Dialogue abstract generation method based on large language model LLM auxiliary tag and two-dimensional centrality and storage medium thereof
By generating intent labels and performing confidence filtering using a large language model (LLM), and combining two-dimensional centrality calculation, the BART-large model is used to generate dialogue summaries. This solves the problems of topic confusion and omission of key information in multi-topic dialogues, and achieves more accurate and interpretable summary generation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- FUZHOU LIANCHUANG ZHIYUN INFORMATION TECH CO LTD
- Filing Date
- 2026-02-10
- Publication Date
- 2026-05-26
AI Technical Summary
Existing dialogue summarization techniques struggle to identify topic boundaries when handling multi-topic dialogues, leading to topic confusion or omission of key information. Furthermore, they fail to effectively leverage the synergistic advantages of large language models (LLMs) and lightweight models, resulting in insufficient stability and interpretability of the generated summaries.
Intent labels are generated and confidence-based filtering is performed using a large language model (LLM) to construct a high-quality training set. Combined with a two-dimensional centrality calculation method, a BART-large model is used to generate summaries, explicitly modeling topic structure and evaluating discourse importance to achieve accurate dialogue summarization.
It improves the accuracy and completeness of dialogue summaries, enhances the interpretability and scene adaptability of the model, solves the problem of topic confusion in multi-topic dialogues, and achieves more accurate key information filtering and summary generation.
Smart Images
Figure CN122087104A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of natural language processing technology, specifically to a dialogue summarization method and its storage medium based on auxiliary labels and two-dimensional centrality of a large language model (LLM). Background Technology
[0002] With the widespread adoption of instant messaging, online meetings, and customer service interactions, dialogue data has experienced explosive growth. Extracting core information quickly from massive amounts of dialogue has become a key requirement in the field of natural language processing. Dialogue summarization technology, by compressing lengthy, multi-turn conversations into concise, coherent text, can significantly improve information retrieval efficiency and has irreplaceable value in scenarios such as intelligent customer service debriefing, meeting minutes generation, and social chat summarization.
[0003] Existing dialogue summarization technologies mainly fall into three categories, but all have significant limitations: Traditional end-to-end pre-trained models rely on pre-trained models such as BART and T5, directly training and generating models through parallel "dialogue-summary" corpora. This approach depends on the model's implicit learning of dialogue semantics but lacks explicit topic structure modeling capabilities. When frequent topic shifts occur in the dialogue, the model struggles to identify topic boundaries, leading to topic confusion or omission of key subtopics in the summaries. Furthermore, its ability to distinguish the importance of utterances is insufficient, often treating secondary comments and core decision-making information equally, resulting in redundant summaries or missing key information.
[0004] Rule-based and feature-engineered methods generate summaries by manually designing topic segmentation rules and importance features. These methods offer strong interpretability but rely excessively on human experience, exhibiting poor generalization ability and adaptability across different scenarios. Furthermore, hand-designed features struggle to capture deep semantic connections within dialogues, limiting the quality of the summaries to the completeness of the feature design.
[0005] Large Language Model (LLM) direct generation methods: These methods leverage the strong semantic understanding capabilities of large models such as GPT and LLaMA to generate summaries guided by prompts. While these methods perform reasonably well in short dialogue scenarios, they suffer from insufficient stability, with the generation results significantly affected by the wording of prompts and the length of the dialogue. Furthermore, they are computationally expensive, making deployment on resource-constrained edge devices difficult. Additionally, the decision-making process is untraceable, posing application risks in scenarios requiring accountability.
[0006] Furthermore, existing technologies generally fail to fully leverage the synergistic advantages of large-scale language models (LLMs) and lightweight models. They often employ a single dimension in assessing discourse importance, neglecting the local value of discourse within its subtopics, resulting in incomplete subtopic information. Therefore, constructing a dialogue summarization method that can utilize the semantic understanding capabilities of LLMs for structured assistance, achieve accurate summarization through explicit modeling of topic structure and dual-dimensional importance, and simultaneously balance efficiency and interpretability, has become a pressing technical challenge in this field. Summary of the Invention
[0007] To address the aforementioned problems, the present invention aims to provide a dialogue summarization method and its storage medium based on large language model LLM auxiliary tags and two-dimensional centrality, so as to improve the accuracy, completeness and interpretability of dialogue summaries.
[0008] The technical solution of the present invention is as follows: A dialogue summarization method based on large-scale language model LLM auxiliary labels and two-dimensional centrality includes the following steps: Step 1: Generate intent labels for dialogue utterances using a large language model (LLM), and construct a training set (TS) with high-quality labels after confidence filtering; Step 2: Using the pre-trained model as the backbone, train a deep learning network model F based on the training set TS, which integrates the auxiliary labels of the large language model LLM and two-dimensional centrality. The two-dimensional centrality includes global centrality and local centrality. The global centrality is associated with the intention label weight, and the local centrality is associated with the importance of the discourse within the subtopic. Using BART-large as the backbone model, train a deep learning network model F based on the training set TS, which integrates the auxiliary labels of the large language model LLM and two-dimensional centrality. The deep learning network model F includes a dialogue preprocessing submodule, a semantic representation encoding submodule, an intent-driven subtopic segmentation submodule, an intent-driven two-dimensional centrality calculation submodule, a semantic weighted adjustment submodule, and an autoregressive decoding submodule. The dialogue preprocessing submodule performs structured processing on the training data to provide standardized input for subsequent encoding; The semantic representation encoding submodule obtains the semantic vector representation of the discourse through the BART-large encoder; The intent-driven subtopic segmentation submodule segments the dialogue into subtopics based on the filtered intent tag sequence. The intent-driven two-dimensional centrality calculation submodule calculates weights from two dimensions: the coreness of the subtopic in the dialogue and the importance of the discourse in the subtopic. The global centrality calculation introduces a global target vector based on intent label weights. The semantic weighted adjustment submodule optimizes the token-level semantic representation based on two-dimensional weights. The autoregressive decoding submodule generates a summary and completes model training to learn the semantic relationship between the dialogue and the summary; Step 3: Input the dialogue to be processed into the trained deep learning network model F, and output a dialogue summary.
