Topic-enhanced and dialog-centric summarization of dialogues

Abstract

This invention relates to a dialogue summarization method and system based on topic enhancement and dialogue centering. The method includes: extracting user dialogue and user dialogue summaries, labeling them, and constructing a training set; using the training set and an embedded topic model to train a deep learning network model M based on topic enhancement and dialogue centering. Model M obtains the utterance-level representation and utterance-level topic feature representation of the dialogue, and uses these as input. A feature-aware meta-network is used to remove noise, multi-head attention is used to capture the semantic relationships between features, and a gating mechanism is used for filtering and fusion to obtain a topic-enhanced dialogue context feature representation, which is then weighted to obtain the final representation; the final representation is used as input to generate a dialogue summary to train the model, thereby learning the semantic relationship between the user dialogue and the user dialogue summary; the user dialogue is input into the trained model M, and a summary of the user dialogue is output. This method and system are beneficial for improving the accuracy of dialogue summarization.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a dialogue summarization method and system based on topic enhancement and dialogue center. Background Technology

[0002] Dialogue is one of the most common forms of communication on social networks, and summarizing or abstracting dialogue content has become an important task. The purpose of dialogue summarization is to extract key information from the discourse of two or more participants, as perceived by a third party. This task not only helps provide new participants with a quick context but also helps people grasp the central idea or search for key content, thereby improving efficiency and productivity. In recent years, social networks have become increasingly intertwined with people's lives, enabling them to communicate, make friends, work, learn, and entertain themselves. With the ever-increasing number of users, the amount of communication data generated by social networks is also growing rapidly. Dialogue summarization techniques can be used to extract the main information from dialogues, allowing for quick and efficient understanding of relevant information. However, unlike news or articles in traditional document summaries, dialogue data has unique characteristics, including different language styles, fragmented information, flexible discourse structures, and unclear topic boundaries, presenting numerous challenges for dialogue summarization.

[0003] Early work on dialogue summarization addressed this problem by converting dialogues into word graphs and selecting appropriate paths within the graph as summary sentences based on complex rules. Template-based methods were also used, collecting templates from human-written summaries and generating abstract summaries by filling in blanks with appropriate words selected from the dialogue. However, the summaries they generated lacked fluency and diversity, thus proving far from practical use. With the advent of neural encoder-decoder models, input can be projected into dense semantic representations, and summaries with new words can be generated by progressively sampling from the vocabulary until a special marker indicating the end of a representation is output. Based on these models, generative text summarization has made significant progress, from untrained PGN, Fast-Abs, and HRED to pretrained models like BART and Pegasus, driving the development of abstract text summarization.

[0004] Injecting preprocessing features involves explicitly utilizing additional features from the dialogue context, using human annotations or external tagging tools as part of the input. These preprocessing features include acquiring intra-discourse features and inter-discourse features.

[0005] In-discourse features can be categorized into word-level, phrase-level, and discourse-level features. Word-level features include TF-IDF weights, part-of-speech (POS) tags, and named entity tags. Phrase-level features include key phrases / keywords, which emphasize prominent n-grams in the original dialogue, helping to overcome the challenge of information dispersion and thus yielding more detailed summaries. Discourse-level features enhance summaries through features such as speaker or role, user intent, and dialogue behavior. Lin et al. proposed two interactive methods to extract key information from other roles to enhance role-oriented dialogue summarization. Kim et al. used the COMET and PARACOMET commonsense knowledge models to extract commonsense knowledge from dialogues, injecting this knowledge into state-of-the-art language models (e.g., BART) for abstract dialogue summarization.

[0006] Inter-utterance features can be categorized into partitioning and graph methods. Partitioning involves extracting or segmenting the entire dialogue into relatively independent fragments. Information within each fragment is more concentrated, resulting in less interference with summary generation. Simultaneously, these features reduce GPU memory requirements with shorter input lengths, making them particularly suitable for long dialogue summarization. Krishna et al.'s Cluster2Sent extracts important utterances, clusters related utterances together, and generates a summary sentence for each cluster, yielding a semi-structured summary suitable for clinical conversations. Apart from Shang et al.'s unsupervised use of K-means for utterance clustering, most methods require labeled oracle partitions for supervised training of the extractor or classifier. Typically, partitions are concatenated as input to an abstract summarization model, or the summaries generated from each segment are concatenated or ordered to form the final summary. Summary of the Invention

[0007] The purpose of this invention is to provide a dialogue summarization method and system based on topic enhancement and dialogue center, which helps to improve the accuracy of dialogue summarization.

[0008] To achieve the above objectives, the technical solution adopted by this invention is: a dialogue summarization method based on topic enhancement and dialogue center, comprising the following steps:

[0009] Step A: Extract user dialogues and user dialogue summaries, and label the user dialogues and user dialogue summaries. Each complete user dialogue corresponds to a user dialogue summary, and a training set DS is constructed.

[0010] Step B: Train a deep learning network model M based on topic enhancement and dialogue centering using the training set DS and the embedded topic model ETM. The encoder of the deep learning network model M includes an input submodule, a topic feature extraction submodule, a center selection submodule, and a feature dynamic fusion submodule. The deep learning network model M obtains the utterance-level representation of the dialogue text through the input submodule and mines the potential topic feature information in the dialogue text through the topic feature extraction submodule to obtain the utterance-level topic feature representation. The feature dynamic fusion submodule takes the utterance-level representation and the utterance-level topic feature representation as input, uses a feature-aware meta-network to remove noise to enhance the feature representation ability, uses multi-head attention to capture the semantic correlation between features, and finally uses a gating mechanism for filtering and fusion to obtain the topic-enhanced dialogue context feature representation. The center selection submodule uses an unsupervised clustering method to calculate the center weight of the utterance and assigns weights to the topic-enhanced dialogue context feature representation to obtain the final representation. The decoder of the deep learning network model M uses the final representation as input to generate a summary of the dialogue text to train the model, thereby learning the semantic relationship between user dialogue and user dialogue summary.

