Extraction-Generation Hybrid Language Multi-Document News Summarization Methods, Systems, Devices, Media, and Products

By constructing sentence-sentence isomorphism graphs and sentence-word heteromorphism graphs, and integrating inter-sentence relationships and intra-sentence word-level semantic representations, key sentences are extracted to generate summaries. This solves the problems of information loss and coherence in cross-language multi-document news summaries, and achieves efficient key information coverage and interpretable summary generation.

CN123068626APending Publication Date: 2026-11-10YUXI POWER SUPPLY BUREAU OF YUNNAN POWER GRID
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611368407.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-04
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies for generating cross-language multi-document news summaries suffer from problems such as difficulties in cross-language semantic alignment, translation cascade errors, information loss, and insufficient summary coherence.

Method used

We employ an extractive-generative hybrid language multi-document news summarization method. By segmenting sentences, words, and cross-language unified semantic encoding, we construct sentence-sentence isomorphism graphs and sentence-word heteromorphism graphs, integrate inter-sentence relation representations and intra-sentence word-level semantic representations, extract key sentences, and generate target language summaries.

Benefits of technology

It improves the coverage of key information, the coherence of the summary, and the interpretability of the generation process, while reducing translation cascade errors and information loss caused by long inputs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123068626A_ABST
    Figure CN123068626A_ABST
Patent Text Reader

Abstract

This application discloses an extraction-generation hybrid language multi-document news summarization method, system, device, medium, and product, relating to the field of natural language processing. The method includes: acquiring a hybrid language multi-document cluster consisting of multiple news documents in at least two languages, obtaining a sentence set and a vocabulary set; constructing sentence-sentence isomorphism graphs and sentence-word heteromorphism graphs; obtaining inter-sentence relation representations based on an isomorphism graph attention network and intra-sentence word-level semantic representations based on a heteromorphism graph attention network; fusing the inter-sentence relation representations and intra-sentence word-level semantic representations, extracting multiple key sentences based on importance scores, and forming a key segment set; encoding and decoding the key segment set based on a multilingual sequence-to-sequence generator to obtain a target language news summary. This application can structurally model cross-document and cross-language associations, reduce input length and translation cascade errors, and improve key information coverage, summary coherence, and the interpretability of the generation process.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of natural language processing, and in particular to an extraction-generation hybrid language multi-document news summarization method, system, device, medium, and product. Background Technology

[0002] Hybrid-language multi-document news summarization aims to generate concise, complete, and semantically coherent summaries in the target language from multiple news documents written in different languages ​​around the same event or topic. This task faces challenges such as cross-language semantic alignment, cross-document redundancy suppression, long input modeling, and key information selection.

[0003] Related methods often employ a "translate first, then summarize" paradigm, which involves translating documents from different languages ​​into the target language before calling a single-language multi-document summarization model. This paradigm facilitates the reuse of existing models, but semantic shifts, amplified ambiguities, and information loss during the translation stage can propagate to the summarization stage, creating cascading errors. Furthermore, the input length after concatenating multiple documents can easily exceed the context window of the generating model, resulting in insufficient coverage of key information.

[0004] Another type of extraction-generation method first filters candidate sentences and then generates a summary. However, relying solely on single sentence similarity or a single graph structure makes it difficult to simultaneously reflect the discourse-level relationships between sentences and the word-level semantic composition within sentences. Furthermore, the global importance judgment of the extractor and the dynamic content selection of the generator at each decoding time step may cause target offset, affecting the factual coverage and coherence of the summary. Therefore, a technical solution is needed that can simultaneously model inter-sentence relationships and intra-sentence word-level structure, and coordinate the extraction and generation stages. Summary of the Invention

[0005] The purpose of this application is to provide an extraction-generative hybrid language multi-document news summarization method, system, device, medium, and product that can structurally model cross-document and cross-language relationships, reduce input length and translation cascade errors, and improve the coverage of key information, the coherence of the summary, and the interpretability of the generation process.

[0006] To achieve the above objectives, this application provides the following solution: Firstly, this application provides an extractive-generative hybrid language multi-document news summarization method, including: Obtain a mixed-language multi-document cluster consisting of multiple news documents in at least two languages, and perform sentence segmentation, word segmentation, and cross-language unified semantic encoding on each news document to obtain a sentence set and a vocabulary set.

[0007] Sentence-sentence isomorphism graphs are constructed based on sentence sets, and sentence-word heteromorphism graphs are constructed based on sentence sets and vocabulary sets.

[0008] Inputting sentence-sentence isomorphism graphs into isomorphism graph attention networks yields inter-sentence relation representations, while inputting sentence-word heteromorphism graphs into heteromorphism graph attention networks yields intra-sentence word-level semantic representations.

[0009] By integrating inter-sentence relation representations and intra-sentence word-level semantic representations, a comprehensive representation and importance score for each candidate sentence are obtained.

[0010] Multiple key sentences were extracted based on importance scores and combined into a set of key segments.

[0011] The key fragment set is encoded and decoded based on a multilingual sequence-to-sequence generator to obtain a news summary in the target language.

[0012] Secondly, this application provides an extractive-generative hybrid language multi-document news summarization system, including: sentence set and vocabulary set units, multi-view construction unit, representation unit, fusion unit, key sentence extraction unit, and summary generation unit.

[0013] The sentence set and vocabulary set units are used to obtain a mixed-language multi-document cluster consisting of multiple news documents containing at least two languages. Each news document is segmented into sentences, words, and subjected to cross-language unified semantic encoding to obtain the sentence set and vocabulary set.

