversation styles, and can effectively recommend the development direction and content of interactive learning conversations. The embodiment of the present invention can be widely used in the field of conversation recommendation, by constructing a conversation recommendation dataset including a course recommendation dataset, a conversation context interaction intention sequence recommendation dataset, a conversation context source language / response pair recommendation dataset, and a user history conversation recommendation dataset according to a corpus, and generating a first recommendation result according to the course recommendation dataset, generating a second recommendation result according to the first recommendation result and the conversation context interaction intention sequence recommendation dataset, generating a third recommendation result according to the second recommendation result and the conversation context source language / response pair recommendation dataset, and generating a fourth recommendation result according to the third recommendation result and the user history conversation recommendation dataset, thereby improving the effectiveness, pertinence, and richness of the interactive learning conversation recommendation results.

[0220] Another aspect of the embodiment of the present invention further provides a conversation recommendation system, including:

[0221] The first module is used to construct a conversation recommendation dataset based on the corpus; wherein the conversation recommendation dataset includes a course recommendation dataset, a conversation context interaction intention sequence recommendation dataset, a conversation context source language / response pair recommendation dataset and a user history conversation recommendation dataset;

[0222] A second module is used to generate a first recommendation result according to the course recommendation dataset;

[0223] A third module is used to generate a second recommendation result based on the first recommendation result and the session context interaction intention sequence recommendation data set;

[0224] A fourth module is used to generate a third recommendation result according to the second recommendation result and the conversation context source language / response pair recommendation data set;

[0225] A fifth module, configured to generate a fourth recommendation result based on the third recommendation result and the user historical session recommendation data set;

[0226] The sixth module is used to generate a session recommendation report according to the first recommendation result, the second recommendation result, the third recommendation result and the fourth recommendation result.

[0227] It can be understood that the contents of the above method embodiments are all applicable to the present system embodiments, the functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0228] The embodiment of the present invention further provides an electronic device, the electronic device includes a memory and a processor, the memory stores a computer program, and the processor implements the above-mentioned conversation recommendation method when executing the computer program. The electronic device can be any intelligent terminal including a tablet computer, a car computer, etc.

[0229] It can be understood that the contents of the above method embodiments are all applicable to the present device embodiments, the functions specifically implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0230] See also Figure 3 , Figure 3 The hardware structure of an electronic device of another embodiment is illustrated, and the electronic device includes:

[0231] The processor 301 may be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present invention;

[0232] The memory 302 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 302 can store an operating system and other application programs. When the technical solution provided in the embodiment of this specification is implemented by software or firmware, the relevant program code is stored in the memory 302, and the processor 301 calls and executes the conversation recommendation method of the embodiment of the present invention;

[0233] Input / output interface 303, used to implement information input and output;

[0234] The communication interface 304 is used to realize the communication interaction between the device and other devices. The communication can be realized through a wired manner (such as USB, network cable, etc.) or a wireless manner (such as mobile network, WIFI, Bluetooth, etc.);

[0235] A bus 305 that transmits information between the various components of the device (e.g., the processor 301, the memory 302, the input / output interface 303, and the communication interface 304);

[0236] The processor 301 , the memory 302 , the input / output interface 303 and the communication interface 304 are connected to each other in communication within the device via the bus 305 .

[0237] An embodiment of the present invention further provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the above-mentioned conversation recommendation method is implemented.

[0238] It can be understood that the contents of the above method embodiments are all applicable to the present storage medium embodiments, the functions specifically implemented by the present storage medium embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0239] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0240] It should be noted that in various specific embodiments of the present invention, when it comes to the need to perform relevant processing based on data related to user identity or characteristics such as user information, user behavior data, user historical data, and user location information, the user's permission or consent will be obtained first, and the collection, use, and processing of these data will comply with relevant laws, regulations, and standards. In addition, when the embodiment of the present invention needs to obtain the user's sensitive personal information, it will obtain the user's separate permission or consent through a pop-up window or jump to a confirmation page, and after clearly obtaining the user's separate permission or consent, it will obtain the necessary user-related data for the normal operation of the embodiment of the present invention.

[0241] The embodiments described in the embodiments of the present invention are intended to more clearly illustrate the technical solutions of the embodiments of the present invention, and do not constitute a limitation on the technical solutions provided by the embodiments of the present invention. Those skilled in the art can appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of the present invention are also applicable to similar technical problems.

[0242] Those skilled in the art will appreciate that the technical solutions shown in the figures do not limit the embodiments of the present invention and may include more or fewer steps than shown in the figures, or a combination of certain steps, or different steps.

[0243] The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separated, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0244] Those skilled in the art will appreciate that all or some of the steps in the methods disclosed above, and the functional modules / units in the systems and devices may be implemented as software, firmware, hardware, or a suitable combination thereof.