[0011] Step C: Input the user dialogue into the trained deep learning network model M and output a summary of the user dialogue.

[0012] Furthermore, step B specifically includes the following steps:

[0013] Step B1: Initialize and encode all training samples in the training set DS to obtain the sentence [mask] sequence M, token sequence N, and utterance-level representation H of the dialogue. u ;

[0014] Step B2: Combine the token sequence N obtained in Step B1 with the utterance-level representation H u Extracting word-level topic feature representations H using an externally trained Embedded Topic Model (ETM). wt Aligned discourse-level topic feature representations H are obtained through stacking and dimensional expansion. ut ;

[0015] Step B3: Combine the sentence [mask] sequence M obtained in Step B1 with the discourse-level representation H u The discourse centers C of the dialogue are obtained using clustering tools, and the degree centrality score S of each discourse in the dialogue is calculated. s Based on the degree centrality score of the discourse, the central discourse (U) and the redundant discourse (N) are selected and divided. The central weight W of each utterance is calculated using the degree centrality score. c ;

[0016] Step B4: Convert the utterance-level representation H obtained in Step B1 and Step B2 into a single utterance-level representation. u Discourse-level thematic feature representation H ut By dynamically learning and adjusting the fusion weights of discourse-level representation and discourse-level topic feature representation through a meta-network, a weighted discourse-level representation is obtained. and discourse-level thematic features representation Using multi-head attention to fuse weighted discourse-level representations and weighted discourse-level topic feature representation Obtain the feature representation H with topic enhancement t Then, the feature representation H of the topic enhancement is obtained through a gating mechanism. t And discourse level representation H u Filtering and fusion are performed to obtain topic-enhanced dialogue context feature representation H. dt ;

[0017] Step B5: Calculate the center weight W for each utterance obtained in Steps B3 and B4. c H and topic-enhanced dialogue context feature representation dt The discourse center representation H obtained through weighted fusion is thematically enhanced. topic The final representation H is obtained by weighted summation and concatenation. final ;

[0018] Step B6: Convert the final representation H obtained in step B5 into a single representation. final The input is fed into the decoder, where it is decoded through a self-attention mechanism, layer normalization, topic attention, and a feedforward neural network to obtain the decoded output representation.

[0019] Step B7: The decoded output obtained in B6 is represented by the softmax function to obtain the probability distribution P∈[0,1] of generating the next word; then, according to the target loss function loss, the gradient of each parameter in the deep learning network model is calculated by the backpropagation method, and the parameters are updated by the stochastic gradient descent method.

[0020] Step B8: When the change in loss value generated by the deep learning network model in each iteration is less than the set threshold, or when the maximum number of iterations is reached, terminate the training process of the deep learning network model.

[0021] Furthermore, step B1 specifically includes the following steps:

[0022] Step B11: Traverse the training set DS and process the dialogue text in it into sentences to obtain the sentence [mask] sequence M. Each training sample in the training set DS is represented as ds = (D, S); where D is the text content of the dialogue and S is the summary text content of the dialogue.

[0023] Dialogue D is represented as:

[0024] D = [u1, u2, ..., u n ]

[0025] Among them, u t Let t be the t-th sentence in the dialogue, where t = 1, 2, ..., n, and n is the number of sentences in the dialogue.

[0026] words u t Represented as:

[0027]

[0028] in, Let i be the i-th word in the t-th sentence, i = 1, 2, ..., m, where m is the utterance u. t The length of the text;

[0029] Step B12: Convert the dialogue D = [u1, u2, ..., u] obtained in step B11. n Concatenate each sentence in the sequence, inserting a [CLS] marker before each sentence and a [SEP] marker between each sentence to construct the initial dialogue sequence X. d ;

[0030] The initial sequence of the dialogue is represented as:

[0031] X d =[CLS]+u1+[SEP]+u2+[SEP]+…+u n

[0032] Where n represents the number of utterances in the dialogue;

[0033] Step B13: Place X d The tokenizer of the BART model performs padding and length truncation to obtain the dialogue token sequence N, which is then used as input to the BART model encoder. Based on the [CLS] and [SEP] tags, the tokens are then used in X. d The position in the sequence is used to segment the feature representation of each sentence, thereby obtaining the feature representation of each sentence.

[0034] The dialogue token sequence N is calculated as follows:

[0035] N = BART tokenizer (X d )

[0036] Among them, BART tokenizer () represents the tokenizer of the BART model;

[0037] Discourse-level representation Hu The calculation is as follows:

[0038]

[0039]

[0040]

[0041] Among them, BART enc () represents the BART model encoder. For the i-th word in the u-th sentence The corresponding word vectors, l is the text length of the nth sentence, and d is the dimension of the feature vector.