[0014] The multi-view building unit is used to construct sentence-sentence isomorphism graphs based on sentence sets, and to construct sentence-word heteromorphism graphs based on sentence sets and vocabulary sets.

[0015] The representation unit is used to input sentence-sentence isomorphism graphs into the isomorphism graph attention network to obtain inter-sentence relation representations, and to input sentence-word heteromorphism graphs into the heteromorphism graph attention network to obtain intra-sentence word-level semantic representations.

[0016] The fusion unit is used to fuse inter-sentence relation representations and intra-sentence word-level semantic representations to obtain a comprehensive representation and importance score for each candidate sentence.

[0017] The key sentence extraction unit is used to extract multiple key sentences based on importance scores and form a set of key segments.

[0018] The summary generation unit is used to encode and decode a set of key segments based on a multilingual sequence-to-sequence generator to obtain a news summary in the target language.

[0019] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described extract-generative hybrid language multi-document news summarization method.

[0020] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described extraction-generative hybrid language multi-document news summarization method.

[0021] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described extraction-generative hybrid language multi-document news summarization method.

[0022] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides an extraction-generative hybrid language multi-document news summarization method, system, device, medium, and product. By segmenting each news document into sentences, words, and performing cross-language unified semantic encoding, it obtains sentence sets and vocabulary sets. It does not rely on the traditional "translate first, then summarize" paradigm, but directly performs cross-language unified semantic encoding on multi-language documents, eliminating translation cascading errors. Based on the sentence sets, it constructs sentence-sentence isomorphism graphs, and based on the sentence sets and vocabulary sets, it constructs sentence-word heteromorphism graphs, achieving complementary dual-view modeling and cross-document information flow. It integrates inter-sentence relationship representations and intra-sentence word-level semantic representations to obtain candidate... The system comprehensively represents and scores the importance of sentences, extracts multiple key sentences based on their importance scores, and forms a set of key segments. The inter-sentence relationship representation provides information on the contextual position of sentences in the overall discourse, while the intra-sentence word-level semantic representation provides information on the fine-grained semantic integrity of sentences. The fusion of these two approaches achieves information complementarity. Extraction based on importance scores ensures that the set of key segments covers the core information of each document. The set of key segments is then encoded and decoded using a multilingual sequence-to-sequence generator to obtain a news summary in the target language, effectively improving the coverage of key information, the coherence of the summary, and the interpretability of the generation process. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating an extraction-generative hybrid language multi-document news summarization method provided in one embodiment of this application.

[0025] Figure 2 This is an overall flowchart of an extraction-generative hybrid language multi-document news summarization method provided in one embodiment of this application.

[0026] Figure 3The diagram illustrates the structural relationships between a sentence-sentence isomorphic view, a sentence-word heteromorphic view, a key sentence extractor, and a summary generator, as provided in an embodiment of this application.

[0027] Figure 4 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0028] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] like Figure 1 As shown, an extraction-generative hybrid language multi-document news summarization method is provided, including the following steps 201 to 206.

[0031] Step 201: Obtain a mixed-language multi-document cluster consisting of multiple news documents containing at least two languages, and perform sentence segmentation, word segmentation, and cross-language unified semantic encoding on each news document to obtain a sentence set and a vocabulary set.

[0032] Step 202: Construct a sentence-sentence isomorphism graph based on the sentence set, and construct a sentence-word heteromorphism graph based on the sentence set and the vocabulary set.

[0033] Step 203: Input the sentence-sentence isomorphism graph into the isomorphism graph attention network to obtain the inter-sentence relationship representation, and input the sentence-word heteromorphism graph into the heteromorphism graph attention network to obtain the intra-sentence word-level semantic representation.

[0034] Step 204: Integrate the inter-sentence relation representation and the intra-sentence word-level semantic representation to obtain the comprehensive representation and importance score of each candidate sentence.

[0035] Step 205: Extract multiple key sentences based on importance scores and form a set of key segments.

[0036] Step 206: Encode and decode the key fragment set based on the multilingual sequence-to-sequence generator to obtain the target language news summary.

[0037] By implementing steps 201 to 206 above, each news document is segmented into sentences, words, and subjected to cross-language unified semantic encoding to obtain a sentence set and a vocabulary set. This approach directly encodes multilingual documents across languages, eliminating translation cascading errors, without relying on the traditional "translate first, then summarize" paradigm. A sentence-sentence isomorphism graph is constructed based on the sentence set, and a sentence-word heteromorphism graph is constructed based on the sentence set and vocabulary set, achieving complementary dual-view modeling and cross-document information flow. The inter-sentence relationship representation and intra-sentence word-level semantic representation are fused to obtain a comprehensive representation and importance score for each candidate sentence. Multiple key sentences are extracted based on the importance score and formed into a key fragment set. The inter-sentence relationship representation provides contextual information about the sentence's position in the overall text, while the intra-sentence word-level semantic representation provides fine-grained semantic integrity information about the sentence itself. The fusion of these two aspects achieves information complementarity. Extraction based on importance scores ensures that the key fragment set covers the core information of each document. The key fragment set is then encoded and decoded using a multilingual sequence-to-sequence generator to obtain a news summary in the target language, effectively improving the coverage of key information, the coherence of the summary, and the interpretability of the generation process.

[0038] In another exemplary embodiment of this application, each news document is segmented into sentences, words, and subjected to cross-language unified semantic encoding to obtain a set of sentences and a set of words, including the following steps.

