Multi-source document automatic abstracting method and device, equipment and storage medium

Abstract

The invention provides a multi-source document automatic abstracting method and device, equipment and a storage medium, and the method comprises the steps: extracting information from a multi-source document, and constructing an initial knowledge graph; obtaining an inference rule based on the initial knowledge graph, performing automatic inference according to the inference rule, and adding or labeling an inference result in the initial knowledge graph to obtain a first knowledge graph; taking an entity needing to generate an automatic abstract as a key entity, extracting an event corresponding to the key entity from the multi-source document, constructing a relationship between the extracted event and the key entity, and adding the constructed relationship into the first knowledge graph to obtain a second knowledge graph; traversing all events associated with the key entity in the second knowledge graph, calculating importance scores of all the events, and screening part of the events as initial abstracts according to the importance scores; and inputting the initial abstract into a pre-trained editing network to obtain a final abstract.

Description

Technical Field

[0001] The present invention relates to the field of document recognition technology, and in particular to a multi-source document automatic summarization method, device, equipment and storage medium. Background Art

[0002] Automatic web document summarization is an important topic related to information retrieval, natural language processing and artificial intelligence research. Its purpose is to create a method that can understand and refine original text information so that people can obtain the content of the document in a short time.

[0003] In the related technologies, early automatic document summarization was mainly based on statistical methods and was implemented by using simple rules, such as identifying the most frequently occurring words or selecting key sentences containing important information to generate summaries. This easily ignored the context and structure of the document, resulting in fragmented or incoherent summary information. Automatic document summarization technology based on machine learning has been widely studied in recent years. This type of technology uses machine learning models to learn the rules of summary generation from annotated data, but the above technical means still have the following defects: it is easy to miss important information in the original text, and if information needs to be extracted or integrated from multiple documents, it is difficult to handle complex information relationships and conflicts across documents.

[0004] Based on the above analysis of the development status of this technology field, the existing technology lacks a solution for automatic summarization of multi-source documents based on knowledge graphs and optimization through editing networks. Summary of the invention

[0005] The purpose of the present invention is to provide a multi-source document automatic summarization method, device, equipment and storage medium, aiming to solve the above-mentioned problems in the prior art.

[0006] According to a first aspect of an embodiment of the present invention, a multi-source document automatic summarization method is provided, comprising:

[0007] Extract information from multi-source documents and build an initial knowledge graph;

[0008] Obtaining inference rules based on the initial knowledge graph, performing automatic inference according to the inference rules, and adding or annotating the inference results to the initial knowledge graph to obtain a first knowledge graph;

[0009] The entity for which automatic summary is to be generated is taken as the key entity, events corresponding to the key entity are extracted from multi-source documents, the relationship between the extracted event and the key entity is constructed, and the constructed relationship is added to the first knowledge graph to obtain the second knowledge graph;

[0010] Traverse all events associated with the key entity in the second knowledge graph, calculate the importance scores of all events, and select some events as the initial summary according to the importance scores;

[0011] The initial summary is fed into a pre-trained editing network to obtain the final summary.

[0012] According to a second aspect of an embodiment of the present invention, a multi-source document automatic summarization device is provided, comprising:

[0013] The initial building module is used to extract information from multi-source documents and build an initial knowledge graph;

[0014] An automatic reasoning module, used to obtain reasoning rules based on the initial knowledge graph, perform automatic reasoning according to the reasoning rules, and add or annotate the reasoning results to the initial knowledge graph to obtain a first knowledge graph;

[0015] An event association module is used to take the entity for which automatic summary is to be generated as the key entity, extract the events corresponding to the key entity from the multi-source documents, construct the relationship between the extracted event and the key entity, and add the constructed relationship to the first knowledge graph to obtain the second knowledge graph;

[0016] An initial summary generation module is used to traverse all events associated with the key entity in the second knowledge graph, calculate the importance scores of all events, and select some events as initial summaries according to the importance scores;

[0017] The final summary generation module is used to input the initial summary into the pre-trained editing network to obtain the final summary.

[0018] According to a third aspect of an embodiment of the present invention, an electronic device is provided, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, the steps of the multi-source document automatic summarization method provided in the first aspect of the present disclosure are implemented.

[0019] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which a program for implementing information transmission is stored. When the program is executed by a processor, the steps of the multi-source document automatic summarization method provided in the first aspect of the present disclosure are implemented.

[0020] The technical solution provided by the embodiment of the present invention includes the following beneficial effects: obtaining information from multi-source documents to establish an easy-to-understand initial knowledge graph from rich data sources; and automatically reasoning and supplementing the initial knowledge graph with event information in sequence to enrich the information contained in the initial knowledge graph and avoid missing important information in multi-source documents; screening important events as initial summaries, and optimizing the expression of the initial summaries through editing networks to enhance the readability of the generated final summary.

