A method and system for classifying event topics based on conference reports

Abstract

The present invention relates to an event theme classification method and system based on conference reports, which belongs to the field of natural language processing technology and solves the problem of inaccurate event theme classification due to insufficient text feature extraction. The method includes: parsing historical conference reports to obtain an event text set; constructing dynamic prompt words for each event text in the event text set and passing them into a large model, and adding the corresponding multiple inference texts into the event text set to obtain an augmented text set; using the augmented text set to train a dual-channel feature fusion BERT model to obtain an event theme classification model; the dual-channel module in the dual-channel feature fusion BERT model includes a multi-head attention mechanism channel and a text convolution channel; parsing the conference report to be predicted to obtain multiple events to be predicted, passing each event to be predicted into the event theme classification model, and predicting the corresponding event theme category. Accurate classification of event theme categories of conference reports is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and in particular to a method and system for classifying event topics based on conference reports. Background Art

[0002] With the rapid development of information technology, how to efficiently and accurately classify meeting topics from government and group meeting reports has become an urgent problem to be solved.

[0003] Traditional machine learning topic classification methods rely on manual feature extraction and expert experience to design features, which is not only time-consuming and labor-intensive, but also difficult to handle high-dimensional data and complex nonlinear relationships.

[0004] The deep learning BERT model uses a bidirectional encoder. Its core advantage lies in its innovative bidirectional training mechanism. It learns deep language representation through pre-training on a large amount of unlabeled text, allowing the model to capture rich contextual information and language nuances. However, since BERT is a language model based on word granularity, it is easy to ignore grammatical information between words and long-distance text features in sentences. It cannot handle complex semantic text well, resulting in classification errors. Summary of the invention

[0005] In view of the above analysis, an embodiment of the present invention aims to provide a method and system for event topic classification based on conference reports, so as to solve the problem of inaccurate event topic classification of conference reports due to insufficient text feature extraction.

[0006] On the one hand, an embodiment of the present invention provides an event topic classification method based on conference reports, comprising the following steps:

[0007] Parse historical conference reports to obtain an event text set; construct dynamic prompt words for each event text in the event text set and pass them into the big model, and add the corresponding multiple inference texts into the event text set to obtain an augmented text set;

[0008] The dual-channel feature fusion BERT model is trained using the augmented text set to obtain an event topic classification model; the dual-channel module in the dual-channel feature fusion BERT model includes a multi-head attention mechanism channel and a text convolution channel;

[0009] The conference report to be predicted is parsed to obtain multiple events to be predicted, and each event to be predicted is passed into the event topic classification model to predict the corresponding event topic category.

[0010] Based on the further improvement of the above method, the dual-channel feature fusion BERT model includes: BERT network, dual-channel module, feature fusion module and classification prediction module in sequence; the BERT network receives the training text, inputs the last layer of text feature vector into the dual-channel module, and after the feature fusion module fuses the global multi-dimensional features output by the multi-head attention mechanism channel in the dual-channel module and the local key features output by the text convolution channel to obtain the text fusion features, it is passed to the classification prediction module to output the event topic category.

[0011] Based on the further improvement of the above method, a dynamic prompt word is constructed for each event text in the event text set, including:

[0012] Set the role and task objectives of the prompt word;

[0013] Each event text is introduced through the first placeholder as sample information of the conference report content;

[0014] Based on the sample information, text generation rules and target encouragement are set, wherein the event topic category marked by the current event text is introduced through the second placeholder.

[0015] Based on the further improvement of the above method, after receiving the text feature vector, the multi-head attention mechanism channel passes through the multi-head attention sub-module and the first forward network sub-module in sequence, in which residual connections and layer normalization are added to the multi-head attention sub-module and the first forward network sub-module to output global multi-dimensional features.

[0016] Based on the further improvement of the above method, the text convolution channel adopts a text convolutional neural network, which includes: embedding layer, convolution layer, maximum pooling layer and feature integration layer in sequence; after the embedding layer adjusts the dimension of the received text feature vector, multiple convolution result features are extracted through multiple convolution kernels of different sizes in the convolution layer, and the maximum pooling layer extracts the maximum value of each convolution result feature, which is then integrated into local key features by the feature integration layer.

