A method, apparatus and device for generating dialogue summaries

Abstract

This invention discloses a dialogue summarization method, apparatus, and device. The method includes, upon receiving a summarization request, extracting multiple key sentences from the dialogue document corresponding to the request using weighted methods; encoding each key sentence to generate multiple feature vectors; upon receiving a decoding start message, sequentially decoding each feature vector to generate an initial dialogue summary corresponding to the dialogue document; and updating the initial dialogue summary using non-key sentences from the dialogue document to generate a target dialogue summary. This effectively reduces the length of the data input and the computational burden on the model while considering the different contributions of different fields to key information, thus improving the flexibility of dialogue summarization generation.

Description

Technical Field

[0001] This invention relates to the field of document processing technology, and in particular to a method, apparatus, and device for generating dialogue summaries. Background Technology

[0002] Currently, various forms of dialogue scenarios are increasingly common, such as meetings, television series, interviews, and casual conversations. Dialogue summarization can extract key information from complex dialogue data, thereby reducing the difficulty for people to understand the data.

[0003] The main approaches to dialogue summarization, besides directly applying document summarization models to the conversation setting, include conversation-tailored models aimed at achieving state-of-the-art performance, such as hierarchical modeling of conversations. The rich structured information within dialogues has also been explored and utilized, including external information such as dialogue behavior, key point / entity sequences, topic fragments, phase development, discourse relationships, and common-sense knowledge, to help understand the global context of dialogue.

[0004] However, existing dialogue summarization methods mostly use datasets from past unstructured text summarization models for training and fine-tuning, which are not specifically designed for dialogue summarization tasks. They also face significant processing pressure for long dialogue sequences and cannot flexibly consider the differences in the contribution of different fields to key information, resulting in low flexibility in dialogue summarization generation. Summary of the Invention

[0005] This invention provides a dialogue summarization method, apparatus, and device, which solves the technical problems of existing dialogue summarization methods, which mostly use datasets from past unstructured text summarization models for training and fine-tuning. These methods are not specific to dialogue summarization tasks, and they face significant processing pressure for long dialogue sequences. Furthermore, they cannot flexibly consider the differences in the contribution of different fields to key information, resulting in low flexibility in dialogue summarization generation.

[0006] The first aspect of this invention provides a dialogue summary generation method, comprising:

[0007] When a summary request is received, multiple key sentences are extracted from the dialogue document corresponding to the summary request using a weighted method.

[0008] Each of the key sentences is encoded to generate multiple feature vectors;

[0009] When the decoding start information is received, each of the aforementioned feature vectors is decoded sequentially to generate an initial dialogue summary corresponding to the dialogue document;

[0010] The initial dialogue summary is updated using non-key sentences from the dialogue document to generate the target dialogue summary.

[0011] Optionally, the step of extracting multiple key sentences from the dialogue document specified in the summary request in a weighted manner when a summary request is received includes:

[0012] When a summary request is received, the dialogue document specified in the summary request is segmented into multiple dialogue statements and initial weights are set.

[0013] A word segmentation tool is used to perform word segmentation on each of the dialogue statements to determine multiple initial words corresponding to each of the dialogue statements.

[0014] The initial weights are adjusted based on a preset regular expression and multiple initial words;

[0015] Calculate the sentence score for each of the aforementioned dialogue statements;

[0016] Based on the statement scores, select multiple key sentences from the multiple dialogue statements, with the number of requests corresponding to the summary request being equal to the number of key sentences.

[0017] Optionally, the regular expression includes specific word expressions and numeric character expressions; the step of adjusting the initial weights according to the preset regular expression and multiple initial words includes:

[0018] Retrieve initial words that match the feature word expression as first words, and determine the dialogue statement associated with the first words as the first query statement;

[0019] Retrieve initial words that match the numeric character expression as second words, and determine the dialogue statement associated with the second words as the second query statement;

[0020] Increase the initial weight corresponding to the first query statement by a first preset increment;

[0021] The initial weight of the second query statement is increased by a second preset range; wherein the first preset range is greater than the second preset range.

[0022] Optionally, the step of calculating the sentence score corresponding to each of the dialogue sentences includes:

[0023] Each initial word in each of the dialogue statements is selected as a word to be processed, and the dialogue statement in which it is located is determined as a query statement.

[0024] Calculate the importance measure corresponding to each of the words to be processed;

[0025] Calculate the proportion of each word to be processed in each of the dialogue statements;

[0026] The score for each query statement is calculated by combining the importance metric and the occurrence ratio with a preset score calculation formula.

[0027] The formula for calculating the fraction is as follows:

[0028]

[0029] Wherein, BM25F(Doc,Q) is the statement score of dialogue statement Q in the dialogue document Doc, and dialogue statement Q is composed of the initial vocabulary q1 to q25F(Doc,Q). r Composition; IDF(q) i ) represents the i-th word to be processed. i Importance metric, TF(q) i ,f j ) represents the i-th word to be processed. i In the j-th dialog statement f, excluding query statement Q j The proportion of occurrences in the query statement Q, r is the total number of initial words in the query statement Q, u is the number of dialogue statements in the dialogue document Doc excluding the query statement Q, and w j As the initial weights for each dialogue statement, UL j avgUL represents the length of the j-th dialog statement excluding query statement Q, and avgUL represents the total length of all dialog statements f. j The average length, k is a preset positive parameter, b is an adjustable parameter, 0 <b<1。

[0030] Optionally, the target encoder includes an embedding layer, a multi-head self-attention mechanism layer, and a feedforward neural network layer; the step of encoding each of the key sentences to generate multiple feature vectors includes:

[0031] The first key sentence is input into the embedding layer as the statement to be encoded, according to the order of the dialogue document;

[0032] The embedding layer converts multiple words in the sentence to be encoded into high-dimensional vectors and transmits them to the multi-head self-attention mechanism layer.

[0033] After establishing the dependency relationship between each of the high-dimensional vectors through the multi-head self-attention mechanism layer, the vectors are transmitted to the feedforward neural network layer.