[0021] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate one or more embodiments of this specification or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0023] Figure 1 is a flowchart of a multi-source document automatic summarization method according to an embodiment of the present invention;

[0024] Figure 2 is a schematic diagram of a multi-source document automatic summarization device according to an embodiment of the present invention;

[0025] Figure 3 is a schematic diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the following will be combined with the drawings in one or more embodiments of this specification to clearly and completely describe the technical solutions in one or more embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this document.

[0027] Method Example 1

[0028] According to an embodiment of the present invention, a multi-source document automatic summarization method is provided. Figure 1 is a flowchart of a multi-source document automatic summarization method according to an embodiment of the present invention. Figure 1 As shown, the multi-source document automatic summarization method according to an embodiment of the present invention specifically includes:

[0029] In step S110, information is extracted from multi-source documents to construct an initial knowledge graph, which specifically includes:

[0030] Use the pre-trained named entity recognition (NER) model to extract entities from multi-source documents. The recognition of entity i can be expressed by the conditional random field formula, as shown in Formula 2:

[0031]

[0032] Among them, P(y|x) represents the conditional probability distribution, x represents the input sequence, y represents the labeled sequence, n represents the length of the input sequence, k and l represent the number of feature functions, λ and μ represent feature-related parameters, i and j represent index subscripts, Z(x) represents the normalization factor, f and f' represent feature functions, and exp represents the exponential function.

[0033] Use the pre-trained BERT model to identify the relationship between entities. Based on the identified entities and relationships, use AI technology to form an initial recognition graph in the form of a graph. Use the minimum spanning tree algorithm to determine the connection weights between entities in the initial knowledge graph. Use Formula 3 to represent the weight of the edge:

[0034]

[0035] Among them, w ij represents the connection weight between entity i and entity j, d ij Represents the distance between entity i and entity j;

[0036] Preferably, the initial knowledge graph is established for long-term use, and the importance of entities is examined over time using time series. The knowledge graph is optimized and evolved to adapt to the changes. This can be achieved by calculating the centrality of entities in different periods. The centrality score is calculated using Formula 4:

[0037] C d(i) =∑(j∈N) A ij Formula 4;

[0038] Among them, C d(i) represents the centrality score of entity i, N represents the set of all nodes in the network, A ij A represents the weight of the edge connecting entity i and entity j, i≠j, and when two entities are connected ij =1, otherwise equal to 0.

[0039] Multi-source documents refer to multiple documents on the same topic, which are used to understand semantic relationships across documents and determine which information is useful for summaries. Building an initial knowledge graph is a rough extraction and integration of multi-source documents, which can not only identify the entities and relationships of the documents, but also understand the situations of these entities and relationships in each document.

[0040] In step S120, an inference rule based on the initial knowledge graph is obtained, automatic inference is performed according to the inference rule, and the inference result is added or annotated in the initial knowledge graph to obtain a first knowledge graph, which specifically includes:

[0041] In the initial knowledge graph, causal reasoning, DPLL logical reasoning and Bayesian probabilistic reasoning are performed in sequence according to the inference rules.

[0042] In the process of constructing the initial knowledge graph, some causal relationships have been organized, and the causal graph method can be used for reasoning. For example, if the reasoning rule includes "A causes B", then when B is not observed in the initial knowledge graph but A is observed, it can be inferred that B may occur. In other words, it is assumed that there is a causal relationship between the two entities, and potential reasoning is performed through the state of one of them occurring or not occurring.

[0043] Logical reasoning is based on Boolean logic. For example, for two propositions A and B, there are four basic conjunctions: "and", "or", "not", and "implies":

[0044] AND: If both A and B are true, the result of “AND” is true, otherwise it is false;

[0045] Or: If at least one of A and B is true, the result of "or" is true, otherwise it is false;

[0046] Not: If A is true, then the result of "not" is false, otherwise it is true;

[0047] Implication: If A is true and B is false, the result of "implication" is false, otherwise it is true. A implies B is equivalent to not A or B;

[0048] If the inference rule contains "if A happens, then B must happen", if we know that "A happened", we can infer that "B happened". In order to be able to infer, we need to add edges to the initial knowledge graph to represent information. For example, if it happened, add True information to the edge, and then use the graph reasoning algorithm DPLL for reasoning.