[0009] Furthermore, step 1 specifically involves: Step 1.1: Generate intent labels for dialogue utterances using a large-scale language model (LLM), and construct a high-quality labeled training set (TS) after confidence-based filtering; for the input dialogue text sequence... ,in This represents the i-th utterance, and for each utterance... The context information within the fixed windows before and after the dialogue is combined to generate a bidirectional context-compressed summary. This summary is not used as the final dialogue summary, but only as the input feature for subsequent intent classification. Step 1.2: Define five types of dialogue intent labels: COMMENT, ANSWER, DEVELOP-TOPIC, INTRODUCTION-TOPIC, and CHANGE-TOPIC. COMMENT indicates a comment on the preceding text and does not involve topic transition; ANSWER indicates an answer to the preceding question and does not involve topic transition; DEVELOP-TOPIC indicates the development of a sub-topic of the current topic and does not involve topic transition; INTRODUCTION-TOPIC indicates the introduction of a related but different topic and is considered topic transition; CHANGE-TOPIC indicates a complete change of topic and is considered topic transition. Input the bidirectional context summary of each utterance generated in Step 1 into the intent labeling module based on a large-scale language model (LLM). The intent labeling module uses a nested decision tree discrimination process to output the initial intent label corresponding to each utterance. Step 1.3: Filter the initial intent labels generated in Step 1.2 based on their confidence: Generate each utterance by sampling it independently multiple times, calculate the confidence of the label voting consistency, the model's self-assessment confidence level, and the generated text fuzzy word flags, and calculate the comprehensive confidence based on these; when the comprehensive confidence is lower than the preset threshold, mark the intent label of the utterance as "UNCERTAIN"; Step 1.4: Combine the dialogue text sequence U with the filtered intent label sequence Perform association to form training samples Multiple training samples are collected to construct the training set TS.
[0010] Furthermore, in step 1.1, for each utterance... The text extracts the utterances from the first two rounds of the preceding fixed window and the last three rounds of the following fixed window as context; it then calls the Large Language Model (LLM) and inputs a structured prompt to guide the LLM to analyze the utterances based on the extracted context. The semantic content is compressed and refined to generate a bidirectional contextual compressed summary.
[0011] Furthermore, in step 1.2, the nested decision tree discrimination process is based on the priority order of whether it is a direct response (ANSWER), whether it is a subjective comment (COMMENT), and whether the focus of the dialogue has shifted. If it has not shifted, it is a DEVELOP-TOPIC. If it has shifted, it is further determined whether it is an INTRODUCTION-TOPIC or a CHANGE-TOPIC of the same topic domain.
[0012] Furthermore, in step 1.3, the confidence screening of the initial intent tags specifically involves: generating a tag set by independently sampling each utterance k=3 times under the same prompt template. : ; Voting Consistency Confidence In the formula: ; Among them majority voting labels; Furthermore, each time a large language model is invoked to generate intent labels, the prompt template is designed to require the model to simultaneously output three parts of information: the intent label, the self-assessed confidence level, and the explanatory text generated for that label; the self-assessed confidence level parsed from the model response is denoted as... Simultaneously, fuzzy word detection is performed on the explanatory text generated by the model: if preset fuzzy words appear in the text, a language confidence flag is added. Record it as 0, otherwise record it as 1; The formula for calculating the overall confidence level is: ; Furthermore, step 2 specifically includes: Step 2.1: The dialogue preprocessing submodule processes all training samples in the training set TS, splits the dialogue into speech sequences according to the speech, retains the speech order and numbers, removes meaningless symbols, filters invalid speech, and outputs a structured speech list to provide input for the semantic representation encoding submodule; Step 2.2: The semantic representation encoding submodule uses a pre-trained BART-large model encoder to encode the utterance list output in Step 2.1. It constructs an input sequence by inserting [CLS] tokens before all utterances and [SEP] tokens at the end of each utterance. This input sequence is then padded and truncated by the token segmenter to obtain a token sequence, which is input into the encoder to obtain the vector representation of each token. ; belonging to the same discourse token vector Perform average pooling to obtain the discourse semantic vector representation ; Step 2.3: The intent-driven subtopic segmentation submodule takes the utterance list output in Step 2.1 and the filtered intent tag sequence L obtained in Step 1.4 as input. It traverses from the first utterance. When it encounters an utterance with the intent tag INTRODUCTION-TOPIC or CHANGE-TOPIC, it segments it into a new subtopic segment. For utterances marked "UNCERTAIN", it adds them to the current subtopic segment. Finally, it outputs a list of subtopic segments. : ; in This is the k-th subtopic segment; Step 2.4: The intent-driven two-dimensional centrality calculation submodule calculates the global centrality weight and local centrality weight of each utterance to obtain the joint importance weight; Step 2.5: The semantic weighted adjustment submodule adjusts the vector representation of each token in the corresponding utterance based on the joint importance weights obtained in step B4, to obtain the weighted token vector. ; Step 2.6: The autoregressive decoding submodule uses the weighted vector output in step B5. Compared with the original token vector The weighted sum is used as input to generate a summary through the BART-large decoder; the word probability distribution is obtained using the softmax function, and the model parameters are updated through backpropagation based on the negative log-likelihood loss function; training is terminated when the change in the model loss value is less than a set threshold or the maximum number of iterations is reached.
