A participant-intention-based dialogue summary generation method

Abstract

The present application belongs to the technical field of natural language processing, and particularly relates to a dialogue summary generation method based on participant intention. In order to make the generated summary more comprehensive and complete, the method comprises the following steps: constructing a training data set of a dialogue summary model; analyzing and semantically representing the intention of a dialogue participant to construct an intention-based enhanced pseudo paraphrase data set; constructing an intention attention-based dialogue summary model; training the dialogue summary model, and finally inputting a new dialogue text into the summary model to generate a summary. The method can not only narrow the gap between the format and language style of dialogue text and structured text, but also effectively condense the intention of dialogue participants, improve the deep understanding of the model on the transformation of roles and language information in the dialogue, and generate high-quality dialogue summaries.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of natural language processing, and particularly relates to a dialogue summary generation method based on participant intention. BACKGROUND

[0002] With the development of Internet communication technology, the problem of text information overload is increasingly serious, and people rely more on online meetings, consultations and exchanges on the network, so that a large amount of dialogue text of different types is left on the Internet. The dialogue text is different from the structured narrative text, has the characteristics of multi-role interaction, dialogue super-long, topic jump and the like, so that it is difficult for dialogue participants or non-participants to capture the key information and the intention of each role in the dialogue. Therefore, it is urgent to use the summary technology to help them quickly understand the core content of the dialogue.

[0003] Most of the current works are to improve the quality of the model generating dialogue summary from the methods of dialogue features, fusion of external knowledge and data enhancement. However, the generated summary still has the problem of loss of core viewpoints and key information of dialogue participants, and the reason is that the model is not accurate in identifying the intention of dialogue participants and is not comprehensive in condensing the dialogue intention. Therefore, the existing method of injecting dialogue features by PLMS to enhance the model understanding of dialogue is difficult to generate a comprehensive and complete summary. SUMMARY

[0004] In view of the above problems, the application provides a dialogue summary generation method based on participant intention.

[0005] In order to achieve the above purpose, the application adopts the following technical scheme:

[0006] A dialogue summary generation method based on participant intention, comprising the following steps:

[0007] Step 1, constructing a training data set of dialogue summary model;

[0008] Step 2, analyzing and semantically representing the intention of dialogue participants, and constructing an intention-based enhanced pseudo paraphrase data set;

[0009] Step 3, constructing an intention attention-based dialogue summary model;

[0010] Step 4, training the dialogue summary model, and sending new dialogue text into the model to generate a summary.

[0011] Further, the step 1 of constructing a training data set of dialogue summary model has the following specific steps:

[0012] Step 1.1, downloading dialogue data set from an open source dialogue summary data set website;

[0013] Step 1.2, preprocessing the dialogue dataset, extracting dialogue data and summary data in the dialogue dataset, removing special symbols in the dialogue dataset, obtaining the training dataset {C; S} of dialogue C and summary S.

[0014] Further, the step 2 analyzes and semantically represents the intention of the dialogue participant, and constructs an intention-enhanced pseudo paraphrase dataset, the specific steps are:

[0015] Step 2.1, analyze the intention of the dialogue participant in the preprocessed dialogue dataset, and use the pre-trained model Sentence-Bert to extract the intention feature of the participant;

[0016] Step 2.2, construct an intention-enhanced segmentation block set U based on the results of step 2.1;

[0017] Step 2.3, divide the summary S into sentences, and pair each sentence with the intention-enhanced segmentation block set U and the dialogue C respectively, to form an intention-enhanced pseudo paraphrase dataset {U, C, S i}.

[0018] Further, the step 2.1 uses the pre-trained model Sentence-Bert to extract the intention feature of the participant, the specific steps are:

[0019] Step 2.1.1, use S-Bert to perform semantic feature representation on each sentence as semantic representation of participant intention, the specific implementation is as follows:

[0020] {O1,O2,…,O k}=SBert(d1,d2,…,d k ) (1)

[0021] In the formula, d i represents a paragraph of dialogue, O i represents the semantic representation of the participant's intention, and i=1,2,3...k;

[0022] Step 2.1.2, calculate the similarity between different intention feature representations, and apply it to all sentence pairs to generate a similarity matrix, the formula is as follows:

[0023]

[0024] In the formula, represents the frequency of word x in sentence O i , and sim(O i , O j ) represents the cosine similarity between a pair of sentences O i , O j ;