[0039] Step 301: Segment the news document by calling the corresponding sentence delimitation rules based on the language identifier, resulting in multiple sentences.

[0040] Step 302: Use a word segmenter to convert each sentence into a word sequence to obtain multiple words.

[0041] Step 303: Input the word sequence of each sentence into the same multilingual encoder to obtain the word hiding state.

[0042] Step 304: Pool the hidden sub-word states in the sentence, map them to the same-dimensional cross-lingual semantic space through a shared projection layer, and perform L2 normalization to obtain the initial sentence representation of each sentence in the sentence set.

[0043] Step 305: Aggregate the hidden states of sub-words belonging to the same vocabulary, map them to the same-dimensional cross-linguistic semantic space through a shared vocabulary projection layer, and perform L2 normalization to obtain the initial vocabulary representation of each word in the vocabulary set.

[0044] In another exemplary embodiment of this application, a sentence-sentence isomorphism graph is constructed based on a sentence set, and a sentence-word heteromorphism graph is constructed based on the sentence set and the vocabulary set, including the following steps.

[0045] Step 401: Using each sentence in the sentence set as a node, and sentence pairs with common vocabulary or sentence pairs with semantic similarity reaching a preset threshold in the unified semantic space as undirected edges, a sentence-sentence isomorphism graph is obtained.

[0046] Step 402: Take each sentence in the sentence set as a sentence node and each word in the vocabulary set as a vocabulary node; when the sentence corresponding to the sentence node contains the word corresponding to the vocabulary node, establish a sentence-word connection edge between the sentence node and the vocabulary node to obtain a sentence-word heterogeneous graph.

[0047] In another exemplary embodiment of this application, the isomorphic graph attention network uses a multi-head attention mechanism to weighted aggregate the mapping features of adjacent sentence nodes, and obtains the inter-sentence relationship representation through a feedforward neural network and residual connections.

[0048] In another exemplary embodiment of this application, the heterogeneous graph attention network obtains intra-sentence word-level semantic representations through bidirectional message propagation from sentence nodes to word nodes and from word nodes to sentence nodes.

[0049] In another exemplary embodiment of this application, the multilingual sequence-to-sequence generator is mBART; the key segment set is encoded and decoded based on the multilingual sequence-to-sequence generator to obtain a target language news summary, including the following steps.

[0050] Step 501: Perform mBART encoding on each key sentence in the key segment set to obtain the hidden state of each key sentence.

[0051] Step 502: At the current decoding time step, input the hidden state of each key sentence into the language model head to obtain the condition word generation probability of each key sentence, and input it into the multilayer perceptron to obtain the dynamic sentence weight of each key sentence at the current decoding time step.

[0052] Step 503: The generation probabilities of condition words in each key sentence are weighted and summed according to the dynamic sentence weights to obtain the marginalization generation probability based on the key segment set. Then, the marginalization generation probability is used to decode word by word until the end marker is generated to obtain the target language news summary.

[0053] Specifically, the extraction-generative hybrid language multi-document news summarization method of this application includes the following steps.

[0054] Step 1: Obtain a mixed-language multi-document cluster consisting of multiple news documents in at least two languages. Perform sentence segmentation, word segmentation, and cross-language unified semantic encoding on each news document to obtain a sentence set and a vocabulary set. First, sentence segmentation is performed based on the language identifier of each news document by calling the corresponding sentence boundary rules. The word segmenter of the multi-language pre-trained model is used to convert each sentence into a word sequence. The hidden states of sentence words are pooled and mapped to the same-dimensional cross-language semantic space through a shared projection layer to obtain the initial representation of the sentence. The hidden states of sub-words belonging to the same vocabulary are aggregated to obtain the initial representation of the vocabulary.

[0055] Step 2: Construct a sentence-sentence isomorphism graph based on the sentence set, and construct a sentence-word heteromorphism graph based on the sentence set and the vocabulary set.

[0056] Specifically, in the sentence-sentence isomorphism graph G1=(V1,E1,W1), the node set V1 consists of all sentences within the document cluster, the edge set E1 consists of sentence pairs with common vocabulary or semantic similarity reaching a preset threshold, and the edge weight W1 represents the correlation strength of sentence pairs in a unified cross-linguistic semantic space.

[0057] In the sentence-word heteromorphism graph G2=(V2,E2,W2), V2=V1∪V w V w E1 is the set of word nodes for the document cluster, E2 is the set of connections between sentence nodes and word nodes, and W2 is the TF-IDF weight calculated in the shared semantic space generated by the multilingual pre-trained model.

[0058] Among them, sentence-sentence isomorphism graphs are used to represent semantic or structural associations between sentences across documents, while sentence-word heteromorphism graphs are used to represent word-level semantic associations between sentences and words.

[0059] The initial representation of a word node is obtained by encoding with mBERT; the initial representation of a sentence node is obtained by inputting the word sequence of the sentence into mBERT to obtain the semantic representation, extracting local patterns using a convolutional neural network, capturing contextual dependencies using a bidirectional long short-term memory network, and concatenating the outputs of the two.

[0060] Step 3: Input the sentence-sentence isomorphism graph into the isomorphism graph attention network to obtain the inter-sentence relationship representation, and input the sentence-word heteromorphism graph into the heteromorphism graph attention network. Obtain the intra-sentence word-level semantic representation through bidirectional message propagation from sentence node to word node and from word node to sentence node.