[0013] Furthermore, step 2.1 specifically includes: Step 2.11: Traverse the training set TS, each training sample is represented as... ; Where U represents the text content of the dialogue, as follows: ; Let n be the i-th utterance, n be the number of utterances, and L be the sequence of intent tags. Step 2.12: For each statement Data cleaning is performed to remove meaningless symbols and filter invalid discourse. The cleaned discourse is then segmented and tagged with parts of speech to obtain a standardized discourse sequence. Step 2.13: Number the standardized utterance sequence according to the original order of the utterances in the dialogue, and output a structured utterance list: ,in This is the i-th standardized discourse; Step 2.2 specifically includes: Step 2.21: Construct the input sequence X of the BART-large encoder: ; in, [CLS] is the sequence start marker, and [SEP] is the discourse separation marker; Step 2.22: Process the input sequence using the BART-large model's tokenizer, unify the sequence length, and truncate excessively long sequences to obtain the token sequence N; The token sequence N is calculated as follows: ; in, This represents the word segmenter for the BART-large model. Step 2.23: Input the token sequence N into the BART-large model encoder to obtain the vector of each token. , ; in, This indicates the BART-large model encoder. Let be the vector corresponding to the t-th token, and d be the dimension of the feature vector; Step 2.24: Determine each normalized utterance based on the positions of the [CLS] and [SEP] markers in the input sequence X. The corresponding token range, and the token vectors within that range. Perform average pooling to obtain the discourse semantic vector representation Discourse semantic vector The calculation is as follows: ; in, d is the dimension of the feature vector. This indicates an average pooling operation; In step 2.3, when a utterance with the intent label INTRODUCTION-TOPIC or CHANGE-TOPIC is encountered, a new subtopic segment is segmented; for utterances marked "UNCERTAIN", they are added to the current subtopic segment, and finally, a list of subtopic segments is output. = .
[0014] Furthermore, step 2.4 specifically includes: Step 2.41: Calculate global centrality, including the following steps: Step 2.411: Calculate each subtopic segment semantic center vector ,Will Semantic vectors of all discourses The formula for calculating the average is: ; in, , This indicates an average pooling operation; Step 2.412: Construct the global dialogue target vector Assign different contribution weights to each utterance based on intent tags. ,in: ; Obtained through weighted pooling The formula is: ; Step 2.413: Using cosine similarity Calculate each subtopic segment and The relevance of the subtopic segment is used as the global centrality of the subtopic segment. The formula is:
[0015] Step 2.414: For all sub-topic segments Perform softmax normalization to obtain the normalized global importance weights. The formula is: ; Step 2.42: Calculate local centrality, including the following steps: Step 2.421: Translate the words semantic vectors Embedding vector of its intent label By concatenating the vectors, a fused vector is obtained. The formula is: ; in, The embedding vector is the vector that transforms the intent tag through the embedding layer. For the "UNCERTAIN" tag, its embedding vector is fixed as a zero vector. ; Step 2.422: For the fused vector Perform a nonlinear transformation to obtain the discourse. Importance score The formula is: ; in, For a trainable parameter matrix, For bias terms, Use the Sigmoid activation function; Step 2.423: For the same subtopic segment Score for all statements Softmax normalization is performed to obtain the local attention weights of the discourse within the subtopic. The formula is: ; Step 2.424: Using local attention weights Weighted summation yields subtopic segments Structural perception center vector The formula is: ; in, ; Step 2.425: Calculate discourse cosine similarity As discourse Local centrality The formula is: ; Step 2.426: For the same subtopic segment The local scores of all utterances are Softmax normalized to obtain probabilistic local importance weights. The formula is: ; Step 2.43: Joint weight calculation: Words Global centrality weight of the subtopic segment Local centrality weight of discourse itself Multiply, and you get words. final joint weight The formula is: ; In step 2.5, the weighted token vector By concatenating all weighted token vectors in the original order of the token sequence, a complete weighted token vector sequence is obtained. ; Weighted token vector Represented as: ; in, , The original vector of the token.
[0016] The complete weighted token vector sequence Represented as: , where p is the length of the token sequence; Step 2.6 specifically involves: Step 2.61: Construct the decoder input Decoder_Input, which is a sequence of weighted token vectors. With the original token vector sequence The weighted fusion formula is as follows:
[0017] in, This is a hyperparameter used to control the degree of influence of the weighted vector; Step 2.62: Input the decoder The input is taken into the BART-large model decoder, and the decoder performs decoding through a self-attention mechanism, layer normalization, and feedforward neural network to obtain the decoded output representation; Step 2.63: Perform linear transformation and softmax normalization on the decoded output representation to obtain the probability distribution for generating the next word; Step 2.64: The model uses negative log-likelihood as the loss function to calculate the loss value, and defines the loss value. for:
[0018] in, This represents the first t-1 words that have been generated. Generate the conditional probability of the t-th word for the decoder; Step 2.65: Update the model parameters using the Adam optimizer, calculate the gradient using backpropagation and iteratively optimize; terminate the training process when the change in the loss value of the model in each iteration is less than the set threshold or the maximum number of iterations is reached.
[0019] Furthermore, step 3 specifically includes the following steps: Step 3.1: Preprocess the user dialogue to be processed, following the same steps as in Step 2.1, including utterance splitting, data cleaning, standardization, and outputting a structured utterance list. ; Step 3.2: List of statements Generate a bidirectional context-compressed summary sequence and a filtered intent tag sequence using LLM. ; Step 3.3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require Semantic encoding is performed using the method in step 2.2 to obtain the discourse semantic vector. ; Step 3.4: Based on the discourse list and Calculate the subtopic segments, two-dimensional centrality weights, and weighted token vector sequences using the methods in steps 2.3 to 2.5. ; Step 3.5: Weight the token vector sequence With the original token vector sequence The method in step 2.61 is used to fuse the input into the decoder of the trained deep learning network model F, and the output is a summary of the user dialogue.
[0020] A computer storage medium includes a memory, a processor, and computer program instructions stored in the memory and executable by the processor, wherein when the processor executes the computer program instructions, it can implement the above-described method steps.
