Group discussion theme extraction method and device, electronic equipment and storage medium

Abstract

The invention provides a theme extraction method and device for group discussion, electronic equipment and a storage medium. The method comprises the following steps: acquiring a time sequence text sequence in a group discussion process; based on the text segmentation model, dividing discussion record data in the time sequence text sequence to obtain a plurality of subject areas; and based on a large language model, performing subject extraction on the discussion record data in each subject area, and determining a discussion subject of each subject area. Through the text segmentation model constructed by the PoNet, the BiLSTM, the Dropout layer and the full connection layer, the discussion record data in the time sequence text sequence is divided to obtain a plurality of subject regions, so that the text structure is clearer. The semantic comprehension and analysis capability of the large language model is utilized to automatically extract and generate the topic sentence for each topic paragraph as the representation of the topic, so that automatic text processing and topic analysis are realized, and the efficiency and accuracy of discussing topic determination are improved.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular, to a method, device, electronic device and storage medium for extracting topics of group discussion. Background Art

[0002] Group discussion has always been a key means to promote innovation and solve complex problems. By bringing together the wisdom and experience of multiple individuals, group discussion provides diverse perspectives and ideas for problem-solving. In the process of sharing knowledge, experience and ideas, the group can jointly generate more creative solutions and find the most consensus and feasible solutions through sufficient discussion and debate.

[0003] A large number of discussion records involve the exchange of opinions, the collision of viewpoints and the formation of consensus among multiple participants, and at the same time bring huge text data. Such data contains rich topic information. The length of these text information is relatively long, and it contains many colloquial words and referential ellipsis content, resulting in decision-makers needing to spend a lot of time to understand the topic of the conversation, which brings great information processing challenges to decision-makers. How to achieve efficient and accurate extraction of discussion topics is an important issue that the industry urgently needs to solve at present. Summary of the Invention

[0004] The present invention provides a method, device, electronic device and storage medium for extracting topics of group discussion, so as to solve the problem of efficient and accurate extraction of discussion topics during the group discussion process.

[0005] The present invention provides a method for extracting topics of group discussion, including the following steps: Obtain a time-series text sequence during the group discussion process, where the time-series text sequence includes multiple discussion record data, and the discussion record data includes discussion participant information, discussion time information, and discussion content information; Based on a text segmentation model, divide the discussion record data in the time-series text sequence to obtain multiple topic regions. The text segmentation model is constructed based on a position-aware network PoNet, a bidirectional long short-term memory network BiLSTM, a Dropout layer, and a fully connected layer, and the text segmentation model is trained based on discussion record samples and the topic start sentence labels marked in the discussion record samples; Based on a large language model, extract topics from the discussion record data in each topic region to determine the discussion topics of each topic region.

[0006] According to a method for extracting topics of group discussion provided by the present invention, the step of dividing the discussion record data in the time-series text sequence based on the text segmentation model to obtain multiple topic regions includes: According to the time sequence, input the research record data in the time-sequence text sequence into the text segmentation model item by item to determine whether each research record data is the starting sentence of the research topic; Based on the starting sentences of the research topics determined in each research record data, divide the research record data in the time-sequence text sequence to obtain multiple topic regions.

[0007] According to a method for extracting topics of group discussion provided by the present invention, based on the text segmentation model, determining whether the research record data is the starting sentence of the research topic includes: Input the research record data into the PoNet for local semantic relationship enhancement to obtain the research record data with enhanced context representation; Input the research record data with enhanced context representation into the BiLSTM for feature extraction to obtain global time-sequence dependence features; Input the global time-sequence dependence features into the Dropout layer for regularization processing to obtain the globally time-sequence dependence features after regularization processing; Input the globally time-sequence dependence features after regularization processing into the fully connected layer for classification to determine whether the research record data is the starting sentence of the research topic.

[0008] According to a method for extracting topics of group discussion provided by the present invention, before inputting the research record data into the PoNet for local semantic relationship enhancement, it further includes: Input the research record data into the Bert model for high-dimensional semantic feature extraction to obtain the embedded vector representation of the research record data.

[0009] According to a method for extracting topics of group discussion provided by the present invention, the large language model is initialized based on the parameters of the migrated BART-Chinese model on the basis of the BigBird model architecture.

[0010] According to a method for extracting topics of group discussion provided by the present invention, after determining the research topics of each topic region, it further includes: Based on the clustering algorithm, determine the density relationship between each research topic, and based on the density relationship, group the research record data corresponding to similar topics into the same topic cluster; Based on the graphical interface, visually display the obtained multiple topic clusters.