[0025] Step 2.1.3, the utterances with similar intentions of the participants are aggregated into intention segmentation fragments through sorting, determining segmentation boundaries and aggregation operations, i.e., the dialogue C is divided into multiple intention segmentation blocks b i : C intent = {b1, b2, …, bn}. n

[0026] Further, the step 2.2 constructs the intention-enhanced segmentation blocks based on the results of the step 2.1, and the specific steps are as follows:

[0027] Step 2.2.1, the intention segmentation blocks b i and the summary S are calculated for the Rouge-1 value, based on the results of the step 2.1, when n = 1, the intention segmentation block b1 is taken as the intention-enhanced segmentation block; when n > 1, the intention segmentation blocks b i (1 < i ≤ n) are respectively calculated for the ROUGE correlation with the summary S, and the correlation of the intention segmentation blocks b i with the summary S is evaluated through the F1 value in the Rouge-1 evaluation index:

[0028]

[0029]

[0030]

[0031] In the formula, n-gram is an n-gram, {b i} represents the intention segmentation blocks, {S} represents the summary, Count match (n-gram) represents the number of n-grams appearing in the intention segmentation blocks and the summary, and Count(n-gram) represents the number of n-grams appearing in the summary;

[0032] Step 2.2.2, m intention segmentation blocks b i with the F1 value not being 0 are selected to constitute the intention-enhanced segmentation blocks, and if the ROUGE values are all 0, the intention segmentation block b1 is taken as the intention-enhanced segmentation block, i.e., the intention-enhanced segmentation block set U = {u1, …, un} is constituted. m

[0033] Further, the step 3 constructs the dialogue summary model based on the intention attention, and the specific steps are as follows:

[0034] Step 3.1, the Encoder part of the pre-training model BART based on the Transformer is used to realize the encoding of the intention-enhanced segmentation blocks u i and the dialogue C, the intention-enhanced segmentation blocks u i = {x 0,i , x​​1,i ,…,x n,i token, x in} j,i Dialogue C = {x0, x1, ..., x} n token, x in} j The data are encoded separately using encoders, with the hidden state of the encoders serving as the encoded representation:

[0035] {h 0,i ,h 1,i ,…,h n,i}=E({x 0,i ,x 1,i ,…,x n,i}) (6)

[0036] {h1,h2,…,h n}=E({x1,x2,…,x n}) (7)

[0037] In the formula, E represents the encoder, x 0,i It is a special token added at the beginning of the segment intended for enhancement, and represented by the encoder using h. 0,i To describe the segmented block u intended for enhancement i S i =h 0,i ;

[0038] Step 3.2: Use the intent attention module to obtain the intent-enhanced segmentation blocks {S1,S2,…,S}. m The weight distribution of} is specifically obtained by using an LSTM with a self-attention mechanism to obtain the participant intent representation I with fused attention weights:

[0039] S i =LSTM(h 0,i ,S i-1 ), i∈[1,m] (8)

[0040] Q,K,V=W q S i W k S i W v S i (9)

[0041]

[0042] In the formula, LSTM() is the intent information aggregated by the intent-enhanced segmentation block through a long short-term memory network; Q, K, and V represent the query, key, and value, respectively; V is the vector representation obtained after linear transformation of the input sequence, used to calculate the attention weights; and W... q W k, W v is the parameter matrix of the neural network, d k is the dimension of K, is a scaling factor to prevent the numerical value of the inner product from affecting the learning of the neural network, and softmax() is an activation function to obtain the attention weight matrix after the activation function;

[0043] Step 3.3, introduce the cross-attention module in the transformer to focus on the participant's intention and the dialogue, and the cross-attention layer calculates the importance of the participant's intention in the dialogue α I :

[0044] v I = tanh(WI + b) (11)

[0045] v C = tanh(WC + b) (12)

[0046]

[0047] In the formula, the tanh() function performs a nonlinear transformation on the result to obtain the importance representation, I is the participant's intention representation fused with the attention weight, C is the representation of the dialogue after being encoded by the encoder, W and b are training hyperparameters, and q is a randomly initialized context vector;

[0048] The importance coefficient α I is multiplied by the dialogue representation h j formed by the multi-head attention layer, and continues to be passed forward:

[0049]

[0050] Step 3.4, use the decoder Decoder to generate the summary, input l-1 predicted tokens into the decoder to generate the lth token, the specific implementation is as follows:

[0051] {y1,…,y l-1} = D({t1,…,t l-1}, E(C)) (15)

[0052] P(t l |t ＜l , C) = Soft max(W p y l-1 ) (16)

[0053] In the formula, D represents the decoder, t i represents the token generated at the i-th step, y i represents the prediction result at the i-th step, and W pis the parameter to be learned.