[0042] Furthermore, step B2 specifically includes the following steps:

[0043] Step B21: Using the token sequence N obtained in step B13, extract word-level topic feature representation H through an externally trained Embedded Topic Model (ETM). wt Determine if the word corresponding to each token in the token sequence N exists in the vocabulary obtained by ETM training. If it exists, obtain its word-level topic feature representation H through ETM. wt ;

[0044] Word-level topic feature representation H wt The calculation is as follows:

[0045]

[0046] in, This represents the topic feature representation of the i-th word in the token sequence N;

[0047] Step B22: Utilize the word-level topic feature representation H obtained in step B21 wt The discourse-level representation H obtained in step B13 u Aligned discourse-level topic feature representations H are obtained through stacking and dimensional expansion. ut ;

[0048] Discourse-level topic feature representation H ut The calculation formula is as follows:

[0049]

[0050] Where func(·) represents the dimension expansion function, Let d represent the i-th sentence containing n words, where d is the dimension of the feature vector.

[0051] Furthermore, step B3 specifically includes the following steps:

[0052] Step B31: For the sentence [mask] sequence M obtained in step B11 and the discourse-level representation H obtained in step B13... u Clustering tools were used to obtain the discourse centers C and the center distance matrix D of the dialogue. c ;

[0053] Discourse center C and center distance matrix D c The calculation is as follows:

[0054]

[0055]

[0056] Among them, C k It indicates the central theme of the discourse in the dialogue. Let represent the i-th utterance, l be the utterance length, and d be the dimension of the feature vector; in the center distance matrix D c The element d in the i-th row and j-th column ij Let i represent the distance from utterance i to center j, where i∈[1,n] and k∈[1,k].

[0057] Step B32: Based on the obtained discourse center C and discourse level representation H u Calculate the degree centrality score S for each utterance in the dialogue. s Based on the discourse centrality score, the central discourse (U) and non-central redundant discourse are selected and divided.

[0058] The degree centrality score of discourse S s The calculation formula is as follows:

[0059]

[0060] The central discourse is divided as follows:

[0061]

[0062] Where mean(·) is the mean function;

[0063] Step B33: Calculate the central weight W for each utterance based on its centrality score and the central utterance. c ;

[0064] The central weight of discourse W c The calculation is as follows:

[0065]

[0066] Here, α and β are adjustable weight parameters.

[0067] Furthermore, step B4 specifically includes the following steps:

[0068] Step B41: For the utterance-level representation H obtained in steps B13 and B22 u Discourse-level thematic feature representation H ut By constructing a meta-network to learn and adjust the fusion weights of discourse-level representation and discourse-level topic feature representation, a weighted discourse-level representation is obtained. and discourse-level thematic features representation

[0069] Weighted discourse representation and discourse-level thematic features representation The calculation is as follows:

[0070]

[0071]

[0072]

[0073]

[0074]

[0075]

[0076]

[0077]

[0078] Where MLP stands for Multilayer Perceptron, mean(·) is the mean function, and n is the number of utterances in the dialogue;

[0079] Step B42: Fuse weighted utterance-level representations using a multi-head attention mechanism. Weighted topic feature representation It can better capture the complex relationship between discourse and topic features, resulting in topic-enhanced feature representation H. t ;

[0080] Theme-enhanced feature representation H t The calculation is as follows:

[0081]

[0082]

[0083]

[0084] In this context, MHA represents multi-head attention mechanism, concat(·) represents vector concatenation operation, num is the number of parallel heads, Q(query), K(key), and V(value) are the query term, key term, and value term in the attention mechanism, respectively, and Q is the weighted utterance representation. Both K and V are weighted topic feature representations. l is the text length of the utterance, and d is the dimension of the feature vector; For trainable attention parameters, head i This represents the output of the i-th attention function, where i ∈ [1, num].

[0085] Step B43: Enhance the feature representation H of the topic through a gating mechanism t And discourse level representation H u Filtering and fusion are performed to obtain topic-enhanced dialogue context feature representation H. dt ;

[0086] Topic-enhanced dialogue context feature representation H dt The calculation is as follows:

[0087]

[0088]

[0089] In this context, ":" indicates a concatenation operation. l represents the utterance length, d represents the dimension of the feature vector, and W1 and W2 represent the parameters of the trainable gating network.

[0090] Furthermore, step B5 specifically includes the following steps:

[0091] Step B51: Calculate the discourse center weights W obtained in steps B33 and B43. c H and topic-enhanced dialogue context feature representation dt The discourse center representation H obtained through weighted fusion is thematically enhanced. topic ;

[0092] Theme-enhanced discourse center representation H topic The calculation is as follows:

[0093]

[0094] in, l is the utterance length, and d is the dimension of the feature vector;

[0095] Step B52: Enhance the discourse center representation of the obtained topic H topicWith discourse-level representation H u Re-weight and sum the results, then concatenate each sentence to obtain the final representation H. final ;

[0096] Ultimately, H represents final The calculation is as follows:

[0097]

[0098]

[0099] in, p is the text length of the dialogue, d is the dimension of the feature vector, and n is the number of utterances. To define the hyperparameters, concat(·) represents the vector concatenation operation.

[0100] Furthermore, the specific implementation method of step B6 is as follows:

[0101] The final representation H obtained in step B52 final The input is fed into the decoder, and the decoder performs decoding through a self-attention mechanism, layer normalization, topic attention, and feedforward neural network to obtain the decoded output representation H”’;

[0102] H' = SelfAttention(LayNorm(H) final ))

[0103] H"=TopicAttention(LayNorm(H'+H final ))

[0104] H″′=FFN(LayNorm(H′+H”))

[0105] Where LayNorm(·) represents layer normalization, SelfAttention(·) represents self-attention mechanism, TopicAttention(·) represents topic attention, and FFN(·) represents feedforward neural network.