[0034] The feedforward neural network layer performs a nonlinear transformation on each of the high-dimensional vectors according to the dependency relationship to generate feature vectors.

[0035] Unprocessed key sentences are selected as new sentences to be encoded according to the order of the dialogue document;

[0036] Jump to execute the step of converting multiple words in the sentence to be encoded into high-dimensional vectors through the embedding layer and transmitting them to the multi-head self-attention mechanism layer, until multiple feature vectors equal to the number of key sentences are generated.

[0037] Optionally, the step of sequentially decoding each of the feature vectors upon receiving the decoding start information to generate an initial dialogue summary corresponding to the dialogue document includes:

[0038] When the decoding start information is received, the hidden state is initialized using the preset target decoder, and the first feature vector is selected as the decoding vector;

[0039] The initial state is constructed using the hidden state and the decoding vector;

[0040] The target decoder generates a first summary word based on the initial state and the decoding vector, and caches it in the word sequence.

[0041] Select the unprocessed feature vector as the new decoding vector;

[0042] The target decoder uses the word sequence at the current time and the decoding vector to decode, generate second summary words, and cache them in the word sequence;

[0043] The process jumps to the step of selecting unprocessed feature vectors as new decoding vectors until all feature vectors have been processed, and the word sequence at the current time is determined as the initial dialogue summary corresponding to the dialogue document.

[0044] Optionally, the step of updating the initial dialogue summary using non-key sentences from the dialogue document to generate the target dialogue summary includes:

[0045] Filter all statements in the dialogue document except for the key sentences as non-key sentences;

[0046] The initial dialogue summary is input into a preset first GRU model to generate summary semantic information;

[0047] The non-key sentences are input into a preset second GRU model to generate contextual semantic information;

[0048] The context semantic information and the summary semantic information are concatenated to obtain a concatenated vector;

[0049] The concatenated vector is decoded to generate a target dialogue summary.

[0050] Optionally, the method further includes:

[0051] Retrieve multiple dialogue summary data pairs;

[0052] The target encoder and target decoder are obtained by training a preset initial encoder and a preset initial decoder using multiple dialogue summary data pairs.

[0053] A second aspect of the present invention provides a dialogue summary generation apparatus, comprising:

[0054] The key sentence extraction module is used to extract multiple key sentences from the dialogue document corresponding to the summary request in a weighted manner when a summary request is received.

[0055] The vector encoding module is used to encode each of the key sentences and generate multiple feature vectors;

[0056] The vector decoding module is used to decode each of the feature vectors sequentially when the decoding start information is received, so as to generate an initial dialogue summary corresponding to the dialogue document.

[0057] The summary update module is used to update the initial dialogue summary using non-key sentences from the dialogue document to generate the target dialogue summary.

[0058] A third aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the dialogue summary generation method as described in any of the first aspects of the present invention.

[0059] As can be seen from the above technical solutions, the present invention has the following advantages:

[0060] Upon receiving a summary request, multiple key sentences are extracted from the corresponding dialogue document using a weighted approach. Each key sentence is encoded to generate multiple feature vectors. When decoding start information is received, each feature vector is decoded sequentially to generate an initial dialogue summary corresponding to the dialogue document. The initial dialogue summary is then updated using non-key sentences from the dialogue document to generate the target dialogue summary. This effectively reduces the length of the data input and the computational burden on the model, while also considering the different contributions of different fields to key information, thus improving the flexibility of dialogue summary generation. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1This is a flowchart of the steps of a dialogue summary generation method provided in Embodiment 1 of the present invention;

[0063] Figure 2 This is a flowchart of the steps of a dialogue summary generation method provided in Embodiment 2 of the present invention;

[0064] Figure 3 This is a schematic diagram of the key sentence extraction process using BM25F as the extraction algorithm in an embodiment of the present invention;

[0065] Figure 4 A flowchart illustrating a specific example of a dialogue provided in an embodiment of the present invention;

[0066] Figure 5 This is a flowchart illustrating a dialogue summary generation process according to an embodiment of the present invention;

[0067] Figure 6 This is a structural block diagram of a dialogue summary generation device provided in Embodiment 3 of the present invention. Detailed Implementation

[0068] The main approaches to dialogue summarization, besides directly applying document summarization models to the conversation setting, include conversation-tailored models aimed at achieving state-of-the-art performance, such as hierarchical modeling of conversations. The rich structured information within dialogues has also been explored and utilized, such as dialogue behaviors, key point / entity sequences, topic fragments, phase development, and discourse relationships. External information, such as common-sense knowledge, is also incorporated to help understand the global context of dialogue.

[0069] Most existing dialogue summarization methods based on pre-trained neural networks use Transformer-based models as the pre-trained neural network. However, since dialogue summarization is typically used in scenarios such as meetings and customer service conversations, the dialogues generated in these scenarios are often quite long. This results in the data used to train and fine-tune a dialogue summarization model being very long, frequently exceeding the input length limit of Transformer-based models, leading to poor summary quality.

[0070] If we consider using algorithms to extract key sentences from a dialogue as input for a downstream model, we often think of using TF-IDF or BM25 algorithms. However, TF-IDF does not consider stop words during calculation, and stop words may constitute a large portion of the vocabulary in a document, thus affecting the results. Furthermore, the BM25 algorithm treats the entire document as a single entity when calculating key sentence scores, meaning it cannot consider the weight information between different fields within the document. In other words, certain key words and data may not be prioritized, resulting in poor quality of the obtained key information and consequently, poor quality of the trained dialogue summarization model.

[0071] To address this, embodiments of the present invention provide a dialogue summarization method, apparatus, and device to solve the technical problem that existing dialogue summarization methods mostly use datasets from past unstructured text summarization models for training and fine-tuning, which are not specific to dialogue summarization tasks, have significant processing pressure for long-sequence dialogues, and cannot flexibly consider the differences in the contribution of different fields to key information, resulting in low flexibility in dialogue summarization generation.

[0072] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0073] Please see Figure 1 , Figure 1 The flowchart illustrates the steps of a dialogue summary generation method provided in Embodiment 1 of the present invention.