[0021] Compared with the prior art, the present invention has the following beneficial effects: This invention addresses the issue of topic confusion in multi-topic dialogues by introducing an LLM-assisted intent annotation and confidence filtering mechanism. It utilizes high-quality intent tags to explicitly delineate topic boundaries, achieving more accurate topic structure modeling. By designing an intent-driven, two-dimensional centrality weight calculation method that considers both the coreness of the subtopic and the local contribution of the discourse within subtopics, it overcomes the shortcomings of traditional single-centrality methods, achieving more comprehensive and accurate key information filtering. Employing a collaborative architecture of "LLM offline generation of structured tags + lightweight, fine-tunable BART model executing core generation tasks," it effectively utilizes LLM's deep semantic understanding capabilities while avoiding its direct summary generation shortcomings, achieving more efficient and controllable use of LLM capabilities. Subtopic division is based on interpretable intent tag rules, and the two-dimensional weight calculation process is transparent, enhancing the model's interpretability and meeting the transparency requirements of decision-making processes in key scenarios. Through training on multi-scenario dialogue data and a universal five-category intent tag design, the model exhibits stronger scenario adaptability and generalization ability, stably generating coherent and complete summaries across various dialogue types. Attached Figure Description
[0022] Figure 1 This is an architecture diagram (one) of a dialogue summarization model based on LLM auxiliary tags and two-dimensional centrality in an embodiment of the present invention. Figure 2 This is the second architectural diagram of the dialogue summarization model based on LLM auxiliary tags and two-dimensional centrality in an embodiment of the present invention. Figure 3 This is the third architectural diagram of the dialogue summarization model based on LLM auxiliary labels and two-dimensional centrality in an embodiment of the present invention. Detailed Implementation
[0023] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.
[0024] See Figure 1-3 A dialogue summarization method based on large-scale language model LLM auxiliary labels and two-dimensional centrality includes the following steps: Step 1: Generate intent labels for dialogue utterances using a large language model (LLM), and construct a training set (TS) with high-quality labels after confidence filtering; Step 2: Using the pre-trained model as the backbone, train a deep learning network model F based on the training set TS, which integrates the auxiliary labels of the large language model LLM and two-dimensional centrality. The two-dimensional centrality includes global centrality and local centrality. The global centrality is associated with the intention label weight, and the local centrality is associated with the importance of the discourse within the subtopic. Using BART-large as the backbone model, train a deep learning network model F based on the training set TS, which integrates the auxiliary labels of the large language model LLM and two-dimensional centrality. The deep learning network model F includes a dialogue preprocessing submodule, a semantic representation encoding submodule, an intent-driven subtopic segmentation submodule, an intent-driven two-dimensional centrality calculation submodule, a semantic weighted adjustment submodule, and an autoregressive decoding submodule. The dialogue preprocessing submodule performs structured processing on the training data to provide standardized input for subsequent encoding; The semantic representation encoding submodule obtains the semantic vector representation of the discourse through the BART-large encoder; The intent-driven subtopic segmentation submodule segments the dialogue into subtopics based on the filtered intent tag sequence. The intent-driven two-dimensional centrality calculation submodule calculates weights from two dimensions: the coreness of the subtopic in the dialogue and the importance of the discourse in the subtopic. The global centrality calculation introduces a global target vector based on intent label weights. The semantic weighted adjustment submodule optimizes the token-level semantic representation based on two-dimensional weights. The autoregressive decoding submodule generates a summary and completes model training to learn the semantic relationship between the dialogue and the summary; Step 3: Input the dialogue to be processed into the trained deep learning network model F, and output a dialogue summary.
[0025] The present invention will be further described below with reference to a specific embodiment: A dialogue summarization method based on large-scale language model LLM auxiliary labels and two-dimensional centrality includes the following steps: Step 1: Construct a training set TS with high-quality labels; Step 1.1: Generate intent labels for dialogue utterances using a large-scale language model (LLM), and construct a high-quality labeled training set (TS) after confidence-based filtering; for the input dialogue text sequence... ,in This represents the i-th utterance, and for each utterance... The context information within the fixed windows before and after the dialogue is combined to generate a bidirectional context-compressed summary. This summary is not used as the final dialogue summary, but only as the input feature for subsequent intent classification. For each statement The text extracts the utterances from the first two rounds of the preceding fixed window and the last three rounds of the following fixed window as context; it then calls the Large Language Model (LLM) and inputs a structured prompt to guide the LLM to analyze the utterances based on the extracted context. The semantic content is compressed and refined to generate a bidirectional contextual compressed summary.
[0026] Step 1.2: Define five types of dialogue intent labels: COMMENT, ANSWER, DEVELOP-TOPIC, INTRODUCTION-TOPIC, and CHANGE-TOPIC. COMMENT indicates a comment on the preceding text and does not involve topic transition; ANSWER indicates an answer to the preceding question and does not involve topic transition; DEVELOP-TOPIC indicates the development of a sub-topic of the current topic and does not involve topic transition; INTRODUCTION-TOPIC indicates the introduction of a related but different topic and is considered topic transition; CHANGE-TOPIC indicates a complete change of topic and is considered topic transition. Input the bidirectional context summary of each utterance generated in Step 1 into the intent labeling module based on a large-scale language model (LLM). The intent labeling module uses a nested decision tree discrimination process to output the initial intent label corresponding to each utterance. The nested decision tree discrimination process determines intent according to the priority order of "whether it is a direct response (ANSWER), whether it is a subjective comment (COMMENT), whether the focus of the dialogue has shifted (if not, it is a DEVELOP-TOPIC; if it has shifted, it is further determined whether it is an INTRODUCTION-TOPIC or a CHANGE-TOPIC)".