[0061] Specifically, the isomorphic graph attention network employs a multi-head attention mechanism to weighted aggregate the mapping features of adjacent sentence nodes, and fuses the original sentence semantics with the aggregated inter-sentence context through a feedforward neural network and residual connections. In one parameter setting, the isomorphic graph attention network includes 1 to 4 network layers, each layer having 2 to 16 attention heads. The outputs of each attention head are concatenated or averaged and projected to a unified dimension. Residual connections, layer normalization, and random deactivation are applied to each layer.

[0062] The heterogeneous graph attention network sets type-specific linear transformation matrices for sentence nodes and word nodes. In each iteration, sentence node information is first passed to word nodes, and then the updated word node information is back-aggregated to sentence nodes. In one parameter setting, the heterogeneous graph attention network includes 1 to 4 bidirectional message propagation layers, and sets 2 to 8 attention heads for both sentence node-to-word node and word node-to-sentence node relationships, and adds the weights of sentence-word connection edges to the attention coefficient calculation of the corresponding relationship.

[0063] Step 4: Integrate the inter-sentence relationship representation and intra-sentence word-level semantic representation to obtain the comprehensive representation and importance score of each candidate sentence. Select the top K candidate sentences based on the importance score and form a key segment set according to the order of the selected sentences in the original news document.

[0064] Specifically, the sentence representations obtained from the isomorphic graph and the heteromorphic graph are concatenated as vectors. The concatenated composite sentence representation is then subjected to a binary classification importance score, where K is a threshold for the number of key sentences determined based on the validation set summary evaluation metric. The composite sentence representation is then processed by a multilayer perceptron to obtain logarithmic sentence importance values. The predicted probability is obtained by using the Sigmoid function. The reasoning stage is in accordance with Select the top K candidate sentences from highest to lowest.

[0065] Step 5: Input the set of key segments into the multilingual sequence generator. For each decoding time step, determine the condition word generation probability and dynamic sentence weight of each key sentence. Perform weighted marginalization on the condition word generation probability of each key sentence to obtain the target language news summary.

[0066] Specifically, the multilingual sequence to sequence generator is mBART; at the t-th decoding time step, the hidden states obtained for each key sentence are input into the language model head and the multilayer perceptron respectively to obtain the condition word generation probability and the sentence weight that dynamically changes with the generated summary prefix. Marginalization generation is achieved by summing the weighted conditional probabilities of all key sentences.

[0067] In Steps 3-5, attention propagation is performed on the two views using isomorphic GAT and heterogeneous GAT respectively, sentence representations from the two views are fused and Top-K key sentences are selected; the key sentences are arranged into a key fragment set in the original text order, input into the mBART generator, and marginalized generation is performed based on the dynamic sentence weights at each decoding time step to obtain the target language summary.

[0068] Step 6: In the model training phase, a joint loss function is constructed based on the extractor loss, generator loss, and extract-generation consistency loss. This function is used to jointly optimize the isomorphic graph attention network, heteromorphic graph attention network, and multilingual sequence-to-sequence generator. During model training, multilingual multi-document news documents and their corresponding target language news summaries from the training samples are input into the extract-generation hybrid multilingual multi-document news summarization model. Sentence-to-sentence isomorphic graphs and sentence-to-word heteromorphic graphs are used to obtain inter-sentence relation representations and intra-sentence word-level semantic representations, respectively. The importance of candidate sentences is predicted based on these comprehensive representations. Simultaneously, the multilingual sequence-to-sequence generator is used to predict the target language news summary. The extractor loss is the key sentence binary classification cross-entropy loss, the generator loss is the negative log-likelihood loss of the summary sequence, and the extract-generation consistency loss is the KL divergence between the generator's average dynamic sentence weight distribution across all decoding time steps and the extractor's sentence importance distribution. In each training batch, multi-view encoding, key sentence probability prediction, and summary sequence prediction are performed sequentially. The extractor loss, generator loss, and extract-generation consistency loss are calculated, and the total loss is calculated based on the weights of each loss term. The model parameters are updated through one backpropagation. After completing model training and obtaining the trained model parameters, the news summary inference stage begins. In the inference stage, the joint loss is no longer calculated. Instead, the trained homogeneous graph attention network, heterogeneous graph attention network, importance classifier, and multilingual sequence-to-sequence generator are used to generate a target language news summary based on the input mixed-language multi-document news.

[0069] Specifically, the extractor loss is the key sentence binary classification cross-entropy loss, the generator loss is the negative log-likelihood loss of the summary sequence, and the consistency loss is the KL divergence between the average dynamic sentence weight distribution of the generator across all decoding time steps and the sentence importance distribution of the extractor.

[0070] The true labels for key sentences are constructed using a weakly supervised approach: mBERT is used to semantically encode sentences in the source document cluster and sentences in the target summary, the cosine similarity between the source sentences and the target summary sentences is calculated, and the top K source sentences with the highest similarity are labeled as key sentences. For the i-th source sentence... Take its meaning and extract it from the summary sentences of each target. The maximum cosine similarity is used as the weak supervision score. If i belongs to The largest set of the top K source sentence indices is the key sentence tag. =1, otherwise =0.

[0071] In Step 6, the model is jointly optimized by a weighted sum of the extractor cross-entropy loss, the generator negative log-likelihood loss, and the extract-generator KL divergence consistency loss.

[0072] The extraction-generative hybrid language multi-document news summarization method proposed in this application addresses the problems in existing technologies, such as difficulty in fully modeling cross-language and cross-document relationships, loss of key information due to long inputs, translation cascade errors, and inconsistency between extraction and generation objectives.