[0017] Based on the further improvement of the above method, the feature fusion module passes the received global multidimensional features and local key features into the feature splicing submodule. The feature splicing submodule obtains high-dimensional features through vector splicing, vector alignment subtraction and vector alignment dot multiplication, and then passes them into the second forward network submodule. Residual connection and layer normalization are added to the second forward network submodule to output text fusion features.

[0018] Based on the further improvement of the above method, the classification prediction module inputs the received text fusion features into the fully connected layer for nonlinear operation and then obtains the classification probability through softmax operation, and takes the event topic category corresponding to the maximum classification probability as the prediction result.

[0019] Based on the further improvement of the above method, the role of the prompt word is set as a conference theme report generation expert; the text generation rules include: do not repeat the unit address in the sample information, do not repeat the person name in the sample information, use a text background similar to the sample information, and the generated conference report content has a similarity with the {THEME} theme above 0.8; where {THEME} is the second placeholder.

[0020] Based on the further improvement of the above method, the large model adopts a 32B four-bit quantized base large model.

[0021] On the other hand, an embodiment of the present invention provides an event topic classification system based on conference reports, including:

[0022] The text set construction unit is used to parse historical conference reports to obtain an event text set; construct dynamic prompt words for each event text in the event text set and pass them into the large model, and add the corresponding multiple inference texts into the event text set to obtain an augmented text set;

[0023] The classification model training unit is used to train the dual-channel feature fusion BERT model using the augmented text set to obtain an event topic classification model; the dual-channel module in the dual-channel feature fusion BERT model includes a multi-head attention mechanism channel and a text convolution channel;

[0024] The event theme classification unit is used to parse the conference report to be predicted to obtain multiple events to be predicted, pass each event to be predicted into the event theme classification model, and predict the corresponding event theme category.

[0025] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0026] 1. A dual-channel feature fusion network model based on a pre-trained network is proposed. The model combines a multi-head attention mechanism network and a text convolutional network to enhance the BERT text feature representation. The multi-head attention mechanism gives the model the ability to learn multiple relationships and features in the text in parallel by focusing on multiple information dimensions at the same time. The text convolutional neural network enhances the model's learning of local text feature key information by focusing on the local key information of the text, thereby enhancing the model's text feature capabilities. Finally, feature fusion technology is used to fuse the semantic features of the dual channels, enhancing the text feature capabilities and improving the accuracy of event subject category classification.

[0027] 2. Based on the semantic features of event texts extracted from historical conference reports, dynamic prompt words are constructed, and large models are used as a means to expand the data set, thereby reducing dependence on large amounts of manually annotated data. This not only effectively reduces manpower and material resources, but also significantly improves the efficiency and quality of data processing.

[0028] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components;

[0030] Figure 1 This is a flow chart of an event topic classification method based on conference reports in Example 1 of the present invention;

[0031] Figure 2 Schematic diagram of the structure of the dual-channel feature fusion BERT model in Example 1 of the present invention. DETAILED DESCRIPTION

[0032] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0033] Example 1

[0034] A specific embodiment of the present invention discloses a method for classifying event topics based on conference reports, such as Figure 1 As shown, the following steps are included:

[0035] S1. Parse historical conference reports to obtain an event text set; construct dynamic prompt words for each event text in the event text set and pass them into the large model, and add the corresponding multiple inference texts into the event text set to obtain an augmented text set.

[0036] It should be noted that the meeting report in this embodiment is usually a Word file or a PDF file, and a meeting report contains one or more important events.

[0037] For Word files, the python-docx library in the python third-party package is used to parse the Word files to obtain each paragraph in the Word files; the paragraphs containing the subject identifiers are extracted according to the subject identifiers to obtain the first event text, which is put into the event text set.

[0038] For PDF files, use the PyMuPDF library in the third-party Python package to parse the PDF files and obtain each paragraph in the PDF files; extract the paragraphs containing the topic identifiers according to the topic identifiers to obtain the second event text, and put it into the event text set.