[0074] This invention provides a dialogue summary generation method, comprising:

[0075] Step 101: When a summary request is received, extract multiple key sentences from the dialogue document corresponding to the summary request using a weighted method.

[0076] A conversation document refers to document-type data used to record the content of conversations between users. Each sentence is marked with the speaker on the front end and is introduced with a colon.

[0077] In this embodiment of the invention, when a summary request is received from an external user or other terminal, the dialogue document for which dialogue summary generation is required is extracted from the summary request, or the dialogue document stored locally or in the cloud is retrieved according to the summary request.

[0078] After obtaining the dialogue document, a weighted score is calculated for each dialogue statement based on the dialogue sentences and vocabulary in the document, and the key sentences are extracted based on their weighted scores.

[0079] It should be noted that key sentences refer to dialogue statements that are richer in semantic information and more important, including but not limited to dialogue statements containing certain specific keywords such as transition words, conjunctions, or dialogue topic words, or dialogue statements containing numbers or other specific characters such as letter abbreviations or acronyms.

[0080] Step 102: Encode each key sentence to generate multiple feature vectors;

[0081] After extracting multiple key sentences, they can be encoded according to the words in the key sentences and the relationships between the words to generate feature vectors corresponding to each key sentence.

[0082] Among them, the feature vector refers to the context vector corresponding to the key sentence, which contains important semantic information of the key sentence.

[0083] Step 103: When the decoding start information is received, each feature vector is decoded sequentially to generate the initial dialogue summary corresponding to the dialogue document;

[0084] When multiple feature vectors are obtained, if decoding start information is received, for example... <sos>(start of sequence) At this point, the corresponding decoder can be used to decode each feature vector in sequence, thereby generating a word corresponding to a feature vector at each time step and caching it. After all feature vectors are decoded, the initial dialogue summary corresponding to the dialogue document is generated.

[0085] Step 104: Update the initial dialogue summary using non-key sentences from the dialogue document to generate the target dialogue summary.

[0086] Dialogue summarization refers to transforming one or more dialogues into a more concise form, making it easier for people to read, understand, and review the content. The generated summary text includes the main information and key points of the dialogue, while minimizing its length. Dialogue summarization has wide applications; for example, in online customer service, service representatives handle a large number of user interactions, and dialogue summarization technology can help them understand customer needs more quickly and improve work efficiency.

[0087] In this embodiment, after generating the initial dialogue summary, the semantics of the initial dialogue summary may shift due to the adjustment of the weights of some key sentences. In this case, the remaining sentences in the dialogue document (excluding key sentences, i.e., non-key sentences) can be secondary. Their corresponding contextual semantic information is obtained and fused with the semantic information of the initial dialogue summary before decoding, thereby further updating the initial dialogue summary, generating the target dialogue summary, and ultimately producing a more accurate and coherent dialogue summary.

[0088] In this embodiment of the invention, when a summary request is received, multiple key sentences are extracted from the dialogue document corresponding to the summary request using a weighted method; each key sentence is encoded to generate multiple feature vectors; when a decoding start message is received, each feature vector is decoded sequentially to generate an initial dialogue summary corresponding to the dialogue document; the initial dialogue summary is updated using non-key sentences from the dialogue document to generate a target dialogue summary. This effectively reduces the length of the data input and the computational burden on the model, while considering the differences in the contribution of different fields to key information, thus improving the flexibility of dialogue summary generation.

[0089] Please see Figure 2 , Figure 2 The flowchart illustrates the steps of a dialogue summary generation method provided in Embodiment 2 of the present invention.

[0090] This invention provides a dialogue summary generation method, comprising:

[0091] Step 201: When a summary request is received, the dialogue document specified in the summary request is segmented into multiple dialogue statements and initial weights are set.

[0092] In this embodiment of the invention, when a summary request is received from an external source, it indicates that the user needs to generate a summary for a specific dialogue document. At this point, the currently uploaded dialogue documents can be displayed in a second-level page. A "Create New Document" button can also be provided, allowing users to upload new dialogue documents by clicking it.

[0093] Upon receiving a summary request, the system responds by specifying a dialogue document, segments the document into multiple dialogue statements according to predetermined rules, and assigns an initial weight to each statement. These initial weights may be equal or unequal.

[0094] Specifically, a sentence segmentation tool is used to segment the sentences in each dialogue. In Python, the `sent_tokenize` function from the `nltk` library can be used for sentence segmentation, resulting in the first preprocessed dataset `DSum_Sentences`, which consists of multiple dialogue sentences. Simultaneously, the dataset structure is modified by adding a "weight" attribute to each separated sentence, initialized to 1.

[0095] It should be noted that for ordinary documents, attention is focused more on sections such as the title and the first and last paragraphs. However, for a dialogue scenario, considering that different fields should have different weights, the quality of key sentence extraction is better, resulting in a better summary. The BM25F algorithm can consider different fields having different weights, thus making key sentence extraction more accurate. In this way, important information in the text can be more accurately distinguished during key sentence extraction, thereby improving the accuracy and quality of the summary. The dialogue document in this embodiment of the invention contains clearly distinguishing symbols to differentiate between different speakers.

[0096] Step 202: Use a word segmentation tool to perform word segmentation on each dialogue statement to determine multiple initial words corresponding to each dialogue statement;

[0097] In this embodiment of the invention, in order to retain sentences rich in more effective or even key information and to distinguish them from other sentences, these sentences should also be identified. For this purpose, a word segmentation tool can be used to perform word segmentation operations on each dialogue statement separately, thereby determining multiple initial words corresponding to each dialogue statement.

[0098] In Python, the word_tokenize function from the nltk library can be used to implement word segmentation, which divides each initial word that conforms to the word rules into multiple initial words corresponding to each dialogue statement.

[0099] Step 203: Adjust the initial weights based on the preset regular expression and multiple initial words;

[0100] Optionally, the regular expression includes specific word expressions and numeric character expressions, and step 203 may include the following sub-steps:

[0101] The initial words that match the feature word expression are taken as the first words, and the dialogue statements associated with the first words are determined as the first query statements.