[0245] The terms "first", "second", "third", "fourth", etc. (if any) in the description of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0246] It should be understood that in the present invention, "at least one (item)" refers to one or more, and "plurality" refers to two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can represent: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following items" or similar expressions refers to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0247] In the several embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the above units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0248] The units described above as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0249] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0250] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including multiple instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods of various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store programs.

[0251] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but the scope of the rights of the embodiments of the present invention is not limited thereby. Any modification, equivalent substitution and improvement made by those skilled in the art without departing from the scope and essence of the embodiments of the present invention shall be within the scope of the rights of the embodiments of the present invention.

Claims

1. A conversation recommendation method, characterized in that: The following steps are involved: Constructing a conversation recommendation dataset based on the corpus; wherein the conversation recommendation dataset includes a course recommendation dataset, a conversation context interaction intention sequence recommendation dataset, a conversation context source language / response pair recommendation dataset, and a user history conversation recommendation dataset; Generate a first recommendation result according to the course recommendation dataset; Generate a second recommendation result according to the first recommendation result and the session context interaction intention sequence recommendation data set; generating a third recommendation result according to the second recommendation result and the conversation context source language / response pair recommendation data set; Generate a fourth recommendation result according to the third recommendation result and the user historical session recommendation data set; A session recommendation report is generated according to the first recommendation result, the second recommendation result, the third recommendation result, and the fourth recommendation result.

2. A conversation recommendation method according to claim 1, characterized in that: The step of constructing a conversation recommendation dataset based on the corpus includes the following steps: Constructing the course recommendation dataset according to the course meta information and the course description and catalog content document information; According to the conversation context sequence, construct the conversation context interaction intention sequence recommendation dataset; According to the conversation context source / response pair data, a conversation context source / response pair recommendation dataset is constructed; wherein the conversation context source / response pair recommendation dataset includes a conversation context source / response pair interaction intention representation vector and a conversation context source / response pair self-attention / cross-attention matching matrix; wherein the conversation context source / response pair interaction intention representation vector is obtained by characterizing and classifying the conversation context source / response pair interaction intention through a neural classifier; wherein the conversation context source / response pair self-attention / cross-attention matching matrix is ​​obtained by encoding and characterizing the semantic dependency information in the conversation context source / response pair data through a Transformers encoder; A user history conversation recommendation data set is constructed according to the user history conversation text data; wherein the user history conversation recommendation data set includes a personalized language statistical feature data set and a user conversation language feature data set based on a psychology dictionary.

3. A conversation recommendation method according to claim 2, characterized in that: The step of constructing a conversation recommendation dataset based on the corpus further includes the step of constructing a tokenized conversation text dataset based on the preprocessing of the user's historical conversation text data, which is specifically: The tokenized conversation texts are integrated and summarized to obtain a tokenized conversation text dataset, which is used as input data in the process of extracting conversation language statistical feature data; wherein the tokenized conversation texts are composed of different conversation text elements, each conversation text element is separated by a space, and each conversation text element is a complete token; Constructing a personalized language statistical feature dataset based on the tokenized conversation text dataset; specifically: extracting 6 personalized language statistical features of the user from the tokenized conversation text dataset, and forming the personalized language statistical feature dataset by integrating and aggregating; constructing a user conversation language feature dataset based on a psychology dictionary based on a psychology dictionary by using a linguistic survey and word counting tool or a Wenxin Chinese psychological analysis system; extracting the user conversation language features based on the psychology dictionary from the tokenized conversation text, and forming the user conversation language feature dataset based on the psychology dictionary by integrating and aggregating; Among them, the psychological dictionary is intended to capture the social and psychological states of users.

4. A conversation recommendation method according to claim 1, characterized in that: The step of generating a first recommendation result according to the course recommendation dataset comprises the following steps: Acquire a first recommendation list from the course recommendation dataset according to a first preset rule; Constructing a feature vector of key knowledge points of course description and catalog content documents according to the first recommendation list; Calculate the course similarity based on the feature vectors of the course description and the key knowledge points of the catalog content document; wherein the course similarity is the similarity between the target feature vector of the target course description and the key knowledge points of the catalog content document and the candidate feature vector of the key knowledge points of the course description and content document of any candidate course set in the first recommendation list; Sorting the candidate courses in the first recommendation list according to the course similarity scores, and obtaining the candidate courses whose similarity scores are greater than a user-preset threshold as the course set of the first recommendation result; The candidate courses in the course set of the first recommendation result are sorted according to the similarity scores to generate the first recommendation result.