[0054] Further, the step 4 trains the dialogue summary model, and the specific steps are:

[0055] The model is first trained using the intent-enhanced pseudo paraphrase data and then trained using the source dialogue summary data. After two-stage training, the hyperparameters of the summary model reach the optimal value. In the training process of the dialogue summary model, the training target of the model is to minimize the cross-entropy loss function. The specific training process adopts the teacher forcing strategy: in the training, the reference value of the previous l-1 steps is used as the input of the lth step; in the test, the input is the value of the previous l-1 steps predicted by the decoder. The cross-entropy loss function expression is as follows:

[0056] L = -∑logP(t l |t ＜l ,C) (17)

[0057] In the formula, C is the hidden layer representation of the dialogue after the encoder, t l represents the token generated in the lth step.

[0058] Compared with the prior art, the present application has the following advantages:

[0059] (1) The present application adopts an end-to-end neural network model to model the participant's intention and long dialogue. In the end-to-end framework, the attention weight and the neural network parameter are learned together.

[0060] (2) The present application constructs an intent information-based pseudo paraphrase data set. The pseudo paraphrase data not only adds a structured text prompt to the dialogue, but also increases the semantic diversity of the training data, which is conducive to the understanding and reasoning of the model for unstructured dialogue.

[0061] (3) In the intent attention-based neural network model constructed by the present application, the participant's intention and dialogue information are considered comprehensively, the dialogue representation is enriched in the form of intention perception, which further assists the dialogue summary task and improves the quality of predicted summary generation to a certain extent.

[0062] (4) The method proposed by the present application can provide a solution for participants, doctors and patients, and customer service customers to quickly understand the core content of the dialogue in the fields of meeting summary, doctor-patient dialogue, and customer service dialogue, etc., and has direct practical value and economic benefits in life and work. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 is a dialogue summary generation flowchart based on participant intention;

[0064] Figure 2 is a flowchart of step 1 of constructing a training data set;

[0065] Figure 3 is a flow chart of step 2 constructing an intent-enhanced pseudo paraphrase data set;

[0066] Figure 4 is a flow chart of step 3 constructing an intent attention-based dialogue summary model;

[0067] Figure 5 is a flow chart of step 4 summary generation unit;

[0068] Figure 6 is a dialogue summary generation model diagram constructed in step 3;

[0069] Figure 7 is a schematic diagram of the intent attention module in the present application. DETAILED DESCRIPTION

[0070] The following examples are implemented on the basis of the technical solutions of the present application, and give detailed implementation modes and specific operation processes, but do not limit the protection scope of the present application patent, and any technical solution obtained by equivalent substitution or equivalent transformation shall fall within the protection scope of the present application.

[0071] Example 1

[0072] As shown in Figure 1 , the dialogue summary generation method based on participant intent of the present embodiment includes the following steps:

[0073] S1, constructing a training data set of a dialogue summary model, the specific steps are (as shown in Figure 2 ):

[0074] S101, downloading a dialogue data set from an open source dialogue summary data set website: downloading a SAMSUM casual dialogue data set from an open source data set website, the data set contains rich topic contents such as daily chatting, arranging meetings and discussing problems;

[0075] S102, preprocessing the dialogue data set, extracting dialogue data and summary data in the dialogue data set, removing special symbols such as carriage return and line feed in the dialogue data set, obtaining a training data set {C; S} of dialogue C and summary S, and the dialogue example is shown in Table 1:

[0076] Table 1

[0077]

[0078]

[0079] S2, analyzing and semantically representing the dialogue participant intent, constructing an intent-enhanced pseudo paraphrase data set, the specific steps are (as shown inFigure 3

[0080] S201, analyze the dialogue participant utterance intent in the preprocessed dialogue data set, use the pre-trained model Sentence-Bert to extract participant intent features: from Table 1, it can be seen that each dialogue data has multiple roles participating, and the utterance of the participant contains the respective dialogue intent, which contains the key information and main opinion of the speaker; the semantic representation of the participant intent is realized by using the Sentence-BERT pre-training model, and the encoded tensor is used as the intent feature representation of the participant utterance. The similarity matrix W is obtained by calculating the similarity in C99 algorithm, and after sorting, determining the segmentation boundary and aggregation operation, the dialogue C is divided into multiple segments b, and the specific steps are as follows:

[0081] S20101, use S-Bert to perform semantic feature representation on each sentence as semantic representation of participant intent, and the specific implementation is as follows:

[0082] {O1,O2,…,O k}=SBert(d1,d2,…,d k ) (1)

[0083] In the formula, d i represents a segment of dialogue in the dialogue, O i represents the semantic representation of the participant intent, and i=1,2,3...k;

[0084] S20102, calculate the similarity between different intent feature representations, and such calculation is applied to all sentence pairs to generate a similarity matrix, and the formula is as follows:

[0085]

[0086] In the formula, represents the frequency of the word x in the sentence O i , and sim(O i , O j ) represents the cosine similarity between a pair of sentences O i , O j ;

[0087] S20103, after sorting, determining the segmentation boundary and aggregation operation, the utterances with similar participant intent are aggregated into intent segmentation segments, that is, the dialogue C is divided into multiple intent segmentation blocks (utterance blocks in the intent set) b i : C intent ={b1,b2,…,b n}.

[0088] S202, construct an intent-enhanced segmentation block set U based on the result of S201, and the specific steps are as follows:​

[0089] S20201, divide the intended block b i The Rouge-1 value is calculated with the summary S, because a higher Rouge-1 score usually means a higher score on other metrics such as Rouge-2 or Rouge-L. Based on the results of S201, when n=1, the intent segmentation block b1 is used as the intent enhancement segmentation block; when n>1, the intent segmentation block b... i (1<i≤n) ROUGE correlation degree is calculated with the abstract S respectively, and the intention segmentation block b is evaluated by the F1 value in the ROUGE-1 evaluation index. i Relevance to abstract S:

[0090]

[0091]

[0092]

[0093] In the formula, n-gram is an n-tuple, {b i} represents the intent segmentation block, {S} represents the summary, and Count match (n-gram) represents the number of n-grams that appear simultaneously in the intended segmentation block and the summary, and Count(n-gram) represents the number of n-grams that appear in the summary;

[0094] S20202. Select m intention segmentation blocks b whose F1 values ​​are not 0. i If all ROUGE values ​​of the segments constituting intent enhancement are 0, then the intent segment b1 is used as the intent enhancement segment, i.e., the set of intent enhancement segments U = {u1, ..., u...} m}

[0095] S203. Divide the summary S into sentences, and pair each sentence with the intent-enhanced segmentation block set U and the dialogue C, respectively, to form the intent-enhanced pseudo-paraphrasing dataset {U,C,S}. i The reason why it is called intention-enhanced pseudo-paraphrasing data is that the output sentence is not a complete paraphrase of the input information, but only a part of it; examples of intention-enhanced pseudo-paraphrasing data are shown in Table 2.

[0096] Table 2

[0097]

[0098]

[0099] S3. Construct a dialogue summarization model based on intent attention. The specific steps are as follows (e.g.) Figure 4as shown) :

[0100] The dialogue summary model is improved based on the pre-training model BART to realize the dialogue summary model, such as Figure 6 As shown, the model is composed of an encoder module, an intention attention module, and a decoder module.

[0101] S301, the intention enhanced segmentation block u is realized by using the Encoder part of the pre-training model BART based on Transformer i and the encoding of the dialogue C, the intention enhanced segmentation block u i ={x 0,i ,x 1,i ,…,x n,i} in token, x j,i and the token in dialogue C={x0,x1,…,x n} x j Encode through the encoder respectively, and the hidden layer state of the encoder is taken as the encoding representation respectively:

[0102] {h 0,i ,h 1,i ,…,h n,i}=E({x 0,i ,x 1,i ,…,x n,i}) (6)

[0103] {h1,h2,…,h n}=E({x1,x2,…,x n}) (7)

[0104] In the formula, E represents the encoder, x 0,i is a special token added at the beginning of the intention enhanced segmentation block, and h 0,i after the encoder is used to describe the intention enhanced segmentation block u i , that is, S i =h 0,i ;

[0105] S302, as shown Figure 7 , the intention enhanced segmentation block {S1,S2,…,S m} is obtained by using the intention attention module to obtain the weight distribution, and the participant intention representation I fused with attention weight is obtained through LSTM with self-attention mechanism:

[0106] S i =LSTM(h 0,i ,S i-1 ),i∈[1,m] (8)

[0107] Q,K,V=W q S i ,W k S i ,W v S i (9)

[0108]

[0109] In the formula, LSTM() is the intention-enhanced segmentation block aggregated intention information through the long short-term memory network, Q, K, V represent the query, key and value respectively, and is the vector representation of the input sequence after linear transformation, which is used to calculate the attention weight, W q , W k , W v is the parameter matrix of the neural network, d k is the dimension of K, is a scaling factor to prevent the value of the inner product from being too large to affect the learning of the neural network, and soft max() is an activation function to obtain the attention weight matrix after the activation function.