[0039] It should be noted that the topic identifiers are set according to the compilation rules of the actual conference reports. Exemplarily, the topic identifiers are Chinese numerals plus a punctuation mark, such as "One、", "Two、".

[0040] Furthermore, label the corresponding event topic categories for each event text in the event text set, including: major decision-making matters, major project arrangement matters, major personnel appointment and removal matters, and large-scale fund operation matters.

[0041] Preferably, remove the event texts in the event text set that do not belong to any category and correct the incorrect expressions in the event texts.

[0042] In order to reduce the dependence on manual annotation and reduce the cost and workload of manual annotation, this embodiment uses a large model to generate multiple new event texts with the same category as the input event text to expand the event text set.

[0043] In this embodiment, the large model uses a 32B four-bit quantized base model. For example, the Qwen1.5-32B four-bit quantized base model reduces the storage requirements and computational overhead of the model through quantization technology, enabling it to still run efficiently in resource-constrained environments; and uses the large model inference acceleration library vLLM (Very Large Language Models) to provide efficient management and optimization for large model inference and effectively process large-scale data.

[0044] It should be noted that an efficient prompt will improve the generation ability of the large model. By designing a role-playing function in the prompt, the model belongs to a certain role. By designing a few-shot prompting function in the prompt, the large model can understand the generation direction of the inference content and generate accurate text content.

[0045] Specifically, this embodiment constructs dynamic prompts for each event text in the event text set, including:

[0046] ① Set the role and task objective of the prompt.

[0047] In this embodiment, the role of the prompt is set as an expert in generating conference theme reports. This role setting helps the model understand that it needs to analyze and reason about the text as an expert. The task objective is set to generate multiple data that most conform to the color of human conference reports according to the sample information, effectively guiding the large model to reason and ensuring that the model's inference target and direction are consistent with the task objective.

[0048] ②Introduce each event text as sample information of the conference report content through the first placeholder.

[0049] Traverse the event text set and obtain the content of each event text through the first placeholder, such as {CONTENT}.

[0050] ③ Based on the sample information, set text generation rules and target encouragement, in which the event topic category marked by the current event text is introduced through the second placeholder.

[0051] It should be noted that the event subject category of each event text is obtained through the second placeholder {THEME}; the text generation rules include: do not repeat the unit address in the sample information, do not repeat the person name in the sample information, use a text background similar to the sample information, and the generated conference report content has a similarity of more than 0.8 with the {THEME} theme. In other words, the event text content generated by the large model has the same event subject category as the input event text, and no manual annotation is required.

[0052] The rules for generating text are set to constrain the reasoning direction of the large model, so that the large model can trigger the lf-reflection mechanism in the reasoning process, making the reasoning content of the large model more realistic. In addition, target encouragement is set in the prompt words to motivate the large model to provide better answers towards the goal. By using affirmative instructions and incentive strategies, the enthusiasm of the model and the quality of the answer are improved.

[0053] For example, the constructed dynamic prompt words are as follows:

[0054] Prompt(“““

[0055] As a natural language processing conference theme report generation expert, you now generate 10 pieces of data that best match the color of human conference reports based on the sample information below.

[0056] # Sample information

[0057] Conference report content: {CONTENT}

[0058] #Data generation rules

[0059] ##Do not repeat the unit address in the sample information

[0060] ##Do not repeat the names of people in the sample information

[0061] ##Use a text background similar to the sample information

[0062] ##The generated conference report content has a similarity of more than 0.8 with the theme of {THEME}

[0063] Please generate your most human conference report based on the sample information.

[0064] You are the best and look forward to your success.

[0065] ”””)

[0066] Finally, the multiple inference texts corresponding to each event text are added to the event text set to obtain an augmented text set.

[0067] This step uses the format data semantic features of the event text in the historical conference report to construct dynamic prompt words, and generates a large amount of high-quality semantic data through a large model to solve the problem of category imbalance in historical data and enhance data diversity.