[0106] Furthermore, step B7 specifically includes the following steps:

[0107] Step B71: Normalize the decoded output representation H”' obtained in step B61 using the softmax function to obtain the score distribution P of the generated next word;

[0108] The formula for calculating the score distribution P is as follows:

[0109] P = softmax(W3H″′ + b)

[0110] Where P∈[0,1], W3 is the trainable parameter matrix, and b is the bias vector;

[0111] Step B72: The model calculates the loss value using cross-entropy as the loss function, updates the learning rate using the Adam gradient optimization algorithm, and iteratively updates the model parameters using backpropagation to train the model by minimizing the loss function.

[0112] The formula for minimizing the loss function L is as follows:

[0113]

[0114] Wherein, P(y t |y <t D; θ) represents the corresponding word y taken from P. t The probability value, | represents the conditional part of the conditional probability, that is, given y <t Given the dialogue text D and model parameters θ, generate the target word y. t The probability of y <t =(y1,y2,…,y t-1 ) represents the first t-1 words of the output sequence, S y This represents a summary of the target.

[0115] The present invention also provides a dialogue summarization system based on topic enhancement and dialogue center, 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.

[0116] Compared with existing technologies, this invention offers the following advantages: To effectively identify key information in dialogue and incorporate additional feature information, this invention proposes a novel topic-enhanced context encoder, which consists of a topic feature extraction submodule, a dynamic feature fusion submodule, and a center selection submodule. In the topic feature extraction submodule, this invention introduces an embedded topic model (ETM) to discover interpretable latent topic features from the dialogue. In the dynamic feature fusion submodule, a feature-aware meta-network is designed to reduce noise and enhance feature representations. Then, multi-head attention is used to capture semantic relevance between features, and a gating mechanism is employed for filtering and fusion, thereby enhancing the effective fusion between topic features and dialogue representations. The center selection submodule utilizes an unsupervised method to weight utterances, guiding the model to identify key utterances, thus improving the effectiveness of understanding the dialogue context. Experimental results on two public benchmark datasets demonstrate that the proposed method achieves significant performance improvements over existing state-of-the-art baseline models. Attached Figure Description

[0117] Figure 1 This is a flowchart illustrating the method implementation of an embodiment of the present invention;

[0118] Figure 2 This is an architecture diagram of a deep learning network model M based on topic enhancement and dialogue center in an embodiment of the present invention. Detailed Implementation

[0119] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0120] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0121] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0122] like Figure 1 As shown, this embodiment provides a dialogue summarization method based on topic enhancement and dialogue center, including the following steps:

[0123] Step A: Extract user dialogues and user dialogue summaries, and label the user dialogues and user dialogue summaries. Each complete user dialogue corresponds to a user dialogue summary, and a training set DS is constructed.

[0124] Step B: Train a deep learning network model M based on topic enhancement and dialogue centering using the training set DS and the embedded topic model ETM. The encoder of the deep learning network model M includes an input submodule, a topic feature extraction submodule, a center selection submodule, and a feature dynamic fusion submodule. The deep learning network model M obtains the utterance-level representation of the dialogue text through the input submodule and mines the potential topic feature information in the dialogue text through the topic feature extraction submodule to obtain the utterance-level topic feature representation. The feature dynamic fusion submodule realizes the effective fusion of topic features. This module takes the utterance-level representation and the utterance-level topic feature representation as input, uses a feature-aware meta-network to remove noise to enhance the feature representation ability, uses multi-head attention to capture the semantic correlation between features, and finally uses a gating mechanism for filtering and fusion to obtain the topic-enhanced dialogue context feature representation. The center selection submodule uses an unsupervised clustering method to calculate the center weight of the utterance and assigns weights to the topic-enhanced dialogue context feature representation to obtain the final representation. The decoder of the deep learning network model M uses the final representation as input to generate a summary of the dialogue text to train the model, thereby learning the semantic relationship between user dialogue and user dialogue summary.

[0125] Step C: Input the user dialogue into the trained deep learning network model M and output a summary of the user dialogue.

[0126] In this embodiment, the model structure of the deep learning network model M based on topic enhancement and dialogue center is as follows: Figure 2 As shown. In this embodiment, the specific implementation steps of step B are as follows.

[0127] Step B1: Initialize and encode all training samples in the training set DS to obtain the sentence [mask] sequence M, token sequence N, and utterance-level representation H of the dialogue. u .

[0128] In this embodiment, step B1 specifically includes the following steps:

[0129] Step B11: Traverse the training set DS and process the dialogue text in it into sentences to obtain the sentence [mask] sequence M. Each training sample in the training set DS is represented as ds = (D, S); where D is the text content of the dialogue and S is the summary text content corresponding to the dialogue.

[0130] Dialogue D is represented as:

[0131] D = [u1, u2, ..., u n ]

[0132] Among them, u tLet t be the t-th sentence in the dialogue, where t = 1, 2, ..., n, and n is the number of sentences in the dialogue.

[0133] words u t Represented as:

[0134]

[0135] in, Let i be the i-th word in the t-th sentence, i = 1, 2, ..., m, where m is the utterance u. t The length of the text.

[0136] Step B12: Convert the dialogue D = [u1, u2, ..., u] obtained in step B11. n Concatenate each sentence in the sequence, inserting a [CLS] marker before each sentence and a [SEP] marker between each sentence to construct the initial dialogue sequence X. d .