5. A conversation recommendation method according to claim 1, characterized in that: The step of generating a second recommendation result according to the first recommendation result and the session context interaction intention sequence recommendation data set comprises the following steps: Taking the session context sequence associated with the learning user related to the candidate courses in the course set of the first recommendation result as the second recommendation list; The conversion frequency from one interactive intention to another interactive intention in the conversation context sequence is calculated according to the adjacency matrix, and the residual coefficient of the interactive intention conversion in the conversation context sequence is calculated; wherein the residual coefficient is obtained by obtaining the difference between the residual count and the expected count of the interactive intention conversion in the conversation context sequence, and then obtaining the residual coefficient according to the quotient of the difference and a preset error estimate; According to the comparison between the residual value coefficient and the significance level threshold, identifying the interaction intention conversion mode significantly associated with the target conversation source language; Acquire, from the second recommendation list, a conversation context sequence whose residual value coefficient of the interaction intention conversion mode is greater than a significance level threshold; The conversation context sequences in the second recommendation list are sorted from high to low according to the significance level of the interaction intention conversion pattern to generate a second recommendation result.

6. A conversation recommendation method according to claim 1, characterized in that: The conversation context source / response pair recommendation dataset includes conversation context source / response pairs, a self-attention / cross-attention matching matrix, and a conversation context source / response pair interaction intention representation vector; generating a third recommendation result according to the second recommendation result and the conversation context source / response pair recommendation dataset includes the following steps: taking the conversation context source word / response pairs associated in the conversation context sequence of the second recommendation result as a third recommendation list; Using an attention mechanism based on intent perception, weighted matching is performed on the conversation context source / response pair and the self-attention / cross-attention matching matrix; Extracting the final representation through a convolutional neural network to generate a ranking score for the source / response pair in the conversation context; Filtering out the conversation context source word / response pair whose conversation context source word / response pair ranking score is greater than the conversation context source word / response pair ranking score threshold from the second recommendation results as a third recommendation list set; The conversation context source word / response pairs associated in the conversation context sequence in the second recommendation result are sorted according to the sorting scores of the conversation context source word / response pairs in the third recommendation list set to generate a third recommendation result.

7. A conversation recommendation method according to claim 1, characterized in that: The user history conversation recommendation dataset includes a personalized language statistical feature dataset and a user conversation language feature dataset based on a psychological dictionary; Generating a fourth recommendation result according to the third recommendation result and the user history session recommendation data set comprises the following steps: A fourth recommendation list is formed based on the user history conversation recommendation data set involved in the conversation context source language / response pairs screened out according to the third recommendation result; Constructing a personalized conversation language feature sequence according to the personalized language statistical feature data set and the user conversation language feature data set based on the psychology dictionary; wherein the personalized conversation language feature sequence includes the personalized language statistical feature vector and the user conversation language feature vector based on the psychology dictionary; Calculate a first similarity based on the personalized language statistical feature vector; wherein the first similarity is the cosine similarity between the personalized language statistical feature vector of the target user and the personalized language statistical feature vector of any user in the fourth recommendation list associated user set; the fourth recommendation list associated user set is a set of users other than the target user in the fourth recommendation list associated users; Calculate a second similarity based on the user conversation language feature vector based on the psychology dictionary; wherein the second similarity is a cosine similarity between the user conversation language feature vector based on the psychology dictionary and the user conversation language feature vector based on the psychology dictionary of any user in the user set associated with the fourth recommendation list; Calculating a comprehensive similarity according to the first similarity and the second similarity; The user whose comprehensive similarity is greater than the comprehensive similarity threshold is obtained from the user set associated with the fourth recommendation list, and the response recommendation set of the conversation context source / response pair recommendation data set associated with the user is taken as the fourth recommendation result.

8. A conversation recommendation system, characterized in that: include: The first module is used to construct a conversation recommendation dataset based on the corpus; wherein the conversation recommendation dataset includes a course recommendation dataset, a conversation context interaction intention sequence recommendation dataset, a conversation context source language / response pair recommendation dataset and a user history conversation recommendation dataset; A second module is used to generate a first recommendation result according to the course recommendation dataset; A third module is used to generate a second recommendation result based on the first recommendation result and the session context interaction intention sequence recommendation data set; A fourth module is used to generate a third recommendation result according to the second recommendation result and the conversation context source language / response pair recommendation data set; A fifth module, configured to generate a fourth recommendation result based on the third recommendation result and the user historical session recommendation data set; The sixth module is used to generate a session recommendation report according to the first recommendation result, the second recommendation result, the third recommendation result and the fourth recommendation result.

9. An electronic device, characterized in that: including a processor and a memory; The memory is used to store programs; The processor executes the program to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The storage medium stores a program, and the program is executed by a processor to implement the method according to any one of claims 1 to 7.

Citation Information

Patent Citations

  • Personalized dialogue generation method and system based on long-term and short-term memory information

    CN112115247A

  • Session recommendation method and system based on session data

    CN114282077A

  • Multi-granularity recommendation method, system and device and storage medium

    CN115344787A

  • Personality-based emotion intensity controllable dialogue generation method

    CN116166778A