[0011] The present invention also provides a device for extracting topics of group discussion, including the following modules: A data acquisition module for acquiring a time-series text sequence during a group discussion process, where the time-series text sequence includes multiple discussion record data, and the discussion record data includes discussion participant information, discussion time information, and discussion content information; A theme segmentation module for dividing the discussion record data in the time-series text sequence based on a text segmentation model to obtain multiple theme regions. The text segmentation model is constructed based on a position-aware network PoNet, a bidirectional long short-term memory network BiLSTM, a Dropout layer, and a fully connected layer, and the text segmentation model is trained based on discussion record samples and the theme start sentence labels annotated in the discussion record samples; A theme extraction module for extracting themes from the discussion record data in each theme region based on a large language model to determine the discussion themes of each theme region.

[0012] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and running on the processor. When the processor executes the program, it implements the theme extraction method for group discussion as described in any one of the above.

[0013] The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the theme extraction method for group discussion as described in any one of the above.

[0014] The present invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the theme extraction method for group discussion as described in any one of the above.

[0015] The theme extraction method, device, electronic device, and storage medium for group discussion provided by the present invention use a text segmentation model constructed by PoNet, BiLSTM, a Dropout layer, and a fully connected layer to divide the discussion record data in the time-series text sequence, obtaining multiple theme regions, achieving text segmentation at the theme level, and making the text structure clearer. By using the semantic understanding and analysis capabilities of a large language model, it automatically extracts and generates a theme sentence for each theme paragraph as the representative of the theme, realizing automated text processing and theme analysis, reducing the time and effort required for manual intervention, and improving the efficiency and accuracy of determining discussion themes. Description of the Drawings

[0016] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0017] Figure 1 It is a schematic flowchart of the topic extraction method for group discussion provided by the present invention.

[0018] Figure 2 It is a schematic processing flowchart of the text segmentation model provided by the present invention.

[0019] Figure 3 It is a schematic diagram of topic generation provided by the present invention.

[0020] Figure 4 It is a schematic flowchart of the topic extraction process for group discussion applying the present invention.

[0021] Figure 5 It is a schematic structural diagram of the topic extraction device for group discussion provided by the present invention.

[0022] Figure 6 It is a schematic structural diagram of the electronic device provided by the present invention. Detailed implementation manners

[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Obviously, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments in the present invention belong to the scope of protection of the present invention.

[0024] Figure 1 It is a schematic flowchart of the topic extraction method for group discussion provided by the present invention. As Figure 1 shown, the method includes the following: Step 110: Obtain a sequential text sequence during the group discussion process. The sequential text sequence includes multiple discussion record data, and the discussion record data includes discussion participant information, discussion time information, and discussion content information. Step 120: Based on the text segmentation model, divide the discussion record data in the sequential text sequence to obtain multiple topic regions. The text segmentation model is constructed based on a position-aware network PoNet, a bidirectional long short-term memory network BiLSTM, a Dropout layer, and a fully connected layer. The text segmentation model is trained based on discussion record samples and the topic start sentence labels marked in the discussion record samples. Step 130: Based on the large language model, perform topic extraction on the discussion record data in each topic region to determine the discussion topics of each topic region.

[0025] The execution subject of the topic extraction method for group discussion provided by the present invention can be an electronic device, a component in the electronic device, an integrated circuit, or a chip. The electronic device can be a mobile electronic device or a non-mobile electronic device. Exemplarily, the mobile electronic device can be a mobile phone, a tablet computer, a laptop computer, a handheld computer, an ultra-mobile personal computer (UMPC), a netbook, or a personal digital assistant (PDA), etc., and the non-mobile electronic device can be a server, a Network Attached Storage (NAS), or a personal computer (PC), etc. The present invention does not make specific limitations.

[0026] Taking a computer executing the topic extraction method for group discussion provided by the present invention as an example, the technical solution of the present invention will be described in detail below.

[0027] In step 110, a time-series text sequence in the group discussion process is obtained. The time-series text sequence includes multiple discussion record data, and the discussion record data includes discussion participant information, discussion time information, and discussion content information. In the group discussion process, obtaining the time-series text sequence is a key step, which helps to understand and analyze the dynamic process of the discussion, the interaction of participants, and the evolution of the discussion content. Specifically, the group discussion process can be in the form of a discussion in an online meeting composed of multiple people, or in the form of a chat discussion in a chat group constructed by multiple people. The chat records generated during the meeting communication or chat discussion can be called discussion record data.

[0028] The time-series text sequence is composed of multiple discussion record data, and each record details the key information of the discussion, including discussion participant information, discussion time information, and discussion content information.