[0027] Step 1.3: Filter the initial intent labels generated in Step 1.2 based on their confidence: Generate each utterance by sampling it independently multiple times, calculate the confidence of the label voting consistency, the model's self-assessment confidence level, and the generated text fuzzy word flags, and calculate the comprehensive confidence based on these; when the comprehensive confidence is lower than the preset threshold, mark the intent label of the utterance as "UNCERTAIN"; The confidence screening of the initial intent tags is specifically performed as follows: for each utterance, k=3 independent samples are generated under the same prompt template to produce a tag set. : ; Voting Consistency Confidence In the formula: ; Among them The majority of voting labels; furthermore, each time the Large Language Model (LLM) is invoked to generate intent labels, we design the prompt template to require the model to simultaneously output three parts of information: the intent label, the self-rated confidence level (1–5 points), and the explanatory text generated for that label. The self-rated confidence level parsed from the model response is denoted as... Simultaneously, we perform fuzzy word detection on the explanatory text generated by the model: if the text contains preset fuzzy words (such as "maybe", "possibly", "uncertain", etc.), then the language confidence flag is added. Record it as 0, otherwise record it as 1.
[0028] The formula for calculating the overall confidence level is: ; Step 1.4: Combine the dialogue text sequence U with the filtered intent label sequence Perform association to form training samples Multiple training samples are collected to construct the training set TS.
[0029] Step 2: Train the deep learning network model F; Step 2.1: The dialogue preprocessing submodule processes all training samples in the training set TS, splits the dialogue into speech sequences according to the speech, retains the speech order and numbers, removes meaningless symbols, filters invalid speech, and outputs a structured speech list to provide input for the semantic representation encoding submodule; Further, step 2.11: Traverse the training set TS, each training sample is represented as... ; Where U represents the text content of the dialogue, as follows: ; Let n be the i-th utterance, n be the number of utterances, and L be the sequence of intent tags. Step 2.12: For each statement Data cleaning is performed to remove meaningless symbols and filter out invalid discourse. The cleaned discourse is then segmented and tagged with parts of speech to obtain a standardized discourse sequence. Step 2.13: Number the standardized utterance sequence according to the original order of the utterances in the dialogue, and output a structured utterance list: ,in Let this be the i-th standardized discourse; Step 2.2: The semantic representation encoding submodule uses a pre-trained BART-large model encoder to encode the utterance list output in Step 2.1. It constructs an input sequence by inserting [CLS] tokens before all utterances and [SEP] tokens at the end of each utterance. This input sequence is then padded and truncated by the token segmenter to obtain a token sequence, which is input into the encoder to obtain the vector representation of each token. ; belonging to the same discourse token vector Perform average pooling to obtain the discourse semantic vector representation
[0030] Further, step 2.21: Construct the input sequence X of the BART-large encoder: ;
[0031] in, [CLS] is the sequence start marker, and [SEP] is the discourse separation marker; Step 2.22: Process the input sequence using the BART-large model's tokenizer, unify the sequence length, and truncate excessively long sequences to obtain the token sequence N; The token sequence N is calculated as follows: ;
[0032] in, This represents the word segmenter for the BART-large model. Step 2.23: Input the token sequence N into the BART-large model encoder to obtain the vector of each token. , ;
[0033] in, This indicates the BART-large model encoder. Let be the vector corresponding to the t-th token, and d be the dimension of the feature vector; Step 2.24: Determine each normalized utterance based on the positions of the [CLS] and [SEP] markers in the input sequence X. The corresponding token range, and the token vectors within that range. Perform average pooling to obtain the discourse semantic vector representation Discourse semantic vector The calculation is as follows: ;
[0034] in, d is the dimension of the feature vector. This indicates an average pooling operation; Step 2.3: The intent-driven subtopic segmentation submodule takes the utterance list output in Step 2.1 and the filtered intent tag sequence L obtained in Step 1.4 as input. It traverses from the first utterance. When it encounters an utterance with the intent tag INTRODUCTION-TOPIC or CHANGE-TOPIC, it segments it into a new subtopic segment. For utterances marked "UNCERTAIN", it adds them to the current subtopic segment. Finally, it outputs a list of subtopic segments. : ; in This is the k-th subtopic segment; Subtopic segmentation is performed based on the intent label sequence L, and the specific algorithm is as follows: enter: dialogue ; Intent label sequence ; Output: Subtopic Collection = ; 1: Initialize the first subtopic Its starting words are Set k=1; 2: fori=2 to ndo; 3: For INTRODUCTION-TOPIC or CHANGE-TOPICthen; 4: k = k + 1; 5: with To begin, create a new subtopic. ; 6: ∈ANSWER,COMMENTorDEVELOP-TOPICthen; 7: will Add to current subtopic ; 8:elseif, if For UNCERTAINthen; 9: will Add to current subtopic (Not used as the starting point for new subtopics); 10: End if; 11:endfor; 12: Return to the collection of subtopics Sub= ;
[0035] When encountering utterances with the intent label INTRODUCTION-TOPIC or CHANGE-TOPIC, a new subtopic segment is created; for utterances marked "UNCERTAIN", they are added to the current subtopic segment, and finally, a list of subtopic segments is output. = .