[0102] The initial words that match the numeric character expression are used as the second words, and the dialogue statements associated with the second words are determined as the second query statements.

[0103] Increase the initial weight of the first query statement by the first preset increment;

[0104] The initial weight of the second query statement is increased according to the second preset range; wherein the first preset range is greater than the second preset range.

[0105] In this embodiment of the invention, sentences beginning with specific words or containing data typically contain more important information and are more worthy of user attention. Regular expressions can be used to match the initial words, filtering out the first words that match the feature word expression, and determining the associated dialogue statements as the first query statement. Simultaneously, second words that match numeric character expressions can be retrieved, and the associated dialogue statements of the second words are determined as the second query statement. The initial weight corresponding to the first query statement is increased by a first preset amount, and the initial weight corresponding to the second query statement is increased by a second preset amount.

[0106] In practical implementation, feature word expressions can be associated with transition words such as "but" and "however," as well as words like "so," "finally," and "result." Numeric character expressions can use regular expressions to easily match numbers and surrounding characters.

[0107] Wherein, the first preset amplitude is greater than the second preset amplitude. The first preset amplitude can be set to 50% or 31% to 70%, and the second preset amplitude can be set to 30%. The specific values ​​of these values ​​are not limited in the embodiments of the present invention.

[0108] Step 204: Calculate the sentence score for each dialogue statement.

[0109] Optionally, step 204 may include the following sub-steps:

[0110] Each initial word in each dialogue statement is selected as a word to be processed, and the dialogue statement in which it is located is determined as the query statement.

[0111] Calculate the importance measure for each word to be processed;

[0112] Calculate the proportion of each word to be processed in each dialogue statement;

[0113] The system uses a pre-defined score calculation formula combined with importance metrics and occurrence ratios to calculate the score for each query statement.

[0114] The formula for calculating fractions is as follows:

[0115]

[0116] Wherein, BM25F(Doc,Q) is the statement score of dialogue statement Q in the dialogue document Doc, and dialogue statement Q is composed of the initial vocabulary q1 to q25F(Doc,Q). r Composition; IDF(q) i ) represents the i-th word to be processed. i Importance metric, TF(q) i ,f j ) represents the i-th word to be processed. i In the j-th dialog statement f, excluding query statement Q j The proportion of occurrences in the query statement Q, r is the total number of initial words in the query statement Q, u is the number of dialogue statements in the dialogue document Doc excluding the query statement Q, and w j As the initial weights for each dialogue statement, UL j avgUL represents the length of the j-th dialog statement excluding query statement Q, and avgUL represents the total length of all dialog statements f. j The average length, k is a preset positive parameter, b is an adjustable parameter, 0 <b<1。

[0117] In this embodiment, the importance metric corresponding to each initial word is calculated one by one. Specifically, the IDF (Individual Word Function Value) is used as a measure of the general importance of a word or phrase. In this extractor, a specific word q... i The IDF can be obtained by dividing the total number of occurrences by the number of statements containing the word, and then taking the logarithm to the base 10 of the quotient.

[0118]

[0119] Among them, U num This refers to the total number of dialogue statements in the dialogue document, df(q i ) refers to the word q to be processed. i The number of statements is used as the denominator. If this value is used directly as the denominator, it will result in a denominator of 0 when the value is 0. Therefore, it is usually added to the denominator.

[0120] At the same time, the occurrence ratio of each initial word in each dialogue sentence is calculated one by one, specifically through q i In f j The number of occurrences in q i In f j The quotient of the number of occurrences of all keywords in TF(q i , f j ) is represented as:

[0121]

[0122] t represents the term q to be processed i In f j The number of occurrences, and T represents in f j All terms q to be processed in i The total number of occurrences.

[0123] It should be noted that the variable k is a positive parameter used to standardize the range of term frequencies in word granularity retrieval applications. When k = 0, it is a binary model, and a larger value corresponds to using more raw term frequency information. This parameter controls the rate of increase of the term frequency result in term frequency saturation. The empirical value is 1.2. The smaller the value, the faster the saturation change, and the larger the value, the slower the saturation change.

[0124] b is another adjustable parameter (0 < b < 1) that determines the range of using document length to represent information content: when b is 1, it fully uses document length to weigh the weights of statements, and when b is 0, it means not using document length. The larger b is, the greater the impact of document length on the relevance score, and vice versa. In the case of the same frequency, the relevance between a long document and q i Should be weaker than the relevance between a short document and q i . The empirical value is 0.75.

[0125] As can be seen from the formula, BM25F divides the document into multiple regions and performs weighted statistics on different domains. For example, in the document, different weights are assigned, and in different domains, each index (term frequency, text length, etc.) of each word is weighted and summed to obtain the final result. Figure 3 Figure is a schematic diagram of the key sentence extraction process using BM25F as the extraction algorithm.

[0126] It shows the relationship between the query statement Q (Query), the dialogue document Doc (Document), and the domain (Field). The depth of the color of the domain represents the weight size of the domain. During calculation, each other sentence is regarded as a domain, that is, f in the BM25F formula; the weight corresponding to the sentence is the weight w corresponding to the domain in the BM25F formula; the sentence length is the domain length UL in the BM25F formula. The average length of all sentences is the domain average length avgUL in the BM25F formula.

[0127] It's worth noting that the BM25F algorithm (Best Matching 25 with Fields), an improvement on BM25, was originally used in information retrieval and search engine algorithms. It adds factors such as field weights and field lengths to more accurately calculate document relevance. In BM25F, each document is divided into multiple fields, each with a different weight factor. Besides considering the correspondence between keywords in the query and documents, BM25F further considers the relevance between fields.

[0128] In this embodiment of the invention, each initial word in each dialogue statement is selected as a word to be processed, and the dialogue statement in which it is located is determined as a query statement. The importance metric corresponding to each word to be processed is calculated. The occurrence ratio of each word to be processed in each dialogue statement is calculated. A preset score calculation formula is used to combine the importance metric and the occurrence ratio to calculate the statement score corresponding to each query statement. The calculation is repeated until each dialogue statement is used as a query statement and the corresponding statement score is calculated.