[0029] The discussion participant information is the basic part of the time-series text sequence, which records the identity information of the individuals or teams participating in the discussion. The discussion time information is an important part of the time-series text sequence, which records the time stamp of each discussion record. The discussion content information is the core part of the time-series text sequence, which records the specific viewpoints and discussion content generated during the discussion.

[0030] Specifically, the discussion record data is collected from the original records of the group discussion. These data are sourced from meeting records, online discussion records, chat records, etc., and it is necessary to ensure the integrity and representativeness of the data. To improve the quality of the data, the following measures can be taken to optimize data collection: Diverse data sources: Combine different forms of recorded data (such as audio transcripts, video captions, etc.) to supplement potentially missing information.

[0031] Data deduplication: Remove duplicate records through text matching algorithms to avoid redundant information in the analysis process.

[0032] Annotation and segmentation: Add basic tags (such as timestamps, speaker roles, etc.) during data collection to provide richer context information for subsequent analysis.

[0033] After collecting the seminar record data, it is necessary to preprocess the seminar record data and preprocess the collected text data, including the following steps: Text cleaning: Remove invalid characters (such as punctuation marks, HTML tags, etc.) to ensure the standardization and consistency of the text content.

[0034] Word segmentation: Perform word segmentation on the text to generate a sequence of words. This step can use specialized Chinese word segmentation tools (such as Jieba or HanLP) to improve the accuracy of word segmentation.

[0035] Stop word removal: Delete common but meaningless words (such as "de", "shi", etc.) to reduce noise and improve the efficiency and accuracy of subsequent analysis.

[0036] Formatting: Convert the text into a structured format according to the analysis requirements, such as segmenting the dialogue data by speaker and time.

[0037] In addition, to further improve the preprocessing effect, natural language processing techniques can be used for language error correction and lemmatization operations to ensure that the text content is clearer and more accurate.

[0038] In step 120, based on the text segmentation model, the seminar record data in the temporal text sequence is divided to obtain multiple topic regions. The text segmentation model is constructed based on the position-aware network PoNet, the bidirectional long short-term memory network BiLSTM, the Dropout layer, and the fully connected layer. The text segmentation model is trained based on the seminar record samples and the topic start sentence labels annotated in the seminar record samples.

[0039] Based on the segmentation process of the text segmentation model, each seminar record data in the temporal text sequence can be binary-classified in chronological order to determine whether the seminar record data is the start sentence of a new topic, so as to achieve topic-level text segmentation.

[0040] The text segmentation model specifically adopts a model structure of PoNet + BiLSTM + Dropout + fully connected layer, combines the advantages of local feature modeling and global temporal feature extraction, and precisely classifies the input seminar record data. After topic segmentation, the temporal text sequence will be divided into multiple potential topic paragraphs or topic regions.

[0041] The input of the text segmentation model can be the preprocessed seminar record data, and the output is the binary classification result of each sentence. Optionally, the specific processing process of the text segmentation model can be as Figure 2 shown in the processing flow schematic diagram of the text segmentation model provided by the present invention. The seminar record data is preprocessed to obtain sentence encodings. The sentence encodings are input into the PoNet network.

[0042] Among them, PoNet is a position-aware network used to enhance the local semantic relationship modeling ability between seminar record data. Each seminar record data embedding obtains an enhanced context representation through the PoNet layer; BiLSTM is added to extract global temporal dependency features, and finally Dropout (to prevent overfitting) is used to increase the generalization ability of the model; the output of the Dropout layer is passed into the fully connected layer to map the high-dimensional temporal features to the binary classification space, generating the probability that each seminar record data is the start sentence of the topic. After passing through the text segmentation model, the original text is divided into multiple topic paragraphs.

[0043] For example, for a temporal text sequence of one thousand lines, based on the text segmentation model, analyze sentence by sentence, and determine that the start sentences of the topics in the temporal text sequence are the 1st line, 107th line, and 589th line. Based on the determined three start sentences of the topics, the temporal text sequence can be divided into three parts, that is, three topic paragraphs are obtained.

[0044] In step 130, based on the large language model, the seminar record data in each topic region is subjected to topic extraction to determine the seminar topics of each topic region.

[0045] Utilize the semantic understanding and analysis ability of the large language model to extract and generate topic sentences for each topic paragraph as the representative of the topic. Specifically, it can be as Figure 3 shown in the topic generation schematic diagram provided by the present invention. The seminar record data in multiple topic regions is respectively subjected to topic extraction to obtain multiple seminar topics, realizing the rapid determination process of seminar topics.

[0046] Optionally, since the BigBird model has advantages in processing long texts and can capture long-distance dependencies, a fine-tuned BigBird model can be used for topic generation. Since there is no suitable Chinese pre-trained model for igBird, the parameters of the BART-Chinese model are migrated to initialize the BigBird model so that it can better adapt to the processing of Chinese texts.