[0137] The initial sequence of the dialogue is represented as:

[0138] X d =[CLS]+u1+[SEP]+u2+[SEP]+…+u n

[0139] Where n represents the number of utterances in the dialogue.

[0140] Step B13: Place X d The tokenizer of the BART model performs padding and length truncation to obtain the dialogue token sequence N, which is then used as input to the BART model encoder. Based on the [CLS] and [SEP] tags, the tokens are then used in X. d The position in the sequence is used to segment the feature representation of each sentence, thereby obtaining the feature representation of each sentence.

[0141] The dialogue token sequence N is calculated as follows:

[0142] N = BART tokenizer (X d )

[0143] Among them, BART tokenizer () represents the tokenizer of the BART model.

[0144] Discourse-level representation H u The calculation is as follows:

[0145]

[0146]

[0147]

[0148] Among them, BART enc () represents the BART model encoder. For the i-th word in the u-th sentence The corresponding word vectors, l is the text length of the nth sentence, and d is the dimension of the feature vector.

[0149] Step B2: Combine the token sequence N obtained in Step B1 with the utterance-level representation H u Extracting word-level topic feature representations H using an externally trained Embedded Topic Model (ETM). wt Aligned discourse-level topic feature representations H are obtained through stacking and dimensional expansion. ut .

[0150] In this embodiment, step B2 specifically includes the following steps:

[0151] Step B21: Using the token sequence N obtained in step B13, extract word-level topic feature representation H through an externally trained Embedded Topic Model (ETM). wt Determine if the word corresponding to each token in the token sequence N exists in the vocabulary obtained by ETM training. If it exists, obtain its word-level topic feature representation H through ETM. wt .

[0152] Word-level topic feature representation H wt The calculation is as follows:

[0153]

[0154] in, This represents the topic feature representation of the i-th word in the token sequence N.

[0155] Step B22: Utilize the word-level topic feature representation H obtained in step B21 wt The discourse-level representation H obtained in step B13 u Aligned discourse-level topic feature representations H are obtained through stacking and dimensional expansion. ut .

[0156] Discourse-level topic feature representation H ut The calculation formula is as follows:

[0157]

[0158] Where func(·) represents the dimension expansion function, Let d represent the i-th sentence containing n words, where d is the dimension of the feature vector.

[0159] Step B3: Combine the sentence [mask] sequence M obtained in Step B1 with the discourse-level representation H u The discourse centers C of the dialogue are obtained using clustering tools, and the degree centrality score S of each discourse in the dialogue is calculated. s Based on the degree centrality score of the discourse, the central discourse (U) and the redundant discourse (N) are selected and divided. The central weight W of each utterance is calculated using the degree centrality score. c .

[0160] In this embodiment, step B3 specifically includes the following steps:

[0161] Step B31: For the sentence [mask] sequence M obtained in step B11 and the discourse-level representation H obtained in step B13... u Clustering tools were used to obtain the discourse centers C and the center distance matrix D of the dialogue. c .

[0162] Discourse center C and center distance matrix D c The calculation is as follows:

[0163]

[0164]

[0165] Among them, C k It indicates the central theme of the discourse in the dialogue. Let represent the i-th utterance, l be the utterance length, and d be the dimension of the feature vector; in the center distance matrix D c The element d in the i-th row and j-th column ij Let i represent the distance from utterance i to center j, where i∈[1,n] and j∈[1,k].

[0166] Step B32: Based on the obtained discourse center C and discourse level representation H u Calculate the degree centrality score S for each utterance in the dialogue. s Based on the discourse centrality score, the central discourse (U) and non-central redundant discourse are selected and divided.

[0167] The degree centrality score of discourse S s The calculation formula is as follows:

[0168]

[0169] The central discourse is divided as follows:

[0170]

[0171] Where mean(·) is the mean function.

[0172] Step B33: Calculate the central weight W for each utterance based on its centrality score and the central utterance. c .

[0173] The central weight of discourse W c The calculation is as follows:

[0174]

[0175] Here, α and β are adjustable weighting parameters used to balance the influence of centrality scores and whether a statement is central.

[0176] Step B4: Convert the utterance-level representation H obtained in Step B1 and Step B2 into a single utterance-level representation. u Discourse-level thematic feature representation H ut By dynamically learning and adjusting the fusion weights of discourse-level representation and discourse-level topic feature representation through a meta-network, a weighted discourse-level representation is obtained. and discourse-level thematic features representation Using multi-head attention to fuse weighted discourse-level representations and weighted discourse-level topic feature representation Obtain the feature representation H with topic enhancement t Then, the feature representation H of the topic enhancement is obtained through a gating mechanism. t And discourse level representation H u Filtering and fusion are performed to obtain topic-enhanced dialogue context feature representation H. dt .

[0177] In this embodiment, step B4 specifically includes the following steps:

[0178] Step B41: For the utterance-level representation H obtained in steps B13 and B22 u Discourse-level thematic feature representation H ut By constructing a meta-network to learn and adjust the fusion weights of discourse-level representation and discourse-level topic feature representation, a weighted discourse-level representation is obtained. and discourse-level thematic features representation

[0179] Weighted discourse representation and discourse-level thematic features representation The calculation is as follows:

[0180]

[0181]

[0182]

[0183]

[0184]

[0185]

[0186]

[0187]

[0188] Where MLP stands for Multilayer Perceptron, mean(·) is the mean function, and n is the number of utterances in the dialogue.