[0036] Step 2.4: The intent-driven two-dimensional centrality calculation submodule calculates the global centrality weight and local centrality weight of each utterance to obtain the joint importance weight; Furthermore, step 2.41: calculating global centrality includes the following steps: Step 2.411: Calculate each subtopic segment semantic center vector ,Will Semantic vectors of all discourses The formula for calculating the average is: ;
[0037] in, , This indicates an average pooling operation; Step 2.412: Construct the global dialogue target vector Assign different contribution weights to each utterance based on intent tags. ,in: ;
[0038] Obtained through weighted pooling The formula is: ;
[0039] Step 2.413: Using cosine similarity Calculate each subtopic segment and The relevance of the subtopic segment is used as the global centrality of the subtopic segment. The formula is: ;
[0040] Step 2.414: For all sub-topic segments Perform softmax normalization to obtain the normalized global importance weights. The formula is: ;
[0041] Step 2.42: Calculate local centrality, including the following steps: Step 2.421: Translate the words semantic vectors Embedding vector of its intent label By concatenating the vectors, a fused vector is obtained. The formula is: ;
[0042] in, The embedding vector is the vector that transforms the intent tag through the embedding layer. For the "UNCERTAIN" tag, its embedding vector is fixed as a zero vector. ; Step 2.422: For the fused vector Perform a nonlinear transformation to obtain the discourse. Importance score The formula is: ;
[0043] in, For a trainable parameter matrix, For bias terms, Use the Sigmoid activation function; Step 2.423: For the same subtopic segment Score for all statements Softmax normalization is performed to obtain the local attention weights of the discourse within the subtopic. The formula is: ;
[0044] Step 2.424: Using local attention weights Weighted summation yields subtopic segments Structural perception center vector The formula is: ;
[0045] in, ; Step 2.425: Calculate discourse and cosine similarity As discourse Local centrality The formula is: ;
[0046] Step 2.426: For the same subtopic segment The local scores of all utterances are Softmax normalized to obtain probabilistic local importance weights. The formula is: ;
[0047] Step 2.43: Joint weight calculation: Words Global centrality weight of the subtopic segment Local centrality weight of discourse itself Multiply, and you get words. final joint weight The formula is: ; Step 2.5: The semantic weighted adjustment submodule adjusts the vector representation of each token in the corresponding utterance based on the joint importance weights obtained in step B4, to obtain the weighted token vector. ; Weighted token vector By concatenating all weighted token vectors in the original order of the token sequence, a complete weighted token vector sequence is obtained. ; Weighted token vector Represented as: ; in, , The original vector of the token. ; The complete weighted token vector sequence Represented as: , where p is the length of the token sequence; Step 2.6: The autoregressive decoding submodule uses the weighted vector output in step B5. Compared with the original token vector The weighted sum is used as input to generate a summary through the BART-large decoder; the word probability distribution is obtained using the softmax function, and the model parameters are updated through backpropagation based on the negative log-likelihood loss function; training is terminated when the change in the model loss value is less than a set threshold or the maximum number of iterations is reached.
[0048] Furthermore, step 2.61 involves constructing the decoder input Decoder_Inputt, which is a sequence of weighted token vectors. With the original token vector sequence The weighted fusion formula is as follows: ; in, This is a hyperparameter used to control the degree of influence of the weighted vector; Step 2.62: Input the decoder The input to the BART-large model decoder is decoded through a self-attention mechanism, layer normalization, and a feedforward neural network to obtain the decoded output representation; Step 2.63: Perform linear transformation and softmax normalization on the decoded output representation to obtain the probability distribution for generating the next word; Step 2.64: The model uses negative log-likelihood as the loss function to calculate the loss value, and defines the loss value. for: ; in, This represents the first t-1 words that have been generated. Generate the conditional probability of the t-th word for the decoder; Step 2.65: Update the model parameters using the Adam optimizer, calculate the gradient using backpropagation and iteratively optimize; terminate the training process when the change in the loss value of the model in each iteration is less than the set threshold or the maximum number of iterations is reached.
[0049] Step 3: Generate a dialogue summary; Step 3.1: Preprocess the user dialogue to be processed, following the same steps as in Step 2.1, including utterance splitting, data cleaning, standardization, and outputting a structured utterance list. ; Step 3.2: List of statements Generate a bidirectional context-compressed summary sequence and a filtered intent tag sequence using LLM. ; Step 3.3: List of statements Semantic encoding is performed using the method in step 2.2 to obtain the discourse semantic vector. ; Step 3.4: Based on the discourse list and discourse semantic vector Calculate the subtopic segments, two-dimensional centrality weights, and weighted token vector sequences using the methods in steps 2.3 to 2.5. ; Step 3.5: Weight the token vector sequence With the original token vector sequence Following the method in step 2.61, the decoder of the trained deep learning network model F is input to the fusion process, and the output is a summary of the user dialogue.
[0050] This embodiment also provides a computer storage medium, including a memory, a processor, and computer program instructions stored in the memory and executable by the processor. When the processor executes the computer program instructions, it can implement the above-described method steps.
[0051] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.
Claims
1. A dialogue summarization method based on auxiliary labels of a large-scale language model (LLM) and two-dimensional centrality, characterized in that, Includes the following steps: Step 1: Generate intent labels for dialogue utterances using a large language model (LLM), and construct a training set (TS) with high-quality labels after confidence filtering; Step 2: Using the pre-trained model as the backbone, train a deep learning network model F based on the training set TS, which integrates the auxiliary labels of the large language model LLM and two-dimensional centrality. The two-dimensional centrality includes global centrality and local centrality. The global centrality is associated with the intention label weight, and the local centrality is associated with the importance of the discourse within the subtopic. Using BART-large as the backbone model, train a deep learning network model F based on the training set TS, which integrates the auxiliary labels of the large language model LLM and two-dimensional centrality. The deep learning network model F includes a dialogue preprocessing submodule, a semantic representation encoding submodule, an intent-driven subtopic segmentation submodule, an intent-driven two-dimensional centrality calculation submodule, a semantic weighted adjustment submodule, and an autoregressive decoding submodule. The dialogue preprocessing submodule performs structured processing on the training data to provide standardized input for subsequent encoding; The semantic representation encoding submodule obtains the semantic vector representation of the discourse through the BART-large encoder; The intent-driven subtopic segmentation submodule segments the dialogue into subtopics based on the filtered intent tag sequence. The intent-driven two-dimensional centrality calculation submodule calculates weights from two dimensions: the coreness of the subtopic in the dialogue and the importance of the discourse in the subtopic. The global centrality calculation introduces a global target vector based on intent label weights. The semantic weighted adjustment submodule optimizes the token-level semantic representation based on two-dimensional weights. The autoregressive decoding submodule generates a summary and completes model training to learn the semantic relationship between the dialogue and the summary; Step 3: Input the dialogue to be processed into the trained deep learning network model F, and output a dialogue summary.