[0068] S2. Use the augmented text set to train the dual-channel feature fusion BERT model to obtain the event topic classification model; the dual-channel module in the dual-channel feature fusion BERT model includes a multi-head attention mechanism channel and a text convolution channel.

[0069] It should be noted that if Figure 2 As shown in the figure, the dual-channel feature fusion BERT model includes: BERT network, dual-channel module, feature fusion module and classification prediction module in sequence; among them, the BERT network receives the training text, inputs the last layer of text feature vector into the dual-channel module, and after the feature fusion module fuses the global multi-dimensional features output by the multi-head attention mechanism channel in the dual-channel module and the local key features output by the text convolution channel to obtain the text fusion features, it is passed to the classification prediction module to output the event topic category.

[0070] (1) BERT Network

[0071] Specifically, the BERT network is pre-trained on a public corpus and adopts a multi-layer bidirectional Transformer-Encoder architecture, including word embedding layers, multi-layer Transformer encoders, layer normalization, and residual connections.

[0072] In the embedding layer, the received training text is segmented to obtain a series of tokens, and each token is converted into a word embedding representation, where word embeddings are used to map each word to a vector of fixed dimension, segment embeddings are used to indicate the sentence where the token is located, and positional embeddings are used to retain the position information of the token in the sentence.

[0073] The BERT network uses the Masked Language Model (MLM) as a target, randomly masks 15% of the words in the training sample, and uses the feature representation processed by the bidirectional Transformer-Encoder to generate predictions for the masked words. This step obtains the text feature vector of the last layer of the BERT network, that is, the output of the last layer of the encoder in the multi-layer Transformer encoder, which captures the semantic and syntactic information of each word in the context and contains rich semantic information.

[0074] (2) Dual-channel module

[0075] The dual-channel module includes a multi-head attention mechanism channel and a text convolution channel.

[0076] ①The core of the multi-head attention mechanism channel is to merge the attention matrix generated by multiple attention heads and the corresponding context vector to obtain more comprehensive information. In addition, the multi-head attention mechanism allows the model to capture relevant information from multiple aspects at the same time, thereby enhancing the performance and generalization ability of the model.

[0077] Specifically, after receiving the text feature vector, the multi-head attention mechanism channel passes through the multi-head attention sub-module and the first forward network sub-module in sequence, in which residual connections and layer normalization are added to the multi-head attention sub-module and the first forward network sub-module to output global multi-dimensional features.

[0078] The multi-head attention submodule is used to analyze the text feature vector Perform three linear changes according to the query weight , key weight Sum weight , and get the query matrix , key matrix Sum Matrix , the formula is as follows:

[0079] (1).

[0080] Split the data into In the attention heads, each head processes data independently, and the feature dimensions processed by each head are ,in is the text feature vector feature dimension.

[0081] For each attention head, the dot product method is used to calculate the similarity between each word in the text and other words, and then the attention weight is obtained by scaling and softmax normalization function. , the formula is as follows:

[0082] (2).

[0083] The attention weights and value matrix Multiply them together to get the attention vector. The formula is as follows:

[0084] (3).

[0085] Finally, concatenate the outputs of all multi-head attentions and integrate the information through another linear transformation to obtain the feature vector , the formula is as follows:

[0086] (4)

[0087] in, Represents the weight matrix that integrates the outputs of multiple attention heads.

[0088] After residual connection and layer normalization, the feature vector output by the multi-head attention submodule is for:

[0089] (5)

[0090] in, Representation layer normalization operation.

[0091] Next, after passing through the first forward network submodule, the feature vector output by the first forward network with the ReLU activation function for:

[0092] (6)

[0093] in, and Represents the two-layer weights in the first forward network.

[0094] Finally, after residual connection and layer normalization, the feature vector output by the first forward network submodule is obtained, which is the global multi-dimensional feature output of the multi-head attention mechanism channel. , the formula is as follows:

[0095] (7).

[0096] ② The text convolution channel uses text convolutional neural networks (Text Convolutional Neural Networks, referred to as Text CNN). The text convolutional neural network includes: embedding layer, convolution layer, maximum pooling layer and feature integration layer.