[0189] Step B42: Fuse weighted utterance-level representations using a multi-head attention mechanism. Weighted topic feature representation It can better capture the complex relationship between discourse and topic features, resulting in topic-enhanced feature representation H. t .

[0190] Theme-enhanced feature representation H t The calculation is as follows:

[0191]

[0192]

[0193]

[0194] In this context, MHA represents multi-head attention mechanism, concat(·) represents vector concatenation operation, num is the number of parallel heads, Q(query), K(key), and V(value) are the query term, key term, and value term in the attention mechanism, respectively, and Q is the weighted utterance representation. Both K and V are weighted topic feature representations. l is the text length of the utterance, and d is the dimension of the feature vector; For trainable attention parameters, head i Let i represent the output of the i-th attention function, where i ∈ [1, num].

[0195] Step B43: Enhance the feature representation H of the topic through a gating mechanism t And discourse level representation H u Filtering and fusion are performed to obtain topic-enhanced dialogue context feature representation H. dt .

[0196] The topic-enhanced dialogue context feature representation is calculated as follows:

[0197]

[0198]

[0199] In this context, ":" indicates a concatenation operation. l represents the utterance length, d represents the dimension of the feature vector, and W1 and W2 represent the parameters of the trainable gating network.

[0200] Step B5: Calculate the center weight W for each utterance obtained in Steps B3 and B4. c H and topic-enhanced dialogue context feature representation dt The discourse center representation H obtained through weighted fusion is thematically enhanced. topic The final representation H is obtained by weighted summation and concatenation. final .

[0201] In this embodiment, step B5 specifically includes the following steps:

[0202] Step B51: Calculate the discourse center weights W obtained in steps B33 and B43. c H and topic-enhanced dialogue context feature representation dt The discourse center representation H obtained through weighted fusion is thematically enhanced. topic .

[0203] Theme-enhanced discourse center representation H topic The calculation is as follows:

[0204]

[0205] in, l represents the utterance length, and d represents the dimension of the feature vector.

[0206] Step B52: Enhance the discourse center representation of the obtained topic H topic With discourse-level representation H u Re-weight and sum the results, then concatenate each sentence to obtain the final representation H. final .

[0207] Ultimately, H represents final The calculation is as follows:

[0208]

[0209]

[0210] in, p is the text length of the dialogue, d is the dimension of the feature vector, and n is the number of utterances. To define the hyperparameters, cocat(·) represents the vector concatenation operation.

[0211] Step B6: Convert the final representation H obtained in step B5 into a single representation. final The input is fed into the decoder, where it is decoded through a self-attention mechanism, layer normalization, topic attention, and a feedforward neural network to obtain the decoded output representation.

[0212] In this embodiment, the specific implementation method of step B6 is as follows:

[0213] The final representation H obtained in step B52 final The input is fed into the decoder, and the decoder performs decoding through a self-attention mechanism, layer normalization, topic attention, and feedforward neural network to obtain the decoded output representation H”’;

[0214] H' = SelfAttention(LayNorm(H) final ))

[0215] H"=TopicAttention(LayNorm(H'+H final ))

[0216] H″′=FFN(LayNorm(H′+H″))

[0217] Where LayNorm(·) represents layer normalization, SelfAttention(·) represents self-attention mechanism, TopicAttention(·) represents topic attention, and FFN(·) represents feedforward neural network.

[0218] Step B7: The decoded output obtained in B6 is represented by the softmax function to obtain the probability distribution P∈[0,1] for generating the next word; then, according to the target loss function loss, the gradient of each parameter in the deep learning network model is calculated by the backpropagation method, and the parameters are updated by the stochastic gradient descent method.

[0219] In this embodiment, step B7 specifically includes the following steps:

[0220] Step B71: Normalize the decoded output representation H”' obtained in step B61 using the softmax function to obtain the score distribution P of the generated next word.

[0221] The formula for calculating the score distribution P is as follows:

[0222] P = softmax(W3H″′ + b)

[0223] Where P∈[0,1], W3 is the trainable parameter matrix, and b is the bias vector.

[0224] Step B72: The model calculates the loss value using cross-entropy as the loss function, updates the learning rate using the Adam gradient optimization algorithm, and iteratively updates the model parameters using backpropagation to train the model by minimizing the loss function.

[0225] The formula for minimizing the loss function L is as follows:

[0226]

[0227] Wherein, P(y t |y <t D; θ) represents the corresponding word y taken from P. t The probability value, | represents the conditional part of the conditional probability, that is, given t <y Given the dialogue text D and model parameters θ, generate the target word y. t The probability of y <t =(y1,y2,…,y t-1 ) represents the first t-1 words of the output sequence, S y This represents a summary of the target.

[0228] Step B8: When the change in loss value generated by the deep learning network model in each iteration is less than the set threshold, or when the maximum number of iterations is reached, terminate the training process of the deep learning network model.

[0229] This embodiment also provides a dialogue summarization system based on topic enhancement and dialogue center, 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.

[0230] The proposed model was compared with other dialogue summarization baseline models, and the evaluation results are shown in Table 1. On the SAMSum benchmark dataset, the proposed model improved the ROUGE-1, ROUGE-2, and ROUGE-L indices by 0.36%, 0.89%, and 0.33%, respectively, compared to the SDDS model. On the DialogSum benchmark dataset, the proposed model improved the ROUGE-1, ROUGE-2, and ROUGE-L indices by 0.62%, 1.22%, and 1.05%, respectively, compared to the SDDS model. On both benchmark datasets, the proposed model significantly outperformed other baseline models, including the state-of-the-art SDDS model.