[0129] Step 205: Select multiple key sentences from multiple dialogue statements, in descending order of sentence score, that correspond to the number of requests for the summary request.

[0130] After calculating the statement score for each dialogue statement, the importance of each dialogue statement is obtained and sorted from high to low according to their statement scores. Based on the number of requests corresponding to the summary request, multiple key sentences are selected from multiple dialogue statements, which is equal to the number of requests.

[0131] In addition, the number of requests can be specified not only through the digest request, but also by setting a preset value in this method if the digest request does not carry the corresponding number of requests.

[0132] In this embodiment of the invention, after steps 201-205, a dataset consisting of multiple key sentences has been obtained, and the BM25F score of each key sentence has been calculated. At this point, feature vectors can be generated and decoded using a pre-trained neural network encoder-decoder approach to obtain the corresponding dialogue summary.

[0133] Pretrained neural networks (PNNs) refer to models trained on large-scale datasets, whose learned feature parameters are saved as initialization parameters and then fine-tuned on other tasks. In this embodiment of the invention, the pretrained language model BART (Bidirectional and Auto-Regressive Transformer) is used as the generative model. This model is based on an autoencoder structure and uses a Transformer encoder-decoder architecture to learn the ability to generate text from a large-scale text corpus through unsupervised training. This pretraining method does not require labeled data and can be applied to various natural language generation tasks.

[0134] Step 206: Encode each key sentence to generate multiple feature vectors;

[0135] Furthermore, the target encoder includes an embedding layer, a multi-head self-attention mechanism layer, and a feedforward neural network layer; step 206 may include the following sub-steps:

[0136] The first key sentence is input into the embedding layer as the statement to be encoded, following the order of the dialogue document.

[0137] The embedding layer converts multiple words in the sentence to be encoded into high-dimensional vectors and transmits them to the multi-head self-attention mechanism layer.

[0138] After establishing the dependencies between the high-dimensional vectors through a multi-head self-attention mechanism layer, the vectors are transmitted to the feedforward neural network layer.

[0139] Feature vectors are generated by performing nonlinear transformations on each high-dimensional vector according to the dependency relationship through a feedforward neural network layer.

[0140] Select unprocessed key sentences as new sentences to be encoded according to the order of the dialogue document;

[0141] The process jumps to the step of converting multiple words in the sentence to be encoded into high-dimensional vectors through the embedding layer and transmitting them to the multi-head self-attention mechanism layer until multiple feature vectors equal to the number of key sentences are generated.

[0142] In the implementation, each key sentence is used as input to the model and encoded by a BART encoder to obtain a feature vector. This feature vector can be seen as a semantic representation of the dialogue, containing rich information related to the key sentences. Specifically, the input is fed into an embedding layer, which converts each word or sub-word into a high-dimensional vector representation to capture the semantic information and relationships between words. After the embedding layer, the input vector is passed to a series of Transformer encoders. Each encoder consists of a multi-head self-attention mechanism and a feedforward neural network. The self-attention mechanism allows the model to establish dependencies between words within the input sequence, thus better understanding semantic associations. The feedforward neural network helps the model perform non-linear transformations. After all the Transformer encoder layers are completed, a set of encoder outputs is obtained. These outputs can be seen as context vectors encoding the key sentences. The context vectors capture the important semantic information of the key sentences and can be used in the subsequent decoder generation process.

[0143] Step 207: When the decoding start information is received, each feature vector is decoded sequentially to generate the initial dialogue summary corresponding to the dialogue document;

[0144] Optionally, step 207 may include the following sub-steps:

[0145] When the decoding start information is received, the hidden state is initialized using the preset target decoder, and the first feature vector is selected as the decoding vector;

[0146] The initial state is constructed using the hidden state and the decoding vector;

[0147] The target decoder generates the first summary words based on the initial state and the decoding vector, and caches them in the word sequence;

[0148] Select the unprocessed feature vector as the new decoding vector;

[0149] The target decoder uses the word sequence and decoding vector at the current time to decode, generates second summary words, and caches them into the word sequence;

[0150] Jump to the next step of selecting unprocessed feature vectors as new decoding vectors, until all feature vectors have been processed, and determine the word sequence at the current time as the initial dialogue summary corresponding to the dialogue document.

[0151] In the specific implementation, the decoder receives a specific start marker, such as <sos>(Start of sequence) and then automatically generates dialogue summaries step by step. Specifically, the decoder initializes a hidden state as a memory for the model during the decoding process. This hidden state, combined with the first feature vector, constitutes the initial state required to generate the summary. Then, based on the initial state and the decoded vector of the selected first feature vector, the decoder generates the first summary words and caches them in the word sequence, progressively generating the dialogue summary. At each time step, the decoder selects an unprocessed feature vector as a new decoded vector, and generates the next word, i.e., the second summary word, based on the previously generated word sequence and the current decoded vector. This continues until a specific end marker is encountered, such as... <eos>(end of sequence) or reach the maximum length limit.

[0152] Optionally, before performing step 208, this method further includes the following steps:

[0153] When the number of words in the first word is less than or equal to the preset transition threshold, and an initial dialogue summary is generated, the initial dialogue summary is determined as the target dialogue summary.

[0154] In practical implementation, due to the differences in dialogue documents and the varying complexity of the dialogues within them, the number of the first vocabulary words, i.e., the number of transition words or relational words, can be used to determine whether further optimization of the initial dialogue summary is needed. Therefore, before executing step 208, the number of the first vocabulary words can be determined first. If its number is less than or equal to a preset transition threshold, it indicates that the semantic deviation between the initial dialogue summary and the original dialogue document is relatively small. If an initial dialogue summary has already been generated, it can be determined as the target dialogue summary.

[0155] Step 208: Update the initial dialogue summary using non-key sentences from the dialogue document to generate the target dialogue summary.