[0097] Specifically, the embedding layer adjusts the dimension of the received text feature vector to obtain an embedding matrix.

[0098] The convolution layer uses convolution kernels to process the embedding matrix, uses m convolution kernels of different sizes and convolution kernel weights to extract local key features, and Activation function is used to better obtain text features. For each convolution kernel, the convolution result features are obtained by the following formula:

[0099] (8)

[0100] in, Represents the convolution result features extracted by the rth convolution kernel starting from position j in the embedding matrix, represents the size of the rth convolution kernel, represents the bias term of the rth convolution kernel, represents the weight of the rth convolution kernel, .

[0101] For example, three convolution kernels of different sizes are used, namely, 2, 3, and 5.

[0102] Furthermore, the maximum pooling layer uses the maximum pooling for each convolution kernel to extract the most important features and reduce the size of the feature representation. The formula is as follows:

[0103] (9)

[0104] in, Represents the maximum pooled feature of the convolution result of the rth convolution kernel.

[0105] Finally, the pooled features from different convolution kernels are integrated into a feature vector, which is the local key feature output of the text convolution channel. :

[0106] (10).

[0107] (3) Feature fusion module

[0108] The feature fusion module passes the received global multidimensional features and local key features to the feature splicing submodule. The feature splicing submodule obtains high-dimensional features through vector splicing, vector alignment subtraction and vector alignment dot multiplication, and then passes them to the second forward network submodule. Residual connection and layer normalization are added to the second forward network submodule to output text fusion features.

[0109] Specifically, the global multidimensional features and local key features reflect the text feature information from different angles and aspects. First, the global multidimensional features and local key features are directly connected through vector splicing to form a higher-dimensional composite vector, thereby improving the comprehensiveness of text understanding.

[0110] Furthermore, the difference features between vectors are further captured through vector subtraction and vector dot multiplication operations. Among them, vector subtraction refers to the subtraction operation of corresponding elements of each dimension of the vectors of global multidimensional features and local key features, emphasizing the difference between global semantic features and local semantic features, thereby extracting information that may be neglected.

[0111] Vector point multiplication is to perform multiplication operations on the corresponding elements of each dimension of the vectors of global multidimensional features and local key features, further strengthening the complementary information between the two vectors. In this operation, if the global semantic feature vector and the local semantic vector of a certain dimension are both high, the product will also be high, indicating that the dimension is more important in text understanding. On the contrary, if one of the vectors is low in a certain dimension, the product will also be reduced accordingly, thereby reducing the impact of this dimension on the model.

[0112] Through the combined operations of vector counterpoint subtraction and vector counterpoint multiplication, the model can effectively capture and distinguish the global and local semantic features of the text, thereby improving the performance in subsequent classification tasks.

[0113] The high-dimensional feature H output by the feature concatenation submodule is expressed by the following formula:

[0114] (11).

[0115] Next, the high-dimensional feature H passes through the second forward network with ReLU activation function, residual connection and layer normalization in the second forward network submodule, and outputs the text fusion feature. The formula is as follows:

[0116] (12).

[0117] in, represents the text fusion feature, represents the weights in the second forward network.

[0118] (4) Classification prediction module

[0119] The classification prediction module inputs the received text fusion features into the fully connected layer for nonlinear operation and then obtains the classification probability through softmax operation, and takes the event topic category corresponding to the maximum classification probability as the prediction result.

[0120] When training the entire dual-channel feature fusion BERT model, the event topic category corresponding to each training sample in the amplified text set is taken as the true label, compared with the predicted event topic category, and the cross entropy loss function is calculated. The AdamW optimizer is used to continuously optimize the parameters of the entire model. After the training, the event topic classification model is obtained.

[0121] After testing, the dual-channel feature fusion BERT model in this embodiment has high performance indicators: the precision rate reached 98.06%, the recall rate reached 97.53%, and the F1 score reached 97.79%.

[0122] S3. Analyze the conference report to be predicted to obtain multiple events to be predicted, pass each event to be predicted into the event topic classification model, and predict the corresponding event topic category.