[0110] S303, introduce cross attention module in the transformer to focus on the participant's intention and the dialogue, and the cross attention layer calculates the importance of the participant's intention in the dialogue α I :

[0111] v I =tanh(WI+b) (11)

[0112] v C =tanh(WC+b) (12)

[0113]

[0114] In the formula, the tanh() function performs nonlinear transformation on the result to obtain the importance representation, I is the participant's intention representation fused with the attention weight, C is the representation of the dialogue after being encoded by the encoder, W and b are training hyperparameters, and q is a randomly initialized context vector.

[0115] The importance coefficient α I is multiplied by A formed by the multi-head attention layer acting on the dialogue representation h j , and continues to be passed forward:

[0116]

[0117] The decoder Decoder is used to generate the summary, and the l-1 predicted tokens are input into the decoder to generate the lth token, and the specific implementation is as follows:

[0118] {y1,…,y l-1}=D({t1,…,t l-1},E(C)) (15)

[0119] P(t l |t ＜l ,C)=Soft max(W p y l-1 ) (16)

[0120] In the formula, D represents a decoder, t i represents a token generated in the i-th step, y i represents a prediction result in the i-th step, and W p is a parameter to be learned.

[0121] S4, training a dialogue summary model, and sending new dialogue text into the model to generate a summary, specific steps are as shown in the following table (as shown in the following table) : Figure 5

[0122] S401, constructing a cross-entropy loss function, and performing two-stage training on the entire system:

[0123] The model is trained using the intent-enhanced pseudo paraphrase data to optimize the hyperparameters of the summary model; in the training process of the dialogue summary model, minimizing the cross-entropy loss function is used as the training target of the model, and the teacher forcing strategy is used in the specific training process: in the training, the reference value of the previous l-1 steps is used as the input of the i-th step; in the test, the input is the value of the previous l-1 steps predicted by the decoder, and the cross-entropy loss function expression is as follows:

[0124] L=-∑log P(t l |t ＜l ,C) (17)

[0125] In the formula, C is the hidden layer representation of the dialogue after the encoder, and t l represents a token generated in the i-th step.

[0126] S402, inputting the dialogue summary data of the test set into the trained summary model to generate a predicted summary, and Table 3 is an output example of a dialogue:

[0127] Table 3

[0128]

[0129] ​

Claims

1. A method for generating dialogue summaries based on participant intent, characterized in that, Includes the following steps: Step 1: Construct the training dataset for the dialogue summarization model; Step 2: Analyze and semantically represent the intentions of dialogue participants, and construct a pseudo-parsing dataset based on intention enhancement; Step 3: Construct a dialogue summarization model based on intent attention; Step 4: Train the dialogue summarization model by feeding the new dialogue text into the model to generate a summary. Step 2 involves analyzing and semantically representing the intentions of dialogue participants, and constructing a pseudo-parsing dataset based on intention enhancement. The specific steps are as follows: Step 2.1: Analyze the utterance intent of the dialogue participants in the preprocessed dialogue dataset and use the pre-trained model Sentence-Bert to extract the participants' intent features. Step 2.2: Construct an intent-enhanced segmentation block set U based on the results of Step 2.1; Step 2.3: Divide the summary S into sentences, and pair each sentence with the intent-enhanced segmentation block set U and the dialogue C respectively, thus forming the intent-enhanced pseudo-paraphrasing dataset. ; Step 2.1 uses the pre-trained model Sentence-Bert to extract participant intent features. The specific steps are as follows: Step 2.1.1: Use Sentence-BERT to perform semantic feature representation on each sentence as a semantic representation of the participant's intent. The specific implementation is as follows: （1） In the formula, This refers to a passage from a dialogue. A semantic representation of the participant's intention, i=1,2,3...k; Step 2.1.2: Calculate the similarity between different intent feature representations and apply it to all sentence pairs to generate a similarity matrix, as shown in the following formula: （2） In the formula, Indicator In the sentence Frequency of occurrence in Represents a pair of sentences , Cosine similarity between them; Step 2.1.3 involves sorting, determining segmentation boundaries, and aggregation operations to group utterances with similar participant intentions into intention segmentation fragments, i.e., dividing dialogue C into multiple intention segmentation blocks. : .