[0156] Furthermore, step 208 may include the following sub-steps:

[0157] Filter all statements in the dialogue document except for key sentences as non-key sentences;

[0158] Input the initial dialogue summary into the preset first GRU model to generate summary semantic information;

[0159] Input non-key sentences into a pre-defined second GRU model to generate contextual semantic information;

[0160] By concatenating contextual semantic information and summary semantic information, a concatenated vector is obtained.

[0161] The concatenated vector is decoded to generate a summary of the target dialogue.

[0162] Because of weight adjustments made to certain sentences, such as those containing transition words, the final key sentences may contain a large number of sentences with transition words. This could lead to significant semantic deviations in the final generated initial dialogue summary. For example... Figure 4 The main point of the dialogue in the example data is "Mary suggests rescheduling the picnic," but the initial summary is "...Mary tells Tom to go on the picnic without feeling burdened," which creates a serious semantic deviation. Admittedly, sentences containing transition words are crucial, but even "non-critical sentences" can lead to incorrect summaries if not considered in context.

[0163] Therefore, in this embodiment, the initial dialogue summary is updated by fusing contextual semantic information. All sentences except key sentences can be selected from the dialogue document as non-key sentences. The initial dialogue summary is then fed into a preset first GRU model, which performs semantic parsing to generate summary semantic information. Simultaneously, non-key sentences can be input into a preset second GRU model to generate contextual semantic information. By concatenating the contextual semantic information and the summary semantic information, a concatenated vector is obtained. Finally, a BART decoder is used to decode the concatenated vector to generate the target dialogue summary.

[0164] In the specific implementation, S summary For the initial dialogue summary, S context Given a context sequence consisting of non-key sentences, we first use GRU to obtain representations of these two sequences.

[0165] For the semantic information h of the summary summary The representation of:

[0166] h summary =GRU(S) summary )

[0167] For contextual semantic information h context The representation of:

[0168] h context =GRU(S) contett )

[0169] Then, combine these two pieces of information:

[0170]

[0171] Finally, the concatenated vector is input into the BART decoder to obtain the final summary:

[0172] Summary final =BART decoder (h combined )

[0173] GRU (Gated Recurrent Unit) is a recurrent neural network architecture particularly adept at capturing long-range dependencies in time-series or sequential data. Unlike traditional RNNs, GRU utilizes its unique gating mechanism to better store and access earlier information, making it excellent at processing long sequences. When processing dialogues or text summarization, contextual information often spans multiple sentences. GRU's gating mechanism effectively captures and preserves these long-range dependencies. GRU's gating structure allows it to decide which information should be passed to the next time step, allowing the model to focus on the information most relevant to generating the summary.

[0174] This approach combines GRU's ability to capture sequential information with BART's superior text generation performance, resulting in high-quality, coherent summaries. By combining the contextual representations of the generated summaries with those of key sentences, the relationship between the summaries and the original dialogue can be captured more effectively. Further decoding by the BART decoder generates even more accurate and coherent summaries. During processing, using only the context of key sentences instead of the entire original dialogue text as input effectively alleviates the input length issue and reduces the computational burden on the model.

[0175] In another example of the invention, the method further includes:

[0176] Retrieve multiple dialogue summary data pairs;

[0177] Multiple dialogue summary data pairs are used to train a pre-defined initial encoder and a pre-defined initial decoder to obtain a target encoder and a target decoder.

[0178] In this embodiment of the invention, the encoder and decoder of the pre-trained neural network can be constructed and the model trained first, wherein the pre-trained neural network takes the BART model as an example:

[0179] Obtain a dialogue summary dataset DSum, which consists of several dialogue summary data pairs (DS). Each dialogue D is composed of several sentences: D = {U1, U2, ..., U...} n }

[0180] The goal of the key sentence extraction stage is to transform each dialogue D into data D composed of key sentences. K D K ={K1,K2,…,K m Each K comes from D, and m ≤ n, i.e., D K It is a subset of D. After calculating the BM25F score corresponding to each key sentence, the sentences with the highest BM25F scores are selected from high to low as key sentences, and a new dataset is constructed with the corresponding dialogue summary, which serves as the training and fine-tuning data for the next step of the neural network model to generate the summary.

[0181] When training the BART model, a two-stage fine-tuning approach can be employed. The first stage involves supervised learning using key sentences to fine-tune the model and improve the quality of the generated dialogue summaries. The second stage involves end-to-end fine-tuning the model using the entire dialogue and its summaries to further enhance performance. During fine-tuning, the cross-entropy loss function can be used to measure the difference between the generated dialogue summaries and the standard dialogue summaries. Simultaneously, techniques such as gradient accumulation and learning rate decay can be employed to accelerate model training and improve performance. Finally, this trained BART model can be used to automatically generate dialogue summaries.

[0182] Please see Figure 5 , Figure 5 This is a flowchart illustrating a dialogue summary generation process in an embodiment of the present invention.

[0183] In this embodiment of the invention, the specific steps include Stages 1-3:

[0184] Stage 1: Use the BM25F algorithm to calculate the score of each sentence and extract the key sentences from them;

[0185] Stage 2: The key sentences are processed by the encoder and decoder in the neural network to generate an initial summary;

[0186] Stage 3: The initial summary is updated by integrating semantic information from the context of key sentences to obtain the final summary.

[0187] In this embodiment of the invention, when a summary request is received, multiple key sentences are extracted from the dialogue document corresponding to the summary request using a weighted method; each key sentence is encoded to generate multiple feature vectors; when a decoding start message is received, each feature vector is decoded sequentially to generate an initial dialogue summary corresponding to the dialogue document; the initial dialogue summary is updated using non-key sentences from the dialogue document to generate a target dialogue summary. This effectively reduces the length of the data input and the computational burden on the model, while considering the differences in the contribution of different fields to key information, thus improving the flexibility of dialogue summary generation.

[0188] Please see Figure 6 , Figure 6 A structural block diagram of a dialogue summary generation device according to Embodiment 3 of the present invention is shown.

[0189] This invention provides a dialogue summary generation apparatus, comprising:

[0190] The key sentence extraction module 601 is used to extract multiple key sentences from the dialogue document corresponding to the summary request in a weighted manner when a summary request is received.