[0123] According to the file format of the conference report to be predicted, the method of parsing the file in step S1 is used to obtain paragraphs containing topic identifiers to obtain multiple events to be predicted. Each event to be predicted is passed into the event topic classification model, and according to the output probability of each event topic category, the event topic category corresponding to the maximum probability is taken as the event topic category of the conference report to be predicted.

[0124] Compared with the prior art, the event topic classification method based on conference reports provided in this embodiment proposes a dual-channel feature fusion network model based on a pre-trained network. The model combines a multi-head attention mechanism network and a text convolutional network to enhance the BERT text feature representation. The multi-head attention mechanism gives the model the ability to learn the diverse relationships and features in the text in parallel by paying attention to multiple information dimensions at the same time. The text convolutional neural network enhances the model's learning of the key information of local features of the text by paying attention to the local key information of the text, and enhances the model's text feature ability. Finally, the feature fusion technology is used to fuse the semantic features of the dual channels, enhance the text feature ability, and improve the accuracy of event topic category classification. According to the semantic features of the event text extracted from the historical conference report, a dynamic prompt word is constructed, and a large model is used as a means to expand the data set, thereby reducing the reliance on a large amount of manually annotated data, which not only effectively reduces manpower and material resources, but also significantly improves the efficiency and quality of data processing.

[0125] Example 2

[0126] Another embodiment of the present invention discloses an event theme classification system based on conference reports, thereby realizing an event theme classification method based on conference reports in embodiment 1. The specific implementation of each module refers to the corresponding description in embodiment 1. The system includes:

[0127] The text set construction unit is used to parse historical conference reports to obtain an event text set; construct dynamic prompt words for each event text in the event text set and pass them into the large model, and add the corresponding multiple inference texts into the event text set to obtain an augmented text set;

[0128] The classification model training unit is used to train the dual-channel feature fusion BERT model using the augmented text set to obtain an event topic classification model; the dual-channel module in the dual-channel feature fusion BERT model includes a multi-head attention mechanism channel and a text convolution channel;

[0129] The event theme classification unit is used to parse the conference report to be predicted to obtain multiple events to be predicted, pass each event to be predicted into the event theme classification model, and predict the corresponding event theme category.

[0130] Since the event theme classification system based on conference reports in this embodiment and the event theme classification method based on conference reports in the above embodiment can be mutually referenced, it is a repeated description here, so it will not be repeated here. Since the principle of this system embodiment is the same as that of the above method embodiment, this system embodiment also has the corresponding technical effects of the above method embodiment.

[0131] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the computer program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0132] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for classifying event topics based on conference reports, characterized in that: The following steps are involved: Parse historical conference reports to obtain an event text set; construct dynamic prompt words for each event text in the event text set and pass them into the big model, and add the corresponding multiple inference texts into the event text set to obtain an augmented text set; construct dynamic prompt words for each event text in the event text set, including: setting the role and task goal of the prompt words; introducing each event text as sample information of the conference report content through a first placeholder; based on the sample information, set text generation rules and target encouragement, wherein the event theme category marked by the current event text is introduced through a second placeholder; The dual-channel feature fusion BERT model is trained using the augmented text set to obtain an event topic classification model; the dual-channel module in the dual-channel feature fusion BERT model includes a multi-head attention mechanism channel and a text convolution channel; Parse the conference report to be predicted to obtain multiple events to be predicted, pass each event to be predicted into the event topic classification model, and predict the corresponding event topic category; The dual-channel feature fusion BERT model includes: a BERT network, a dual-channel module, a feature fusion module and a classification prediction module in sequence; the BERT network receives training text, inputs the last layer of text feature vectors into the dual-channel module, and after the feature fusion module fuses the global multi-dimensional features output by the multi-head attention mechanism channel in the dual-channel module and the local key features output by the text convolution channel to obtain text fusion features, the features are input into the classification prediction module to output event topic categories; After receiving the text feature vector, the multi-head attention mechanism channel passes through the multi-head attention submodule and the first forward network submodule in sequence, wherein residual connection and layer normalization are added to the multi-head attention submodule and the first forward network submodule to output global multi-dimensional features; The text convolution channel adopts a text convolution neural network, which includes: an embedding layer, a convolution layer, a maximum pooling layer and a feature integration layer in sequence; after the embedding layer adjusts the dimension of the received text feature vector, a plurality of convolution result features are extracted through a plurality of convolution kernels of different sizes in the convolution layer, and the maximum value of each convolution result feature is extracted by the maximum pooling layer and then integrated into a local key feature by the feature integration layer; The feature fusion module passes the received global multi-dimensional features and local key features to the feature splicing submodule, and the feature splicing submodule obtains high-dimensional features through vector splicing, vector alignment subtraction and vector alignment dot multiplication, and then passes them to the second forward network submodule, and the second forward network submodule adds residual connection and layer normalization to output text fusion features; The classification prediction module inputs the received text fusion features into the fully connected layer for nonlinear operation and then performs a softmax operation to obtain the classification probability, and takes the event topic category corresponding to the maximum classification probability as the prediction result.