[0073] The beneficial effects of this application are: 1. By using multiple views composed of sentence-sentence isomorphism and sentence-word heteromorphism, we can simultaneously describe the global relationships between sentences and the fine-grained semantic structure at the word level within sentences, thereby improving the ability to identify key information across languages ​​and documents.

[0074] 2. By extracting key sentences from Top-K, the input of long multi-document documents is compressed into high information density segments, reducing the context length and irrelevant noise of the generator. Furthermore, it does not require all source documents to be pre-translated into a single language, thus reducing translation cascade errors and computational overhead.

[0075] 3. The generator dynamically determines the contribution of different key sentences at each decoding time step, and improves the information coverage, coherence and language fluency of the summary by marginalizing and aggregating the conditional generation probabilities of multiple sentences.

[0076] 4. Consistency loss aligns the generator's dynamic sentence selection with the extractor's global importance judgment, reducing the two-stage target bias and making the generated results traceable back to the selected key sentences, thus improving the model's interpretability.

[0077] The following section provides a detailed explanation of the extract-generative hybrid language multi-document news summarization method proposed in this application.

[0078] like Figure 2 and Figure 3 As shown, an extract-generative hybrid language multi-document news summarization method is described, with the following specific steps: a1. Obtain a multi-document cluster of mixed languages Where M is the number of news documents in the document cluster. , , These represent the language identifiers of the corresponding news documents, and At least not exactly the same; the language set may include English, French, German, and Spanish. Sentence segmentation of news documents in a mixed-language multi-document family is obtained. Extracting the vocabulary set .in, For language identification The first news document, For language identification The second news document, For language identification The Mth news document, This is the first sentence. This is the second sentence. For the nth sentence, The first word, The second word, For the k-th word.

[0079] Specifically, a1 includes the following steps: a1-1. Perform Unicode normalization, whitespace character merging, and webpage tag cleaning on each news document. Select the corresponding sentence boundary recognition rules according to the document language identifier, using period, question mark, exclamation mark, and other terminators as the basis, and combine abbreviation list and number expression rules to eliminate incorrect segmentation. Record the document number and sentence sequence number of each sentence according to the original text position.

[0080] a1-2. The WordPiece segmenter of mBERT is used to segment each sentence. Consecutive words belonging to the same original vocabulary are treated as a single lexical unit. Punctuation and stop words are filtered out, while names of people, places, organizations, dates, and values ​​are retained. After deduplicating lexical units across all documents, a vocabulary set is formed. Sentences exceeding a preset length are segmented and encoded using a sliding window, and the overlapping areas of each window are averaged.

[0081] a1-3. After adding start and end tags to the word sequence of the i-th sentence, input it into the same mBERT encoder to obtain the contextualized hidden state H. i ; For H i The effective lexical units in the text are subjected to average pooling, and then mapped to d dimensions through a shared linear projection layer before being L2 normalized to obtain the initial sentence representation h. i (0) For the j-th word, the hidden states of its sub-words corresponding to each occurrence within the document cluster are first averaged within the word, then averaged over the occurrence counts. This is then mapped to d dimensions via a shared vocabulary projection layer and L2 normalized to obtain the initial vocabulary representation e. (0) Because all languages ​​share the same multilingual encoder and projection parameters, h i (0) and e (0) It resides in a unified cross-linguistic semantic space.

[0082] a2. Constructing the isomorphic graph Each sentence is designated as a node, and undirected edges are created for sentence pairs that share common vocabulary or have a similarity threshold reaching a preset threshold. mBERT is then used to obtain the initial representations of the isomorphic graph sentence nodes. .

[0083] a3. Constructing a heterogeneous graph ,in For the sentence Vocabulary contained Establish sentence-word connection edges, perform word frequency statistics in the mBERT unified semantic space, and calculate TF-IDF weights as the weights of the sentence-word connection edges.

[0084] a4. Initialize vocabulary node representations using mBERT. For sentence nodes, all words in the sentence are first input into mBERT to obtain semantic representations. Then Inputting the data into a convolutional neural network and a bidirectional long short-term memory network, respectively, concatenates the local patterns extracted by the convolutional neural network with the contextual dependencies extracted by the bidirectional long short-term memory network to obtain the initial representation of the sentence nodes in the heterogeneous graph. .

[0085] a5. Perform sentence-to-sentence attention propagation on the isomorphic graph, aggregate the semantic features of adjacent sentences according to attention weights, and obtain updated sentence representations (i.e., inter-sentence relation representations) through a feedforward neural network and residual connections: In the formula, This represents a multi-head attention aggregation operation on an isomorphic graph. Sentence representations that contain inter-sentence relationships; in the isomorphic graph self-attention mechanism, queries, keys, and values ​​all come from the same layer of sentence node representations. This is the initial representation set for sentence nodes. This represents the updated sentence node representation at level 1. These are inter-sentence messages obtained through aggregation using an isomorphic graph attention network. This serves as the initial representation for sentence nodes at level 0. It is a feedforward neural network.

[0086] In a reproducible parameter setting, the isomorphic graph attention network consists of two layers, each with eight attention heads, and each attention head has an output dimension of 64. The outputs of the eight attention heads are concatenated to form a 512-dimensional representation, which is then projected through a linear layer. Attention scoring uses LeakyReLU activation with a negative half-axis slope of 0.2, and the random inactivation rate for attention coefficients and features is set to 0.1. Residual connections and layer normalization are performed sequentially after each layer.