[0191] The vector encoding module 602 is used to encode each key sentence and generate multiple feature vectors;

[0192] The vector decoding module 603 is used to decode each feature vector sequentially when the decoding start information is received, and generate the initial dialogue summary corresponding to the dialogue document.

[0193] The summary update module 604 is used to update the initial dialogue summary using non-key sentences from the dialogue document to generate the target dialogue summary.

[0194] Optionally, the key sentence extraction module 601 includes:

[0195] The statement segmentation submodule is used to segment the dialogue document specified in the summary request into multiple dialogue statements and set initial weights when a summary request is received.

[0196] The word segmentation submodule is used to perform word segmentation on each dialogue statement using a word segmentation tool to determine multiple initial words corresponding to each dialogue statement.

[0197] The weight adjustment submodule is used to adjust the initial weights based on a preset regular expression and multiple initial words;

[0198] The statement score calculation submodule is used to calculate the statement score for each dialogue statement.

[0199] The key sentence selection submodule is used to select multiple key sentences from multiple dialogue statements, in descending order of sentence score, that correspond to the number of requests for the summary request.

[0200] Optionally, regular expressions include specific word expressions and numeric character expressions; the weight adjustment submodule is specifically used for:

[0201] The initial words that match the feature word expression are taken as the first words, and the dialogue statements associated with the first words are determined as the first query statements.

[0202] The initial words that match the numeric character expression are used as the second words, and the dialogue statements associated with the second words are determined as the second query statements.

[0203] Increase the initial weight of the first query statement by the first preset increment;

[0204] The initial weight of the second query statement is increased according to the second preset range; wherein the first preset range is greater than the second preset range.

[0205] Optionally, the statement fraction calculation submodule is specifically used for:

[0206] Each initial word in each dialogue statement is selected as a word to be processed, and the dialogue statement in which it is located is determined as the query statement.

[0207] Calculate the importance measure for each word to be processed;

[0208] Calculate the proportion of each word to be processed in each dialogue statement;

[0209] The system uses a pre-defined score calculation formula combined with importance metrics and occurrence ratios to calculate the score for each query statement.

[0210] The formula for calculating fractions is as follows:

[0211]

[0212] Wherein, BM25F(Doc,Q) is the statement score of dialogue statement Q in the dialogue document Doc, and dialogue statement Q is composed of the initial vocabulary q1 to q25F(Doc,Q). r Composition; IDF(q) i ) represents the i-th word to be processed. i Importance metric, TF(q) i ,f j ) represents the i-th word to be processed. i In the j-th dialog statement f, excluding query statement Q j The proportion of occurrences in the query statement Q, r is the total number of initial words in the query statement Q, u is the number of dialogue statements in the dialogue document Doc excluding the query statement Q, and w j As the initial weights for each dialogue statement, UL j avgUL represents the length of the j-th dialog statement excluding query statement Q, and avgUL represents the total length of all dialog statements f. j The average length, k is a preset positive parameter, b is an adjustable parameter, 0 <b<1。

[0213] Optionally, the target encoder includes an embedding layer, a multi-head self-attention mechanism layer, and a feedforward neural network layer; the vector encoding module 602 is specifically used for:

[0214] The first key sentence is input into the embedding layer as the statement to be encoded, following the order of the dialogue document.

[0215] The embedding layer converts multiple words in the sentence to be encoded into high-dimensional vectors and transmits them to the multi-head self-attention mechanism layer.

[0216] After establishing the dependencies between the high-dimensional vectors through a multi-head self-attention mechanism layer, the vectors are transmitted to the feedforward neural network layer.

[0217] Feature vectors are generated by performing nonlinear transformations on each high-dimensional vector according to the dependency relationship through a feedforward neural network layer.

[0218] Select unprocessed key sentences as new sentences to be encoded according to the order of the dialogue document;

[0219] The process jumps to the step of converting multiple words in the sentence to be encoded into high-dimensional vectors through the embedding layer and transmitting them to the multi-head self-attention mechanism layer until multiple feature vectors equal to the number of key sentences are generated.

[0220] Optionally, the vector decoding module 603 is specifically used for:

[0221] When the decoding start information is received, the hidden state is initialized using the preset target decoder, and the first feature vector is selected as the decoding vector;

[0222] The initial state is constructed using the hidden state and the decoding vector;

[0223] The target decoder generates the first summary words based on the initial state and the decoding vector, and caches them in the word sequence;

[0224] Select the unprocessed feature vector as the new decoding vector;

[0225] The target decoder uses the word sequence and decoding vector at the current time to decode, generates second summary words, and caches them into the word sequence;

[0226] Jump to the next step of selecting unprocessed feature vectors as new decoding vectors, until all feature vectors have been processed, and determine the word sequence at the current time as the initial dialogue summary corresponding to the dialogue document.

[0227] Optionally, the summary update module 604 is specifically used for:

[0228] The process involves: filtering all sentences except key sentences from the dialogue document as non-key sentences; inputting the initial dialogue summary into a preset first GRU model to generate summary semantic information; inputting the non-key sentences into a preset second GRU model to generate context semantic information; concatenating the context semantic information and the summary semantic information to obtain a concatenation vector; and decoding the concatenation vector to generate the target dialogue summary.

[0229] Optionally, the device further includes:

[0230] The data pair acquisition module is used to acquire multiple dialogue summary data pairs;

[0231] The model training module is used to train a preset initial encoder and a preset initial decoder using multiple dialogue summary data to obtain a target encoder and a target decoder.

[0232] This invention provides an electronic device, including a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the dialogue summary generation method as described in any embodiment of this invention.