[0231] Table 1: Evaluation results on the SAMsum and DialogSum datasets

[0232]

[0233]

[0234] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0235] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0236] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0237] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0238] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A dialogue summarization method based on topic enhancement and dialogue center, characterized in that, Includes the following steps: Step A: Extract user dialogues and user dialogue summaries, and label the user dialogues and user dialogue summaries. Each complete user dialogue corresponds to a user dialogue summary, and construct the training set DS. Step B: Train a deep learning network model M based on topic enhancement and dialogue center using the training set DS and the embedded topic model ETM. The encoder of the deep learning network model M includes an input submodule, a topic feature extraction submodule, a center selection submodule, and a feature dynamic fusion submodule. The deep learning network model M obtains the discourse-level representation of the dialogue text through the input submodule and mines the potential topic feature information in the dialogue text through the topic feature extraction submodule to obtain the discourse-level topic feature representation. The feature dynamic fusion submodule takes utterance-level representation and utterance-level topic feature representation as input, uses a feature-aware meta-network to remove noise to enhance feature representation capabilities, utilizes multi-head attention to capture semantic correlations between features, and finally uses a gating mechanism to filter and fuse to obtain topic-enhanced dialogue context feature representation. The center selection submodule uses an unsupervised clustering method to calculate the center weight of the discourse and assigns weights to the topic-enhanced dialogue context feature representation to obtain the final representation; the decoder of the deep learning network model M uses the final representation as input to generate a summary of the dialogue text to train the model, thereby learning the semantic relationship between the user dialogue and the user dialogue summary. Step C: Input the user dialogue into the trained deep learning network model M and output a summary of the user dialogue.

2. The dialogue summarization method based on topic enhancement and dialogue center as described in claim 1, characterized in that, Step B specifically includes the following steps: Step B1: Initialize and encode all training samples in the training set DS to obtain the sentence [mask] sequence of the dialogue. token sequence and discourse level expression ; Step B2: Take the token sequence obtained in Step B1 and discourse level expression Extracting word-level topic feature representations using an externally trained Embedded Topic Model (ETM). Aligned discourse-level topic feature representations are obtained through stacking and dimensional expansion. ; Step B3: Take the sentence [mask] sequence obtained in Step B1. and discourse level expression Identifying the discourse centers of a dialogue using clustering tools Calculate the degree centrality score for each utterance in the dialogue. The central discourse is selected based on its degree centrality score. Non-centralized redundant discourse The center weight of each utterance is calculated using the degree centrality score. ; Step B4: Combine the utterance-level representations obtained in Step B1 and Step B2 Discourse-level thematic feature representation By dynamically learning and adjusting the fusion weights of discourse-level representation and discourse-level topic feature representation through a meta-network, a weighted discourse-level representation is obtained. and discourse-level thematic features representation Utilizing multi-head attention to fuse weighted discourse-level representations and weighted discourse-level topic feature representation Obtaining topic-enhanced feature representations Then, the feature representation of the topic enhancement is obtained through a gating mechanism. and discourse level expression Filtering and fusion are performed to obtain topic-enhanced dialogue context feature representations. ; Step B5: Calculate the center weight of each utterance obtained in Steps B3 and B4. Theme-enhanced dialogue context feature representation Thematically enhanced discourse center representation obtained through weighted fusion The final representation is obtained by weighted summation and concatenation. ; Step B6: The final representation obtained in step B5 The input is fed into the decoder, where it is decoded through a self-attention mechanism, layer normalization, topic attention, and a feedforward neural network to obtain the decoded output representation. Step B7: Use the decoded output representation obtained in B6 to obtain the probability distribution for generating the next word through the softmax function. Then, based on the target loss function, the gradients of each parameter in the deep learning network model are calculated using the backpropagation method, and the parameters are updated using the stochastic gradient descent method. Step B8: When the change in loss value generated by the deep learning network model in each iteration is less than the set threshold, or when the maximum number of iterations is reached, terminate the training process of the deep learning network model.

3. The dialogue summarization method based on topic enhancement and dialogue center as described in claim 2, characterized in that, Step B1 specifically includes the following steps: Step B11: Traverse the training set DS and segment the dialogue text into sentences to obtain the sentence [mask] sequence. Each training sample in the training set DS is represented as ;in, The text content of the dialogue. This is a summary text of the dialogue. dialogue Represented as: in, For the first in the dialogue In short, , The number of utterances contained in the dialogue; words Represented as: in, For the first The first in the sentence One word, , For words The length of the text; Step B12: Obtain the dialogue from step B11 Each sentence in the text is pieced together, and an insertion is made at the beginning. Mark and insert between sentences Marking, constructing the initial sequence of dialogue ; The initial sequence of the dialogue is represented as: in, Indicates the number of utterances in a dialogue; Step B13: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the The tokenizer of the BART model performs padding and length truncation to obtain the token sequence of the dialogue. As input to the BART model encoder, based on the [CLS] and [SEP] labels... The position in the sequence is used to segment the feature representation of each sentence, thereby obtaining the feature representation of each sentence. ; Dialogue token sequence The calculation is as follows: in, This refers to the tokenizer in the BART model. Discourse-level representation The calculation is as follows: in, This represents the BART model encoder. For the first The i-th word in the sentence The corresponding word vectors, , For the first The text length of a sentence is the dimension of the feature vector.