[0087] a6. Perform n rounds of bidirectional message propagation on the heterogeneous graph. In the (t+1)th round, first aggregate from sentence nodes to word nodes, then aggregate the updated word node information back to sentence nodes: In the formula, and These represent type-aware attention aggregation operations from sentence nodes to word nodes and from word nodes to sentence nodes, respectively. This represents the aggregated message from the sentence node to the word node in the (t+1)th round. For the vocabulary nodes after the (t+1)th round of updates, Let t be the vocabulary node representation. For the sentence node representation in round t, This represents the vocabulary node after the (t+1)th round update. Let t be the vocabulary node representation. This represents the aggregated message from the word node to the sentence node in the (t+1)th round. For the sentence node after the (t+1)th round of updates, This represents the sentence node after the (t+1)th round update.

[0088] In a reproducible parameter setting, the heterogeneous graph attention network employs two bidirectional message propagation layers. Each layer sequentially includes relation propagation from sentence nodes to word nodes and from word nodes to sentence nodes. Each relation type has four attention heads, each with an output dimension of 64. The multi-head outputs are concatenated and projected onto a 512-dimensional linear layer. Independent query, key, and value transformation matrices are used for each relation type, and normalized TF-IDF edge weights are added as additional biases to the attention logarithms. The random inactivation rate for each propagation layer is set to 0.1, and residual connections and layer normalization are employed.

[0089] a7. Representing sentences in isomorphic views Sentence representation with heterogeneous views Vector concatenation is performed to obtain a comprehensive sentence representation. The Top-K sentences were selected based on their importance scores. .

[0090] In the formula, This represents vector concatenation. For key sentence indexing, Score the importance of sentences. This refers to the operation of selecting the top K sentences from all candidate sentences in descending order of importance score. Based on... The corresponding sentences are pieced together into a set of key segments in the original text order. .

[0091] Specifically, the comprehensive representation of the i-th candidate sentence (i.e., the above text) The importance of the i-th candidate sentence is classified by its components. First, calculate... ,in, , Both are weight matrices. For activation function, , All are biased; then calculate the predicted probability. The training phase is based on As the positive class probability in cross-entropy loss; during the inference stage, all candidate sentences are classified according to... Sort the data in descending order and select the first K. If the probability of the boundary positions is the same, then determine the order according to the original news document time sequence, document number, and sentence sequence number in turn, so as to ensure that the key sentence selection results can be reproduced.

[0092] a8, Yes Each sentence Perform mBART encoding and autoregressive decoding. In the... Each decoding time step will reveal the hidden state. Input language model head to get The dynamic sentence weights are obtained by inputting them into a multilayer perceptron. Margin out all key sentences: .

[0093] in, For a collection of key fragments The marginalization generation probability of the target summary y under certain conditions. This is a series of multiplications over all decoding time steps from t=1 to T. The summary words are defined for decoding time step t; probabilities are generated based on this marginalization. By decoding word by word until an end marker is generated, a news summary in the target language can be obtained.

[0094] The generator described above can integrate the semantic contributions of different key sentences at each time step, reducing bias towards a single sentence or document.

[0095] a9. Construction of extractor loss The key sentence extraction is modeled as a binary classification task, based on the predicted probability. Calculate the cross-entropy loss with the key sentence label z: .

[0096] In this process, source document sentences and target summary sentences are encoded using mBERT, and cosine similarity is calculated. The top K source sentences with the highest similarity to the target summary sentence are set to z=1, and the remaining sentences are set to z=0, thus obtaining training labels in a weakly supervised manner.

[0097] The process of determining the key sentence label z is as follows: For the source sentence And each sentence in the human target summary The same mBERT encoder and pooling method as the preprocessing stage are used to obtain... and ;Will The largest set Ω consists of the top K source sentences. K When i∈Ω K Season z i =1, otherwise let z =1. i =0. In the event of a tie, the score with the higher ROUGE-L of the target summary sentence is selected first, and then determined according to the original position of the source sentence; at least one positive label is retained for each document cluster.

[0098] a10, Generator Construction Loss and consistency loss : .

[0099] .

[0100] in, This refers to the KL divergence (Kullback-Leibler divergence). The sequence length of the target summary (maximum number of decoding time steps). This is the decoding time step index.

[0101] Consistency loss The formula aligns the time-averaged dynamic weight distribution of the generator for key sentences with the sentence importance distribution output by the extractor using KL divergence.

[0102] a11. Construct the total loss function using the following formula: .

[0103] In the formula, , and These are the weighting coefficients for extraction loss, generation loss, and consistency loss, respectively. Perform backpropagation to complete the joint optimization of the homogeneous GAT, heterogeneous GAT and mBART generators.

[0104] The specific process of joint optimization is as follows: For a training batch, the homogeneous graph attention network and the heterogeneous graph attention network first output a comprehensive sentence representation and... Then, based on the training labels or the current Top-K results, the generator input is formed, and mBART calculates the word sequence probability of the target summary and the dynamic sentence weights at each decoding time step; subsequently, the calculation... , and And weighted summation is After clearing the gradient of the previous batch, Perform one backpropagation. Since the Top-K index is chosen as a discrete operation, the index itself does not participate in the differentiation; as well as The gradient on one side of the extractor distribution is used to update the isomorphic graph attention network, the heteromorphic graph attention network, and the importance classifier. as well as The gradient on one side of the generator distribution is used to update the mBART and the dynamic sentence weight multilayer perceptron, thereby achieving collaborative training under the same objective function.