[0047] The input of the large language model comes from the semantic representation of the topic paragraphs in the text segmentation process, and each paragraph represents the content of a topic or a sub-topic. The output is the topic representation of each paragraph, usually in the form of a generated text summary. These representations reflect the core content of each topic. BigBird is an optimized version of the Transformer model that can process long texts through the global attention mechanism. Compared with the Transformer model, BigBird makes more efficient use of local attention and sparse attention, can capture long-distance dependencies, and is particularly suitable for processing long document tasks. The BigBird model is fine-tuned using labeled topic data so that it can generate topic representations that are more suitable for practical applications.

[0048] The topic extraction method for group discussion provided by the present invention realizes the division of the discussion record data in the time-series text sequence through a text segmentation model constructed by PoNet, BiLSTM, Dropout layer and fully connected layer, obtains multiple topic regions, realizes topic-level text segmentation, and makes the text structure clearer. Utilizing the semantic understanding and analysis capabilities of the large language model, it automatically extracts and generates topic sentences for each topic paragraph as the representative of the topic, realizes automated text processing and topic analysis, reduces the time and effort required for manual intervention, and improves the efficiency and accuracy of determining discussion topics.

[0049] In one embodiment, based on the text segmentation model, the discussion record data in the time-series text sequence is divided to obtain multiple topic regions, including: inputting the discussion record data in the time-series text sequence into the text segmentation model item by item according to the time sequence to determine whether each discussion record data is the starting sentence of a discussion topic; based on the starting sentences of the discussion topics determined in each discussion record data, the discussion record data in the time-series text sequence is divided to obtain multiple topic regions.

[0050] Input the discussion record data item by item into the text segmentation model in chronological order. The model will judge each record to determine whether it is the starting sentence of a discussion topic. From all the discussion record data in the time-series text sequence, determine all the discussion record data marked as starting sentences.

[0051] Starting from the starting sentence, divide the subsequent research record data (until the next starting sentence appears or the sequence ends) into a topic area. Each topic area represents an independent research topic. Divide all the research record data in the time-series text sequence to obtain multiple topic areas.

[0052] Optionally, if two adjacent topic areas are highly similar in content or have logical continuity, they can be considered for merging into a larger topic area.

[0053] In one embodiment, based on the text segmentation model, determining whether the research record data is the starting sentence of a research topic includes: inputting the research record data into the PoNet for local semantic relationship enhancement to obtain the research record data with enhanced context representation; inputting the research record data with enhanced context representation into the BiLSTM for feature extraction to obtain the global temporal dependence features; inputting the global temporal dependence features into the Dropout layer for regularization processing to obtain the globally temporally dependent features after regularization processing; inputting the globally temporally dependent features after regularization processing into the fully connected layer for classification to determine whether the research record data is the starting sentence of a research topic.

[0054] The text segmentation model adopts a model structure of PoNet + BiLSTM + Dropout + fully connected layer, combines the advantages of local feature modeling and global temporal feature extraction, and accurately classifies the input research record data.

[0055] Position-aware Network PoNet: PoNet uses a position encoding mechanism to enhance the local semantic relationship between research record data. After each research record data embedding passes through the PoNet layer, an enhanced context representation is obtained. : ; Among them, PoNet uses a position encoding mechanism and considers the relative positions of the research record data.

[0056] Bidirectional LSTM (BiLSTM): Based on the output of PoNet, BiLSTM is introduced to extract global temporal dependence features: ; Among them, , d is the number of hidden units of the LSTM, and the output dimension of the bidirectional LSTM doubles; are the hidden state and memory state at the last time step respectively.

[0057] Dropout (Preventing Overfitting): In the classification task, to improve the generalization ability of the model, a Dropout layer is added: ; Classification and Output: In the last step, the output of the Dropout layer is fed into the fully connected layer to map the high-dimensional time series features to the binary classification space: ; where , σ is the Sigmoid activation function, which generates the probability that each discussion record data is the starting sentence of the topic.

[0058] Through the above steps, efficient and accurate classification of discussion record data is achieved, providing reliable input data for the topic generation module.

[0059] In one embodiment, before inputting the discussion record data into the PoNet network, it further includes: inputting the discussion record data into the Bert model for high-dimensional semantic feature extraction to obtain the embedded vector representation of the discussion record data.

[0060] Specifically, after collecting the discussion record data, the original text data of the group discussion is converted into a machine-readable high-dimensional vector representation. The BERT technology is used for word embedding to capture the semantic information and context correlation in the text.