4. The dialogue summarization method based on topic enhancement and dialogue center as described in claim 3, characterized in that, Step B2 specifically includes the following steps: Step B21: Use the token sequence obtained in step B13 Extracting word-level topic feature representations using an externally trained Embedded Topic Model (ETM). Determine the token sequence If the vocabulary corresponding to each token exists in the vocabulary obtained from ETM training, then its word-level topic feature representation is obtained through ETM. ; Word-level topic feature representation The calculation is as follows: in, Represents a token sequence Thematic feature representation of the i-th word; Step B22: Utilize the word-level topic feature representation obtained in step B21 The discourse-level representation obtained in step B13 Aligned discourse-level topic feature representations are obtained through stacking and dimensional expansion. ; Discourse-level topic feature representation The calculation formula is as follows: in, This represents the dimensional expansion function. Indicates inclusion The word In short, is the dimension of the feature vector.

5. The dialogue summarization method based on topic enhancement and dialogue center as described in claim 4, characterized in that, Step B3 specifically includes the following steps: Step B31: For the sentence [mask] sequence obtained in step B11 The discourse-level representation obtained in step B13 Identifying the discourse centers of a dialogue using clustering tools Distance matrix between discourse centers ; discourse center and center distance matrix The calculation is as follows: in, It indicates the central theme of the discourse in the dialogue. Let represent the i-th utterance, l be the utterance length, and d be the dimension of the feature vector; in the center distance matrix The element in the i-th row and j-th column This represents the distance from utterance i to center j. , ; Step B32: Based on the obtained discourse center and discourse level expression Calculate the degree centrality score for each utterance in the dialogue. The central discourse is selected based on the discourse centrality score. Non-centralized redundant discourse ; Degree centrality score of discourse The calculation formula is as follows: The central discourse is divided as follows: in, It is a mean function; Step B33: Calculate the central weight of each utterance based on its centrality score and the central utterance. ; The central weight of discourse The calculation is as follows: Here, α and β are adjustable weight parameters.

6. The dialogue summarization method based on topic enhancement and dialogue center as described in claim 5, characterized in that, Step B4 specifically includes the following steps: Step B41: For the discourse-level representations obtained in steps B13 and B22 Discourse-level thematic feature representation By constructing a meta-network to learn and adjust the fusion weights of discourse-level representation and discourse-level topic feature representation, a weighted discourse-level representation is obtained. and discourse-level thematic features representation ; Weighted discourse representation and discourse-level thematic features representation The calculation is as follows: MLP stands for Multilayer Perceptron. Let n be the mean function, and n be the number of utterances in the dialogue. Step B42: Fuse weighted utterance-level representations using a multi-head attention mechanism. Weighted topic feature representation This allows for a better capture of the complex relationship between discourse and thematic features, resulting in thematically enhanced feature representations. ; Feature representation of topic enhancement The calculation is as follows: in, This indicates a multi-head attention mechanism. The concatenation operation, where num represents the number of parallel heads, Q (query), K (key), and V (value) are the query, key, and value in the attention mechanism, respectively, and Q is the weighted utterance-level representation. K and V are both weighted topic feature representations. , where l is the text length of the discourse. The dimension of the feature vector; For trainable attention parameters, This represents the output of the i-th attention function. ; Step B43: Enhance the feature representation of the topic through a gating mechanism and discourse level expression Filtering and fusion are performed to obtain topic-enhanced dialogue context feature representations. ; Topic-enhanced dialogue context feature representation The calculation is as follows: in, For splicing operations, , l is the utterance length, and d is the dimension of the feature vector; These are the parameters of a trainable gating network.

7. The dialogue summarization method based on topic enhancement and dialogue center as described in claim 6, characterized in that, Step B5 specifically includes the following steps: Step B51: Calculate the discourse center weights obtained in steps B33 and B43. Theme-enhanced dialogue context feature representation Thematically enhanced discourse center representation obtained through weighted fusion ; Thematically enhanced discourse center representation The calculation is as follows: in l is the utterance length, and d is the dimension of the feature vector; Step B52: Enhance the discourse center representation of the obtained theme Discourse-level representation Re-weight and sum the results, then concatenate each sentence to obtain the final representation. ; Final expression The calculation is as follows: in, Where p is the text length of the dialogue, d is the dimension of the feature vector, and n is the number of utterances. To define hyperparameters, splicing operation.

8. The dialogue summarization method based on topic enhancement and dialogue center as described in claim 7, characterized in that, The specific implementation method of step B6 is as follows: The final representation obtained in step B52 The input is fed into the decoder, where it is decoded using a self-attention mechanism, layer normalization, topic attention, and a feedforward neural network to obtain the decoded output representation. ; in, Representation layer normalization, This indicates the self-attention mechanism. Indicates thematic attention. This represents a feedforward neural network.

9. The dialogue summarization method based on topic enhancement and dialogue center as described in claim 8, characterized in that, Step B7 specifically includes the following steps: Step B71: Represent the decoded output obtained in step B61 using the softmax function. After normalization, the score distribution of the generated next word is obtained. ; Score distribution The calculation formula is as follows: in, For a trainable parameter matrix, It is the bias vector; Step B72: The model calculates the loss value using cross-entropy as the loss function, updates the learning rate using the Adam gradient optimization algorithm, and iteratively updates the model parameters using backpropagation to train the model by minimizing the loss function.

10. A dialogue summarization system based on topic enhancement and dialogue center, 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.

Images