2. The dialogue summarization method based on large language model LLM auxiliary tags and two-dimensional centrality as described in claim 1, characterized in that, Step 1 is as follows: Step 1.1: Generate intent labels for dialogue utterances using a large language model (LLM), and construct a training set TS with high-quality labels by filtering based on confidence. For the input dialogue text sequence, for the input dialogue text sequence ,in This represents the i-th utterance, and for each utterance... The context information within the fixed windows before and after the dialogue is combined to generate a bidirectional context-compressed summary. This summary is not used as the final dialogue summary, but only as the input feature for subsequent intent classification. Step 1.2: Define five types of dialogue intent labels: COMMENT, ANSWER, DEVELOP-TOPIC, INTRODUCTION-TOPIC, and CHANGE-TOPIC. COMMENT indicates a comment on the preceding text and does not involve topic transition; ANSWER indicates an answer to the preceding question and does not involve topic transition; DEVELOP-TOPIC indicates the development of a sub-topic of the current topic and does not involve topic transition; INTRODUCTION-TOPIC indicates the introduction of a related but different topic and is considered topic transition; CHANGE-TOPIC indicates a complete change of topic and is considered topic transition. Input the bidirectional context summary of each utterance generated in Step 1 into the intent labeling module based on a large-scale language model (LLM). The intent labeling module uses a nested decision tree discrimination process to output the initial intent label corresponding to each utterance. Step 1.3: Filter the initial intent labels generated in Step 1.2 based on their confidence: Generate each utterance by sampling it independently multiple times, calculate the confidence of the label voting consistency, the model's self-assessment confidence level, and the generated text fuzzy word markers, and calculate the comprehensive confidence based on these; when the comprehensive confidence is lower than the preset threshold, mark the intent label of the utterance as "UNCERTAIN"; Step 1.4: Combine the dialogue text sequence U with the filtered intent label sequence Perform association to form training samples Multiple training samples are collected to construct the training set TS.
3. The dialogue summarization method based on large-scale language model LLM auxiliary tags and two-dimensional centrality as described in claim 2, characterized in that, In step 1.1, for each utterance... Extract the utterances from the first two rounds of the fixed window preceding it and the last three rounds of the fixed window following it as context; Invoke the Large Language Model (LLM) and input a structured prompt to guide the LLM to extract contextual dialogue. The semantic content is compressed and refined to generate a bidirectional contextual compressed summary.
4. The dialogue summarization method based on large language model LLM auxiliary tags and two-dimensional centrality as described in claim 2, characterized in that, In step 1.2, the nested decision tree discrimination process is based on the priority order of whether it is a direct response (ANSWER), whether it is a subjective comment (COMMENT), and whether the focus of the dialogue has shifted. If it has not shifted, it is a DEVELOP-TOPIC. If it has shifted, it is further judged whether it is an INTRODUCTION-TOPIC or a CHANGE-TOPIC of different topic domains to determine the intent.
5. The dialogue summarization method based on large language model LLM auxiliary tags and two-dimensional centrality as described in claim 2, characterized in that, In step 1.3, the confidence screening of the initial intent tags specifically involves: generating a tag set by independently sampling each utterance k=3 times under the same prompt template. : ； Voting Consistency Confidence In the formula: ； Among them majority voting labels; Furthermore, each time a large language model is invoked to generate intent labels, the prompt template is designed to require the model to simultaneously output three parts of information: the intent label, the self-assessed confidence level, and the explanatory text generated for that label; the self-assessed confidence level parsed from the model response is denoted as... Simultaneously, fuzzy word detection is performed on the explanatory text generated by the model: if preset fuzzy words appear in the text, a language confidence flag is added. Record it as 0, otherwise record it as 1; The formula for calculating the overall confidence level is: 。 6. The dialogue summarization method based on large language model LLM auxiliary tags and two-dimensional centrality according to claim 2, characterized in that, Step 2 specifically involves: Step 2.1: The dialogue preprocessing submodule processes all training samples in the training set TS, splits the dialogue into speech sequences according to the speech, retains the speech order and numbers, removes meaningless symbols, filters invalid speech, and outputs a structured speech list to provide input for the semantic representation encoding submodule; Step 2.2: The semantic representation encoding submodule uses a pre-trained BART-large model encoder to encode the utterance list output in Step 2.
1. It constructs an input sequence by inserting [CLS] tokens before all utterances and [SEP] tokens at the end of each utterance. This input sequence is then padded and truncated by the token segmenter to obtain a token sequence, which is input into the encoder to obtain the vector representation of each token. ; For those belonging to the same discourse token vector Perform average pooling to obtain the discourse semantic vector representation ; Step 2.3: The intent-driven subtopic segmentation submodule takes the utterance list output in Step 2.1 and the filtered intent tag sequence L obtained in Step 1.4 as input. It traverses from the first utterance. When it encounters an utterance with the intent tag INTRODUCTION-TOPIC or CHANGE-TOPIC, it segments it into a new subtopic segment. For utterances marked "UNCERTAIN", it adds them to the current subtopic segment. Finally, it outputs a list of subtopic segments. : ； in This is the k-th subtopic segment; Step 2.4: The intent-driven two-dimensional centrality calculation submodule calculates the global centrality weight and local centrality weight of each utterance to obtain the joint importance weight; Step 2.5: The semantic weighted adjustment submodule adjusts the vector representation of each token in the corresponding utterance based on the joint importance weights obtained in step B4, thus obtaining the weighted token vector. ; Step 2.6: The autoregressive decoding submodule uses the weighted vector output in step B5. Compared with the original token vector The weighted sum is used as input to generate a summary through the BART-large decoder; the word probability distribution is obtained using the softmax function, and the model parameters are updated through backpropagation based on the negative log-likelihood loss function; training is terminated when the change in the model loss value is less than a set threshold or the maximum number of iterations is reached.