[0061] Converting the original text data into a form that the model can process includes: Text Normalization: The preprocessed text data is further normalized to ensure that the model can correctly understand the input data; Cleaning Invalid Characters: Removing noises such as extra spaces, HTML tags, etc., and preserving the semantic integrity of the text.

[0062] Tokenization: Using the BERT Tokenizer to convert the text into corresponding Token IDs and special tokens ([CLS], [SEP], etc.): ; This step generates the input sequence and its corresponding Token IDs.

[0063] Attention Mask: Generating the corresponding Attention Mask according to the length of the input sequence, which is used to indicate the valid part that the model should focus on, and marking the valid Tokens: setting the Tokens in the non-padding part to 1 and the padding part to 0.

[0064] ; BERT model processing: Input the generated input data (Token ID and Attention Mask) into the pre-trained BERT model to extract high-dimensional semantic features; The BERT model takes input_ids and attention_mask as inputs and generates token-level hidden states: ; ; where n is the length of the input sequence and 768 is the dimension of the BERT hidden layer.

[0065] Paragraph representation extraction: Extract the paragraph-level semantic representation from the output of the BERT model, and use the output of the [CLS] token as the global semantic vector of the paragraph.

[0066] ; Output generation: Finally, the model outputs the global semantic representation of each sentence, providing input data for subsequent analysis steps.

[0067] Through the above steps, the text data of group discussion is successfully converted into a machine-understandable high-dimensional semantic vector representation, providing an accurate input basis for topic segmentation and generation.

[0068] In one embodiment, the large language model is initialized based on the parameters of the migrated BART-Chinese model on the basis of the BigBird model architecture.

[0069] The BigBird model architecture is adopted and is designed specifically for processing long text tasks. BigBird optimizes the attention mechanism of the traditional Transformer and effectively captures long-distance dependencies by combining global, local, and sparse attention.

[0070] Parameter initialization: Since BigBird lacks a pre-trained model suitable for Chinese, the parameters of the migrated BART-Chinese model are used to initialize BigBird to enhance its Chinese text processing ability. BART-Chinese is a pre-trained model optimized for Chinese and can provide reasonable initial weights for BigBird.

[0071] Optionally, labeled topic paragraph data can also be used to fine-tune BigBird to learn the mapping relationship between paragraphs and topic summaries.

[0072] During the training process, the model adjusts its parameters through the cross - entropy loss function, which measures the difference between the generated text and the true labels, enabling the generated topics to accurately reflect the core content of each paragraph: ; Among them, is the true label (i.e., the description of the topic), is the label generated by the model.

[0073] The Adam optimizer is used for parameter update, and the learning rate is adjusted to avoid overfitting.

[0074] The fine - tuned BigBird receives the semantic representation of the topic paragraph and generates the corresponding topic summary. This process can be described as follows: ; Among them, is the topic representation generated by the model, usually a text summary.

[0075] In one embodiment, after determining the discussion topics of each topic area, it further includes: based on the clustering algorithm, determining the density relationship between each discussion topic, and based on the density relationship, classifying the discussion record data corresponding to similar topics into the same topic cluster; based on the graphical interface, visually displaying the obtained multiple topic clusters.

[0076] Specifically, the extracted topics are grouped through the clustering algorithm and visualized to help users better understand the topic structure of the group discussion. Specifically, the DBSCAN algorithm can be used to cluster each topic paragraph to reveal the relationships between different topics, and the UMAP dimensionality reduction technology can be used to map high - dimensional data to a two - dimensional space to generate an intuitive visualization result. At the same time, the clustering result is displayed through the graphical interface to provide a clear and understandable topic overview.

[0077] The input comes from the topic representation extracted by the large - language model, which can be the vector representation of each topic paragraph.

[0078] The output is the topic labels after clustering, indicating the topic cluster to which each paragraph belongs, usually output in the form of topic cluster numbers. And a topic relationship graph, a distance matrix between topic clusters, etc. are drawn to enable users to intuitively understand the similarity and relationship between each topic. The visualization results include a topic relationship graph, a heat map of the distance matrix, etc.

[0079] DBSCAN is a density-based hierarchical clustering algorithm that can automatically determine the number of clusters and identify noise points in the data. It clusters the data through the following steps: Core distance calculation: Measure the density of each point by calculating its core distance; Cluster expansion: Expand from areas of high density and gradually merge points with similar densities into one cluster; Hierarchical clustering structure: Form a hierarchical clustering tree by gradually merging clusters with different densities.

[0080] DBSCAN clusters similar topic paragraphs into the same cluster by considering the density relationships between paragraphs. This algorithm can handle clusters of different shapes and sizes and adapt to the diversity of topic distributions in group discussions.