[0105] The optimizer can use AdamW, with different learning rates set for the extraction network parameters and generator parameters; the global norm of the gradient is pruned, linear warm-up is used for the first few training steps, followed by linear decay, and early stopping is performed based on the ROUGE-L metric of the validation set. During training, sampling can be scheduled between the set of true key sentences and the model's predicted Top-K set according to a preset probability, allowing the generator to gradually adapt to the extraction error during the inference phase.

[0106] The model parameters and effects will be verified next.

[0107] In one specific implementation, the MLMD-news dataset, consisting of English, French, German, and Spanish, is used for both training and testing, with English as the target summarizing language. Set to 1, Set it to 0.1. Set to 0.0001; set the decimator learning rate to 5×10. -6 The generator learning rate is set to 5×10. -5The batch size was set to 8, and K was set to 10. The uniform dimension d of sentence and word representations was set to 512; the isomorphic graph attention network was set to 2 layers with 8 attention heads per layer, and the heteromorphic graph attention network was set to 2 bidirectional message propagation layers with 4 attention heads for each relation; the output dimension of each attention head was set to 64, the random inactivation rate of attention and features was set to 0.1, the upper limit of the global norm of the gradient was set to 1.0, and the number of linear warm-up steps accounted for 10% of the total training steps.

[0108] Table 1. Partial comparison results on the MLMD-news dataset.

[0109] As shown in Table 1, the ROUGE-1, ROUGE-2, and ROUGE-L values ​​of the method in this application are 39.16, 9.64, and 34.02, respectively, which are higher than the extraction-generation methods that directly use mBART and use a combination of TextRank or Centroid and mBART. This indicates that the multi-view key sentence extraction and dynamic marginalization generation can improve the information coverage and semantic coherence of the summary.

[0110] Table 2. Module Performance Verification Table

[0111] In Table 2, after removing the extractor, the three ROUGE metrics decreased by 2.32, 1.51, and 1.80, respectively; after removing the generator, they decreased by 4.99, 1.19, and 2.98, respectively; and after removing the consistency loss, they decreased by 0.35, 0.23, and 0.17, respectively. These results indicate that key sentence extraction, generative rewriting, and extract-generation consistency constraints all have a positive impact on the model's performance.

[0112] When performing sensitivity analysis on the K value, the ROUGE index increases within a certain range as K increases; a better balance between information sufficiency and redundancy control is achieved when K=10; further increasing K will introduce secondary or irrelevant information and interfere with the generator's attention allocation. Therefore, in a specific dataset, the K value can be adaptively selected based on the summary evaluation index on the validation set.

[0113] This application performs sentence segmentation and unified semantic encoding on document clusters containing news documents in multiple languages. It constructs sentence-sentence isomorphic graphs with sentences as nodes and sentence-word heteromorphic graphs with sentences and words as heteromorphic nodes, respectively, and determines graph connections and edge weights within a unified cross-linguistic semantic space. The isomorphic and heteromorphic graphs are input into isomorphic and heteromorphic graph attention networks, respectively, to obtain inter-sentence relation representations and intra-sentence word-level semantic representations. After fusing the two representations, candidate sentences are scored for importance, and the top K sentences are selected and arranged in the original text order to form key segments. These key segments are input into a multilingual sequence generator, which uses dynamic sentence weights and a marginalization mechanism to generate target language summaries, and optimizes the process through joint extraction loss, generation loss, and consistency loss. This application can structurally model cross-document and cross-linguistic relationships, reduce input length and translation cascade errors, and improve key information coverage, summary coherence, and the interpretability of the generation process.

[0114] Based on the same inventive concept, this application also provides an extraction-generation hybrid language multi-document news summarization system for implementing the extraction-generation hybrid language multi-document news summarization method described above. The solution provided by this system is similar to the implementation described in the above method. Therefore, the specific limitations of one or more extraction-generation hybrid language multi-document news summarization system embodiments provided below can be found in the limitations of the extraction-generation hybrid language multi-document news summarization method described above, and will not be repeated here.

[0115] In one exemplary embodiment, an extract-generative hybrid language multi-document news summarization system is provided, comprising: sentence set and vocabulary set units, multi-view construction unit, representation unit, fusion unit, key sentence extraction unit, and summary generation unit.

[0116] The sentence set and vocabulary set units are used to obtain a mixed-language multi-document cluster consisting of multiple news documents containing at least two languages. Each news document is segmented into sentences, words, and subjected to cross-language unified semantic encoding to obtain the sentence set and vocabulary set.

[0117] The multi-view building unit is used to construct sentence-sentence isomorphism graphs based on sentence sets, and to construct sentence-word heteromorphism graphs based on sentence sets and vocabulary sets.

[0118] The representation unit is used to input sentence-sentence isomorphism graphs into the isomorphism graph attention network to obtain inter-sentence relation representations, and to input sentence-word heteromorphism graphs into the heteromorphism graph attention network to obtain intra-sentence word-level semantic representations.

[0119] The fusion unit is used to fuse inter-sentence relation representations and intra-sentence word-level semantic representations to obtain a comprehensive representation and importance score for each candidate sentence.

[0120] The key sentence extraction unit is used to extract multiple key sentences based on importance scores and form a set of key segments.

[0121] The summary generation unit is used to encode and decode a set of key segments based on a multilingual sequence-to-sequence generator to obtain a news summary in the target language.

[0122] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 4 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores extractive-generative hybrid language multi-document news summaries. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements an extractive-generative hybrid language multi-document news summarization method.