[0233] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described device and module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0234] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0235] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0236] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0237] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.< / eos> < / sos> < / sos>

Claims

1. A dialogue summary generation method, characterized in that, include: When a summary request is received, multiple key sentences are extracted from the dialogue document corresponding to the summary request using a weighted method. Each of the key sentences is encoded to generate multiple feature vectors; When the decoding start information is received, each of the aforementioned feature vectors is decoded sequentially to generate an initial dialogue summary corresponding to the dialogue document; The initial dialogue summary is updated using non-key sentences from the dialogue document to generate the target dialogue summary; The step of updating the initial dialogue summary using non-key sentences from the dialogue document to generate the target dialogue summary includes: Filter all statements in the dialogue document except for the key sentences as non-key sentences; The initial dialogue summary is input into a preset first GRU model to generate summary semantic information; The non-key sentences are input into a preset second GRU model to generate contextual semantic information; The context semantic information and the summary semantic information are concatenated to obtain a concatenated vector; The concatenated vector is decoded to generate a target dialogue summary.

2. The method according to claim 1, characterized in that, The step of extracting multiple key sentences from the dialogue document specified in the summary request in a weighted manner when a summary request is received includes: When a summary request is received, the dialogue document specified in the summary request is segmented into multiple dialogue statements and initial weights are set. A word segmentation tool is used to perform word segmentation on each of the dialogue statements to determine multiple initial words corresponding to each of the dialogue statements. The initial weights are adjusted based on a preset regular expression and multiple initial words; Calculate the sentence score for each of the aforementioned dialogue statements; Based on the statement scores, select multiple key sentences from the multiple dialogue statements, with the number of requests corresponding to the summary request being equal to the number of key sentences.

3. The method according to claim 2, characterized in that, The regular expression includes specific word expressions and numeric character expressions; the step of adjusting the initial weights based on the preset regular expression and multiple initial words includes: Retrieve initial words that match the specific word expression as first words, and determine the dialogue statement associated with the first words as the first query statement; Retrieve initial words that match the numeric character expression as second words, and determine the dialogue statement associated with the second words as the second query statement; Increase the initial weight of the first query statement by a first preset margin. The initial weight of the second query statement is increased by a second preset range; wherein the first preset range is greater than the second preset range.

4. The method according to claim 3, characterized in that, The step of calculating the sentence score corresponding to each of the dialogue sentences includes: Each initial word in each of the dialogue statements is selected as a word to be processed, and the dialogue statement in which it is located is determined as the query statement. Calculate the importance measure corresponding to each of the words to be processed; Calculate the proportion of each word to be processed in each of the dialogue statements; The score for each query statement is calculated by combining the importance metric and the occurrence ratio with a preset score calculation formula. The formula for calculating the fraction is as follows: ； in, For dialogue statements In the dialogue document Statement scores in dialogue From initial vocabulary composition; For the i-th word to be processed Importance metrics For the i-th word to be processed In addition to the query statement The j-th dialogue statement outside The proportion of occurrences in the query statement, where r is the percentage of occurrences in the query statement. The total number of initial words, u is the dialogue document. Except for query statements The number of out-of-conversation statements The initial weights for each dialogue statement are: Except for the query statement The length of the j-th dialogue statement outside of the main dialogue. For all dialogue statements The average length, k is a preset positive parameter, b is an adjustable parameter, 0 <b<1。 5. The method according to claim 1, characterized in that, The method involves a preset target encoder, which includes an embedding layer, a multi-head self-attention mechanism layer, and a feedforward neural network layer; the step of encoding each of the key sentences to generate multiple feature vectors includes: The first key sentence is input into the embedding layer as the statement to be encoded, according to the order of the dialogue document; The embedding layer converts multiple words in the sentence to be encoded into high-dimensional vectors and transmits them to the multi-head self-attention mechanism layer. After establishing the dependency relationship between each of the high-dimensional vectors through the multi-head self-attention mechanism layer, the vectors are transmitted to the feedforward neural network layer. The feedforward neural network layer performs a nonlinear transformation on each of the high-dimensional vectors according to the dependency relationship to generate feature vectors. Unprocessed key sentences are selected as new sentences to be encoded according to the order of the dialogue document; Jump to execute the step of converting multiple words in the sentence to be encoded into high-dimensional vectors through the embedding layer and transmitting them to the multi-head self-attention mechanism layer, until multiple feature vectors equal to the number of key sentences are generated.

6. The method according to claim 1, characterized in that, The step of decoding each of the feature vectors sequentially to generate an initial dialogue summary corresponding to the dialogue document upon receiving the decoding start information includes: When the decoding start information is received, the hidden state is initialized using the preset target decoder, and the first feature vector is selected as the decoding vector; The initial state is constructed using the hidden state and the decoding vector; The target decoder generates a first summary word based on the initial state and the decoding vector, and caches it in the word sequence. Select the unprocessed feature vector as the new decoding vector; The target decoder uses the word sequence at the current time and the decoding vector to decode, generate second summary words, and cache them in the word sequence; The process jumps to the step of selecting unprocessed feature vectors as new decoding vectors until all feature vectors have been processed, and the word sequence at the current time is determined as the initial dialogue summary corresponding to the dialogue document.

7. The method according to any one of claims 1-6, characterized in that, The method further includes: Retrieve multiple dialogue summary data pairs; The target encoder and target decoder are obtained by training a preset initial encoder and a preset initial decoder using multiple dialogue summary data pairs.

8. A dialogue summarization generation apparatus, characterized in that, include: The key sentence extraction module is used to extract multiple key sentences from the dialogue document corresponding to the summary request in a weighted manner when a summary request is received. The vector encoding module is used to encode each of the key sentences and generate multiple feature vectors; The vector decoding module is used to decode each of the feature vectors sequentially when the decoding start information is received, so as to generate an initial dialogue summary corresponding to the dialogue document. The summary update module is used to update the initial dialogue summary using non-key sentences from the dialogue document to generate a target dialogue summary; The summary update module is specifically used for: Filter all statements in the dialogue document except for the key sentences as non-key sentences; The initial dialogue summary is input into a preset first GRU model to generate summary semantic information; The non-key sentences are input into a preset second GRU model to generate contextual semantic information; The context semantic information and the summary semantic information are concatenated to obtain a concatenation vector; The concatenated vector is decoded to generate a target dialogue summary.

9. An electronic device, characterized in that, The system includes a memory and a processor, wherein the memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of the dialogue summary generation method as described in any one of claims 1-7.

Images