[0081] Specifically, calculate the similarity between topic paragraphs (such as cosine similarity) and perform clustering based on the similarity. Finally, assign a cluster label to each paragraph. To effectively cluster topic paragraphs, first calculate the similarity between paragraphs (such as using cosine similarity). DBSCAN classifies similar paragraphs into the same topic cluster by analyzing density relationships. Its advantages include: Can handle clusters of different shapes and sizes; Adapt to the diversity of topic distributions; Automatically eliminate noise paragraphs.

[0082] To visualize the clustering results, use UMAP for dimensionality reduction to map high-dimensional topic vectors to a two-dimensional space. Present the clustering results as a two-dimensional scatter plot, where each point represents a topic paragraph and the color of the point indicates the cluster it belongs to. Users can view the distributions of different clusters and the similarities between topics through a graphical interface. Using the distance matrix generated from the clustering results, a similarity matrix between clusters can be generated and the relationships between different topic clusters can be displayed in the form of a heatmap.

[0083] Build a user interaction and interface to enable users to view intuitively. The specific construction process includes: Select a front-end framework: Build a responsive interface using HTML+CSS+JavaScript or React / Vue.js.

[0084] Main module layout: Top navigation bar: Display the project name, load / save buttons, and an entry for parameter adjustment. Main area distribution: The visualization charts (scatter plot, heatmap) are on the left, and the topic summary display area is on the right. Bottom area: Display the current task status (such as data loading progress).

[0085] Implement dynamic interaction of scatter plots: Data preparation: Generate initial topic distribution data (coordinates of points and topic cluster attribution) at the backend using matplotlib or seaborn in Python and transmit it in JSON format. Front-end rendering: Use ECharts or D3.js to load and draw scatter plots at the front end. Specific steps: Set up the basic chart container ( ). Load topic cluster data and draw points of different clusters in different colors. Add mouse hover event: Display paragraph details (paragraph number and topic summary). Add zooming and dragging functions: Implement by setting the zoom and pan options.

[0086] Implement similarity heat maps: Generate a similarity matrix at the backend, calculate the similarity between topics using NumPy or pandas, and store it as a two-dimensional array. Visualize at the front end and use the heat map function of Plotly.js or ECharts to load the similarity matrix. Bind click events to heat map cells: Click on a cell to highlight the corresponding topic cluster.

[0087] Topic summary display module, layout of the right display area: Create a list style through HTML to hierarchically display topic paragraph summaries. Each topic paragraph is displayed in the form of a collapsible card (the summary is displayed by default, and the full content is displayed when clicked to expand). Dynamic update: After the user selects a topic cluster in the scatter plot, send a request to the backend to obtain the paragraph details of that cluster and update the list content. UI design for clustering parameter adjustment function: Add sliders or input boxes in the top navigation bar to allow users to adjust clustering parameters (such as cluster size, distance threshold, etc.).

[0088] Parameter interaction: After the user adjusts the parameters, send the parameter values to the backend via AJAX or WebSocket. The backend recalculates the clustering and returns the updated visualization data. The front end automatically refreshes the scatter plot and heat map.

[0089] Data loading and saving: Data loading, after the user clicks the "Load Data" button, call the backend API to upload the group discussion data file (such as JSON or CSV format). The backend completes topic extraction and clustering analysis and returns the processing results. Result saving, provide a "Save Results" button to download the current analysis results (such as topic paragraphs, similarity matrix) as a file.

[0090] The present invention also provides a schematic diagram of the topic extraction process for group discussion using the present invention. As Figure 4 shown, this process specifically includes input representation, topic segmentation, topic extraction, and topic visualization.

[0091] The input representation preprocesses the text data in the group discussion by using the BERT pre-trained model, the corresponding tokenization tool, and the corresponding vocabulary on the n research record data in the sequential text sequence, and converts the text data into an embedded vector representation for subsequent processing.

[0092] The topic segmentation adopts the model structure of PoNet+BiLSTM+Dropout+fully connected layer, combines the advantages of local feature modeling and global sequential feature extraction, and accurately classifies the input research record data to obtain multiple topic segments.

[0093] The topic extraction is based on the construction of a Chinese BigBird pre-trained model. The parameter weights of the Chinese BART-base pre-trained model are transferred to the original English version of the BigBird pre-trained model to obtain the BigBird-Chinese model for topic extraction, and multiple topics are obtained.

[0094] The topic visualization deeply analyzes and classifies these topic sentences based on the BDSCAN clustering tool. Through this processing process, the originally scattered topic sentences are organized in an orderly manner, forming a clear topic distribution pattern.