[0123] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0124] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0125] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0126] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0127] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0128] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0129] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0130] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0131] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for extracting-generating hybrid language multi-document news summarization, characterized in that, include: Obtain a mixed-language multi-document cluster consisting of multiple news documents containing at least two languages, and perform sentence segmentation, word segmentation, and cross-language unified semantic encoding on each news document to obtain a sentence set and a vocabulary set; A sentence-sentence isomorphism graph is constructed based on the sentence set, and a sentence-word heteromorphism graph is constructed based on the sentence set and the vocabulary set; Inputting sentence-sentence isomorphism graphs into isomorphism graph attention networks yields inter-sentence relation representations, while inputting sentence-word heteromorphism graphs into heteromorphism graph attention networks yields intra-sentence word-level semantic representations. By integrating inter-sentence relation representations and intra-sentence word-level semantic representations, a comprehensive representation and importance score for each candidate sentence are obtained. Multiple key sentences were extracted based on importance scores and combined into a set of key segments; The key fragment set is encoded and decoded based on a multilingual sequence-to-sequence generator to obtain a news summary in the target language.

2. The extraction-generative hybrid language multi-document news summarization method according to claim 1, characterized in that, Each news document is segmented into sentences, words, and subjected to cross-language unified semantic encoding to obtain a sentence set and a vocabulary set, specifically including: The news document is segmented into multiple sentences by calling the corresponding sentence delimitation rules based on its language identifiers. A word segmenter is used to segment each sentence into words, and each sentence is converted into a corresponding word sequence; The word sequence of each sentence is input into the same multilingual encoder to obtain the hidden state of each word. The hidden sub-states in the sentence are averaged and then mapped to a unified-dimensional cross-lingual semantic space through a cross-language shared sentence projection layer and L2 normalized to obtain the initial sentence representation of each sentence in the sentence set. The hidden states of multiple sub-words belonging to the same word are aggregated, mapped to a unified cross-linguistic semantic space through a shared word-level projection layer, and then normalized by L2 to obtain the initial lexical representation of each word in the vocabulary set.

3. The extraction-generative hybrid language multi-document news summarization method according to claim 1, characterized in that, Constructing sentence-sentence isomorphism graphs based on sentence sets, and constructing sentence-word heteromorphism graphs based on sentence sets and vocabulary sets, specifically including: Using each sentence in the sentence set as a node, and sentence pairs with common vocabulary or sentence pairs with semantic similarity reaching a preset threshold in a unified semantic space as undirected edges, a sentence-sentence isomorphism graph is obtained. Each sentence in the sentence set is used as a sentence node, and each word in the vocabulary set is used as a vocabulary node. When the sentence corresponding to a sentence node contains the word corresponding to a vocabulary node, a sentence-word connection edge is established between the sentence node and the vocabulary node to obtain a sentence-word heterogeneous graph.

4. The extraction-generative hybrid language multi-document news summarization method according to claim 3, characterized in that, The isomorphic graph attention network employs a multi-head attention mechanism to weighted aggregate the mapping features of adjacent sentence nodes, and obtains the inter-sentence relationship representation through a feedforward neural network and residual connections.

5. The extraction-generative hybrid language multi-document news summarization method according to claim 3, characterized in that, The heterogeneous graph attention network obtains intra-sentence word-level semantic representations through bidirectional message propagation from sentence nodes to word nodes and from word nodes to sentence nodes.

6. The extraction-generative hybrid language multi-document news summarization method according to claim 1, characterized in that, The multilingual sequence-to-sequence generator is mBART; the key segment set is encoded and decoded based on the multilingual sequence-to-sequence generator to obtain a news summary in the target language, specifically including: Each key sentence in the key segment set is encoded using mBART to obtain the hidden state of each key sentence. At the current decoding time step, the hidden state of each key sentence is input into the language model head to obtain the condition word generation probability of each key sentence, and then input into the multilayer perceptron to obtain the dynamic sentence weight of each key sentence at the current decoding time step. The generation probabilities of condition words in each key sentence are weighted and summed according to the dynamic sentence weights to obtain the marginalization generation probability based on the key segment set. Then, word-by-word decoding is performed according to the marginalization generation probability until the end marker is generated to obtain the target language news summary.

7. An extract-generate hybrid language multi-document news summarization system, characterized in that, include: The sentence set and vocabulary set unit is used to obtain a mixed-language multi-document cluster consisting of multiple news documents containing at least two languages. Each news document is segmented into sentences, words, and cross-language unified semantic encoding to obtain the sentence set and vocabulary set. A multi-view building unit is used to construct sentence-sentence isomorphism graphs based on sentence sets, and to construct sentence-word heteromorphism graphs based on sentence sets and vocabulary sets; The representation unit is used to input sentence-sentence isomorphism graphs into the isomorphism graph attention network to obtain inter-sentence relation representations, and to input sentence-word heteromorphism graphs into the heteromorphism graph attention network to obtain intra-sentence word-level semantic representations. The fusion unit is used to fuse inter-sentence relation representations and intra-sentence word-level semantic representations to obtain a comprehensive representation and importance score for each candidate sentence; The key sentence extraction unit is used to extract multiple key sentences based on their importance scores and form a set of key segments. The summary generation unit is used to encode and decode a set of key segments based on a multilingual sequence-to-sequence generator to obtain a news summary in the target language.

8. A computer device, comprising: A memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that the processor executes the computer program to implement the extract-generative hybrid language multi-document news summarization method according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the extract-generative hybrid language multi-document news summarization method as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the extract-generative hybrid language multi-document news summarization method as described in any one of claims 1-6.