2. The dialogue summary generation method based on participant intent according to claim 1, characterized in that, The specific steps for constructing the training dataset for the dialogue summarization model in step 1 are as follows: Step 1.1: Download the dialogue dataset from the open-source dialogue summary dataset website; Step 1.2: Preprocess the dialogue dataset, extract the dialogue data and summary data from the dialogue dataset, remove special symbols from the dialogue dataset, and obtain the training dataset for dialogue C and summary S. .

3. The dialogue summary generation method based on participant intent according to claim 1, characterized in that, In step 2.2, the intention-enhanced segmentation block set U is constructed based on the results of step 2.

1. The specific steps are as follows: Step 2.2.1, segment the intended blocks Calculate the Rouge-1 value with the summary S. Based on the results of step 2.1, when n=1, segment the intent into blocks. As a segmentation block for intent enhancement; when n > 1, the intent segmentation block will be... ROUGE correlation was calculated between each block and the abstract S, and the intention segmentation blocks were evaluated using the F1 score in the ROUGE-1 evaluation metric. Relevance to abstract S: （3） （4） （5） In the formula, n-gram is an n-tuple. This indicates the intent to segment the block, and {S} represents the summary. This indicates the number of n-grams that appear simultaneously in both the intent segmentation block and the summary. This indicates the number of n-grams appearing in the summary; Step 2.2.2: Select m intention segmentation blocks with F1 values ​​that are not 0. The intention-enhanced segmentation blocks are constructed if all ROUGE values ​​are 0. As a segment for intent enhancement, that is, the set of segments that constitute intent enhancement. .

4. The dialogue summary generation method based on participant intent according to claim 1, characterized in that, Step 3 involves constructing a dialogue summarization model based on intent attention. The specific steps are as follows: Step 3.1: Implement intent-enhanced segmentation blocks using the Encoder part of the Transformer-based pre-trained model BART. Encoding of Dialogue C, Intended Enhancement Segmentation Blocks token in and dialogue token in The data are encoded separately using encoders, with the hidden state of the encoders used as the encoding representation: （6） （7） In the formula, Indicates encoder, It involves adding a special token at the beginning of the segment intended for enhancement, and using the encoded representation. To describe the segmentation blocks intended for enhancement ,Right now = ; Step 3.2: Obtain the intent-enhanced segmentation blocks using the intent attention module. The weight distribution is specifically obtained by using an LSTM with a self-attention mechanism to obtain the participant intent representation I with fused attention weights: （8） （9） （10） In the formula, The intent-enhanced segmentation blocks are aggregated with intent information through a Long Short-Term Memory network. Q, K, and V represent the query, key, and value, respectively. Q is a vector representation of the input sequence after a linear transformation, used to calculate the attention weights. , , This is the parameter matrix of the neural network. For the dimension of K, 1 / This is a scaling factor to prevent excessively large inner product values ​​from negatively impacting the learning of the neural network. The activation function is used to obtain the attention weight matrix. Step 3.3 introduces a cross-attention module into the transformer to focus on participant intent and dialogue. The cross-attention layer calculates the importance of participant intent in the dialogue in the following way. : （11） （12） （13） In the formula, the tanh() function performs a nonlinear transformation on the result to obtain an indication of importance. It is a participant intent representation that incorporates attention weights. It is the representation of the dialogue after being encoded by the encoder, W and b are training hyperparameters, and q is a randomly initialized context vector; Importance coefficient With dialogue The A multiplication, formed by the multi-head attention layer, continues to propagate forward: （14） Step 3.4: Use the decoder to generate a digest, based on the input... The predicted tokens are fed into the decoder to generate the first... Each token is implemented as follows: （15） （16） In the formula, Indicates decoder, This represents the token generated in step i. This represents the prediction result at step i. These are the parameters to be learned.

5. The dialogue summary generation method based on participant intent according to claim 1, characterized in that, The specific steps for training the dialogue summarization model in step 4 are as follows: The model is trained using intention-enhanced pseudo-paraphrased data to optimize the hyperparameters of the summarizing model. During the training of the dialogue summarizing model, minimizing the cross-entropy loss function is used as the training objective. Specifically, a teacher forcing strategy is employed: during training, the teacher forcing data is used... The reference value of the first step is used as the first step. The input for the step; at test time, the input is the previous step predicted by the decoder. The value of the step, the expression for the cross-entropy loss function is as follows: （17） In the formula, C is the hidden layer representation of the dialogue after passing through the encoder. Indicates the first The token generated step by step.

Images