[0095] Next, the topic extraction device for group discussion provided by the present invention will be described. The topic extraction device for group discussion described below can be mutually corresponding and referred to the topic extraction method for group discussion described above.

[0096] As Figure 5 shown, the device includes: A data acquisition module 510 for acquiring a sequential text sequence during the group discussion process. The sequential text sequence includes multiple research record data, and the research record data includes research participant information, research time information, and research content information; A topic segmentation module 520 for dividing the research record data in the sequential text sequence based on a text segmentation model to obtain multiple topic regions. The text segmentation model is constructed based on a position-aware network PoNet, a bidirectional long short-term memory network BiLSTM, a Dropout layer, and a fully connected layer, and the text segmentation model is trained based on research record samples and the topic start sentence labels annotated in the research record samples; A topic extraction module 530 for extracting topics from the research record data in each topic region based on a large language model to determine the research topics of each topic region.

[0097] The topic extraction device for group discussion provided by the present invention realizes the division of the discussion record data in the time-series text sequence through a text segmentation model constructed by PoNet, BiLSTM, Dropout layer and fully connected layer, obtains multiple topic regions, realizes topic-level text segmentation, and makes the text structure clearer. Utilizing the semantic understanding and analysis capabilities of large language models, it automatically extracts and generates topic sentences for each topic paragraph as the representative of the topic, realizes automated text processing and topic analysis, reduces the time and effort required for manual intervention, and improves the efficiency and accuracy of determining discussion topics.

[0098] In one embodiment, the topic segmentation module 520 is specifically configured to: Based on the text segmentation model, divide the discussion record data in the time-series text sequence to obtain multiple topic regions, including: Sequentially input the discussion record data in the time-series text sequence into the text segmentation model according to the time sequence to determine whether each discussion record data is the starting sentence of a discussion topic; Based on the starting sentences of the discussion topics determined in the discussion record data, divide the discussion record data in the time-series text sequence to obtain multiple topic regions.

[0099] In one embodiment, the topic segmentation module 520 is further specifically configured to: Based on the text segmentation model, determine whether the discussion record data is the starting sentence of a discussion topic, including: Input the discussion record data into the PoNet for local semantic relationship enhancement to obtain the discussion record data with enhanced context representation; Input the discussion record data with enhanced context representation into the BiLSTM for feature extraction to obtain global time-series dependency features; Input the global time-series dependency features into the Dropout layer for regularization processing to obtain the global time-series dependency features after regularization processing; Input the global time-series dependency features after regularization processing into the fully connected layer for classification to determine whether the discussion record data is the starting sentence of a discussion topic.

[0100] In one embodiment, the topic segmentation module 520 is further specifically configured to: Before inputting the discussion record data into the PoNet for local semantic relationship enhancement, it further includes: Input the discussion record data into the Bert model for high-dimensional semantic feature extraction to obtain the embedded vector representation of the discussion record data.

[0101] In one embodiment, the topic extraction module 530 is specifically configured to: The large language model is initialized based on the parameters of the migrated BART-Chinese model on the basis of the BigBird model architecture.

[0102] In one embodiment, the theme extraction module 530 is further specifically configured to: After determining the discussion themes of each theme area, the following steps are further included: Based on the clustering algorithm, determine the density relationship between each discussion theme, and based on the density relationship, group the discussion record data corresponding to similar themes into the same theme cluster; Based on the graphical interface, visually display the obtained multiple theme clusters.

[0103] Figure 6 An example of the physical structure diagram of an electronic device is shown as Figure 6 As shown, the electronic device may include: a processor 610, a communication interface 620, a memory 630, and a communication bus 640. Among them, the processor 610, the communication interface 620, and the memory 630 complete mutual communication through the communication bus 640. The processor 610 can call the logical instructions in the memory 630 to execute the theme extraction method for group discussion. The method includes: a text segmentation model constructed by a PoNet, a BiLSTM, a Dropout layer, and a fully connected layer, which realizes the division of the discussion record data in the time series text sequence to obtain multiple theme areas, realizes text segmentation at the theme level, and makes the text structure clearer. Using the semantic understanding and analysis capabilities of the large language model, automatically extract and generate a theme sentence for each theme paragraph as the representative of the theme, realizes automated text processing and theme analysis, reduces the time and effort required for manual intervention, and improves the efficiency and accuracy of determining the discussion theme.

[0104] In addition, when the logical instructions in the above-mentioned memory 630 are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0105] On the other hand, the present invention also provides a computer program product. The computer program product includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the method for extracting the theme of group discussion provided by the above-mentioned various methods. The method includes: a text segmentation model constructed by PoNet, BiLSTM, Dropout layer, and fully connected layer realizes the division of the discussion record data in the time-series text sequence, obtains multiple theme regions, realizes text segmentation at the theme level, and makes the text structure clearer. Utilizing the semantic understanding and analysis capabilities of the large language model, automatically extracts and generates a theme sentence for each theme paragraph as the representative of the theme, realizes automated text processing and theme analysis, reduces the time and effort required for manual intervention, and improves the efficiency and accuracy of determining the discussion theme.