2. The event topic classification method based on conference reports according to claim 1 is characterized in that: The role of the prompt word is set as a conference theme report generation expert; the text generation rules include: do not have repeated unit addresses in the sample information, do not have repeated personnel names in the sample information, use a text background similar to the sample information, and the generated conference report content has a similarity with the {THEME} theme of more than 0.8; wherein {THEME} is the second placeholder.

3. The event topic classification method based on conference reports according to claim 1 is characterized in that: The large model adopts a 32B four-bit quantized base large model.

4. An event topic classification system based on conference reports, characterized in that: include: A text set construction unit is used to parse historical conference reports to obtain an event text set; construct dynamic prompt words for each event text in the event text set and pass them into a large model, and add the corresponding multiple inference texts into the event text set to obtain an augmented text set; construct dynamic prompt words for each event text in the event text set, including: setting the role and task goal of the prompt words; introducing each event text as sample information of the conference report content through a first placeholder; based on the sample information, setting text generation rules and target encouragement, wherein the event theme category marked by the current event text is introduced through a second placeholder; A classification model training unit, used to train a dual-channel feature fusion BERT model using an augmented text set to obtain an event topic classification model; the dual-channel module in the dual-channel feature fusion BERT model includes a multi-head attention mechanism channel and a text convolution channel; An event theme classification unit is used to parse the conference report to be predicted to obtain multiple events to be predicted, and pass each event to be predicted into the event theme classification model to predict the corresponding event theme category; The dual-channel feature fusion BERT model in the classification model training unit includes: a BERT network, a dual-channel module, a feature fusion module and a classification prediction module in sequence; the BERT network receives the training text, inputs the last layer of text feature vectors into the dual-channel module, and after the feature fusion module fuses the global multi-dimensional features output by the multi-head attention mechanism channel in the dual-channel module and the local key features output by the text convolution channel to obtain the text fusion features, the features are input into the classification prediction module to output the event topic category; After receiving the text feature vector, the multi-head attention mechanism channel passes through the multi-head attention submodule and the first forward network submodule in sequence, wherein residual connection and layer normalization are added to the multi-head attention submodule and the first forward network submodule to output global multi-dimensional features; The text convolution channel adopts a text convolution neural network, which includes: an embedding layer, a convolution layer, a maximum pooling layer and a feature integration layer in sequence; after the embedding layer adjusts the dimension of the received text feature vector, a plurality of convolution result features are extracted through a plurality of convolution kernels of different sizes in the convolution layer, and the maximum value of each convolution result feature is extracted by the maximum pooling layer and then integrated into a local key feature by the feature integration layer; The feature fusion module passes the received global multi-dimensional features and local key features to the feature splicing submodule, and the feature splicing submodule obtains high-dimensional features through vector splicing, vector alignment subtraction and vector alignment dot multiplication, and then passes them to the second forward network submodule, and the second forward network submodule adds residual connection and layer normalization to output text fusion features; The classification prediction module inputs the received text fusion features into the fully connected layer for nonlinear operation and then performs a softmax operation to obtain the classification probability, and takes the event topic category corresponding to the maximum classification probability as the prediction result.

Images