7. The dialogue summarization method based on large language model LLM auxiliary tags and two-dimensional centrality as described in claim 6, characterized in that, Step 2.1 specifically includes: Step 2.11: Traverse the training set TS, each training sample is represented as... ; Where U represents the text content of the dialogue, as follows: ; Let n be the i-th utterance, n be the number of utterances, and L be the sequence of intent tags. Step 2.12: For each statement Data cleaning is performed to remove meaningless symbols and filter out invalid discourse. The cleaned discourse is then segmented and tagged with parts of speech to obtain a standardized discourse sequence. Step 2.13: Number the standardized utterance sequence according to the original order of the utterances in the dialogue, and output a structured utterance list: ,in For the i-th standardized discourse; Step 2.2 specifically includes: Step 2.21: Construct the input sequence X of the BART-large encoder: ； in, [CLS] is the sequence start marker, and [SEP] is the discourse separation marker; Step 2.22: Process the input sequence using the BART-large model's tokenizer, unify the sequence length, and truncate excessively long sequences to obtain the token sequence N; The token sequence N is calculated as follows: ； in, This represents the word segmenter for the BART-large model. Step 2.23: Input the token sequence N into the BART-large model encoder to obtain the vector of each token. , ； in, This indicates the BART-large model encoder. Let be the vector corresponding to the t-th token, and d be the dimension of the feature vector; Step 2.24: Determine each normalized utterance based on the positions of the [CLS] and [SEP] markers in the input sequence X. The corresponding token range, and the token vectors within that range. Perform average pooling to obtain the discourse semantic vector representation Discourse semantic vector The calculation is as follows: ； in, d is the dimension of the feature vector. This indicates an average pooling operation; In step 2.3, when a utterance with the intent label INTRODUCTION-TOPIC or CHANGE-TOPIC is encountered, a new subtopic segment is segmented; for utterances labeled "UNCERTAIN", they are added to the current subtopic segment, and finally, a list of subtopic segments is output. .
8. The dialogue summarization method based on large language model LLM auxiliary tags and two-dimensional centrality according to claim 7, characterized in that, Step 2.4 specifically includes: Step 2.41: Calculate global centrality, including the following steps: Step 2.411: Calculate each subtopic segment semantic center vector ,Will Semantic vectors of all discourses in the text The formula for calculating the average is: ； in, , This indicates an average pooling operation; Step 2.412: Construct the global dialogue target vector Assign different contribution weights to each utterance based on intent tags. ,in: ； Obtained through weighted pooling The formula is: ； Step 2.413: Using cosine similarity Calculate each subtopic segment and Relevance, as the global centrality of sub-topic segments The formula is: ； Step 2.414: For all sub-topic segments Perform softmax normalization to obtain the normalized global importance weights. The formula is: ； Step 2.42: Calculate local centrality, including the following steps: Step 2.421: Translate the words semantic vectors Embedding vector of its intent label By concatenating the vectors, a fused vector is obtained. The formula is: ； in, The embedding vector is the vector that transforms the intent tag through the embedding layer. For the "UNCERTAIN" tag, its embedding vector is fixed as a zero vector. ; Step 2.422: For the fused vector Perform a nonlinear transformation to obtain the discourse. Importance score The formula is: ； in, For a trainable parameter matrix, For bias terms, Use the Sigmoid activation function; Step 2.423: For the same subtopic segment Score for all statements Softmax normalization is performed to obtain the local attention weights of the discourse within the subtopic. The formula is: ； Step 2.424: Using local attention weights Weighted summation yields subtopic segments Structural perception center vector The formula is: ； in, ; Step 2.425: Calculate discourse cosine similarity As discourse Local centrality The formula is: ； Step 2.426: For the same subtopic segment The local scores of all utterances are Softmax normalized to obtain probabilistic local importance weights. The formula is: ； Step 2.43: Joint weight calculation: Words Global centrality weight of the subtopic segment Local centrality weight of discourse itself Multiply, and you get words. final joint weight The formula is: ； In step 2.5, the weighted token vector By concatenating all weighted token vectors in the original order of the token sequence, a complete weighted token vector sequence is obtained. ; Weighted token vector Represented as: ； in, , The original vector of the token. ; The complete weighted token vector sequence Represented as: , where p is the length of the token sequence; Step 2.6 specifically involves: Step 2.61: Construct the decoder input Decoder_Input, which is a sequence of weighted token vectors. With the original token vector sequence The weighted fusion formula is as follows: ； in, This is a hyperparameter used to control the degree of influence of the weighted vector; Step 2.62: Input the decoder The input to the BART-large model decoder is decoded through a self-attention mechanism, layer normalization, and a feedforward neural network to obtain the decoded output representation; Step 2.63: Perform linear transformation and softmax normalization on the decoded output representation to obtain the probability distribution for generating the next word; Step 2.64: The model uses negative log-likelihood as the loss function to calculate the loss value, and defines the loss value. for: ； in, This represents the first t-1 words that have been generated. Generate the conditional probability of the t-th word for the decoder; Step 2.65: Update the model parameters using the Adam optimizer, calculate the gradient using backpropagation and iteratively optimize; terminate the training process when the change in the loss value of the model in each iteration is less than the set threshold or the maximum number of iterations is reached.
9. The dialogue summarization method based on large-scale language model LLM auxiliary tags and two-dimensional centrality as described in claim 8, characterized in that, Step 3 specifically includes the following steps: Step 3.1: Preprocess the user dialogue to be processed, following the same steps as in Step 2.1, including utterance splitting, data cleaning, standardization, and outputting a structured utterance list. ; Step 3.2: List of statements Generate a bidirectional context-compressed summary sequence and a filtered intent tag sequence using LLM. ; Step 3.3: List of statements Semantic encoding is performed using the method in step 2.2 to obtain the discourse semantic vector. ; Step 3.4: Based on the discourse list and Calculate the subtopic segments, two-dimensional centrality weights, and weighted token vector sequences using the methods in steps 2.3 to 2.
5. ; Step 3.5: Order the weighted token vector sequence With the original token vector sequence The method in step 2.61 is used to fuse the input into the decoder of the trained deep learning network model F, and the output is a summary of the user dialogue.
10. A computer storage medium, characterized in that, It includes a memory, a processor, and computer program instructions stored in the memory and executable by the processor, which, when executed by the processor, enable the implementation of the steps of the method as described in any one of claims 1-9.