[0106] On another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it realizes the method for extracting the theme of group discussion provided by the above-mentioned various methods. The method includes: a text segmentation model constructed by PoNet, BiLSTM, Dropout layer, and fully connected layer realizes the division of the discussion record data in the time-series text sequence, obtains multiple theme regions, realizes text segmentation at the theme level, and makes the text structure clearer. Utilizing the semantic understanding and analysis capabilities of the large language model, automatically extracts and generates a theme sentence for each theme paragraph as the representative of the theme, realizes automated text processing and theme analysis, reduces the time and effort required for manual intervention, and improves the efficiency and accuracy of determining the discussion theme.

[0107] The device embodiments described above are merely illustrative. The units described 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 to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative effort.

[0108] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of each embodiment of the present invention.

Claims

1. A method for extracting themes from group discussions, characterized in that, Including: Obtain a sequential text sequence during group discussion, where the sequential text sequence includes multiple discussion record data, and the discussion record data includes information about discussion participants, discussion time information, and discussion content information; Based on a text segmentation model, divide the discussion record data in the sequential text sequence to obtain multiple topic regions. The text segmentation model is constructed based on a Position-Aware Network (PoNet), a Bidirectional Long Short-Term Memory Network (BiLSTM), a Dropout layer, and a fully connected layer. The text segmentation model is trained based on discussion record samples and the topic start sentence labels annotated in the discussion record samples; Based on a large language model, extract topics from the discussion record data in each topic region to determine the discussion topics of each topic region.

2. The method for extracting the theme of group discussion according to claim 1, wherein The step of dividing the discussion record data in the sequential text sequence based on the text segmentation model to obtain multiple topic regions includes: Input the discussion record data in the sequential text sequence into the text segmentation model one by one according to the sequence to determine whether each discussion record data is the start sentence of a discussion topic; Based on the start sentences of the discussion topics determined in each discussion record data, divide the discussion record data in the sequential text sequence to obtain multiple topic regions.

3. The method for extracting the theme of group discussion according to claim 2, wherein, Based on the text segmentation model, determining whether the discussion record data is the start sentence of a discussion topic includes: Input the discussion record data into the PoNet to enhance local semantic relationships, obtaining discussion record data with enhanced context representations; Input the discussion record data with enhanced context representations into the BiLSTM for feature extraction to obtain global temporal dependence features; Input the global temporal dependence features into the Dropout layer for regularization processing to obtain globally temporally dependent features after regularization processing; Input the globally temporally dependent features after regularization processing into a fully connected layer for classification to determine whether the discussion record data is the start sentence of a discussion topic.

4. The method for extracting the theme of group discussion according to claim 3, characterized in that, Before inputting the discussion record data into the PoNet to enhance local semantic relationships, it further includes: Input the discussion record data into a Bert model for high-dimensional semantic feature extraction to obtain an embedded vector representation of the discussion record data.

5. The method for extracting the theme of group discussion according to claim 1, characterized in that The large language model is initialized based on the parameters of the migrated BART-Chinese model on the basis of the BigBird model architecture.

6. The method for extracting the theme of group discussion according to claim 1, characterized in that After determining the discussion topics of each topic region, it further includes: Based on a clustering algorithm, determine the density relationship between each discussion topic, and based on the density relationship, group the discussion record data corresponding to similar topics into the same topic cluster; Based on a graphical interface, visually display the obtained multiple topic clusters.

7. A theme extraction device for group discussion, characterized in that Including: A data acquisition module for obtaining a sequential text sequence during group discussion, where the sequential text sequence includes multiple discussion record data, and the discussion record data includes information about discussion participants, discussion time information, and discussion content information; A topic segmentation module, configured to divide the discussion record data in the sequential text sequence based on a text segmentation model to obtain multiple topic regions. The text segmentation model is constructed based on a position-aware network PoNet, a bidirectional long short-term memory network BiLSTM, a Dropout layer, and a fully connected layer, and is trained based on discussion record samples and the topic start sentence labels annotated in the discussion record samples; A topic extraction module, configured to extract topics from the discussion record data in each topic region based on a large language model to determine the discussion topics of each topic region.

8. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the topic extraction method for group discussion according to any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the topic extraction method for group discussion according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the topic extraction method for group discussion according to any one of claims 1 to 6.

Images