[0049] The Bayesian probability inference formula is P(A|B)=[P(B|A)×P(A)] / P(B), which represents the conditional probability of A occurring when B occurs. Suppose a news report says that an earthquake occurred in a certain city. Based on the probability of earthquakes in the city’s history and the accuracy of news reports on earthquakes, we can infer whether an earthquake really occurred in this city.

[0050] In the embodiment of the present invention, the inference rules may be derived from multiple sources, including but not limited to the experience of domain experts, literature and statistical data analysis, etc.

[0051] In step S130, the entity for which automatic summary is to be generated is taken as the key entity, events corresponding to the key entity are extracted from the multi-source documents, the relationship between the extracted event and the key entity is constructed, and the constructed relationship is added to the first knowledge graph to obtain the second knowledge graph, which specifically includes:

[0052] Generate summaries from the perspective of entities. Generally, summaries are generated in sentences. Generating summaries from the perspective of entities can sort out more relevant content. For example, for a summary of political news, it can be generated around several key political figures or organizations, connecting their related events in multi-source documents.

[0053] The first knowledge graph performs automatic reasoning based on the rough integration of multi-source document information. Compared with the initial knowledge graph, it contains more information. At this time, the graph contains relevant information about certain events, but it is not detailed enough. Therefore, it is necessary to extract events corresponding to key entities from multi-source documents to optimize the current graph.

[0054] Use the trained neural network model to extract events corresponding to key entities. Preferably, support vector machines or hidden Markov models can also be used. The model training is based on the annotated corpus, which annotates the information of events and their participants. The identified events are associated with entities in the first knowledge graph by adding nodes or edges. Each node can be an entity or an event. If a node participates in an event, then there will be an edge from the entity node to the event node in the network.

[0055] Preferably, for each event node, additional information including the time and place of the event will also be recorded to facilitate subsequent application in the summary; the connection between nodes and events is called an entity event network, which is established in the first knowledge network diagram in the embodiment of the present invention, and the relevant parts can also be separated.

[0056] In step S140, all events associated with the key entity in the second knowledge graph are traversed, importance scores of all events are calculated, and some events are selected as initial summaries according to the importance scores, specifically including:

[0057] The importance score of the event relative to the key entity is calculated using Formula 1:

[0058] s(e) = w1×I(e) + w2×T(e) + w3×D(e) Formula 1;

[0059] Among them, s(e) represents the importance score of event e, I(e) represents the number of entities involved in event e, that is, the number of directly connected entities, T(e) represents the time corresponding to event e, D(e) represents the distribution of events in multi-source documents, such as the frequency of occurrence, etc., and w1, w2 and w3 represent weight parameters.

[0060] Based on the importance score from high to low, a preset number of events are screened as candidate events, and the candidate events are sorted in chronological order from far to near to obtain an initial summary; importance score is used as the first principle and time is used as the second principle. If you do not pay attention to sorting in chronological order, time confusion may lead to poor logic.

[0061] In step S150, the initial summary is input into the pre-trained editing network to obtain the final summary, which specifically includes:

[0062] The text description in the initial abstract needs to be optimized to improve readability and abstract effectiveness;

[0063] Generate the final summary through the pre-trained Seq2Seq editing network;

[0064] Among them, the Seq2Seq editing network includes an encoder and a decoder. The encoder obtains the sequence of initial summaries, converts the sequence into a vector, receives the vector and the partial summary generated in the previous time step through the decoder, and generates the word corresponding to the current time step. Because only one word is generated in each time step, the word is added to the partial summary as a new partial summary and input into the decoder, and it is iterated repeatedly until the final summary is generated.

[0065] In the training process of the editing network, the original summary and the edited and annotated summary are used as sequence pairs (SR, SE) to train the model, where SR = {r1, r2, ..., rn}, SE = {e1, e2, ..., em}, each item can be regarded as a sentence or word, depending on the needs. The cross entropy loss function is used in the training process. For the sequence pair, the goal of the editing network is to maximize the conditional probability of generating the sequence SE. The conditional probability L is shown in Formula 5:

[0066] L = Σ(log P(ek|e1:e(k-1), SR)) for k = 1 to m Formula 5;

[0067] Where P represents the form of conditional probability, e1:e(k-1) represents e1,e2,...,e(k-1), k=1to m represents k from 1 to m, and k represents the index subscript;

[0068] Edit the network model to generate the final summary That is, maximize the conditional probability This is done by greedy selection or beam search from left to right.

[0069] Method Example 2

[0070] Assume that there are four news documents from different websites, all of which are about Apple producing new iPhones. These four documents are considered as multi-source documents.

[0071] Extract entities such as "Apple", "iPhone", "TimCook" from multi-source documents, identify the relationships between entities, such as "Apple" is the producer of "iPhone", "TimCook" is the CEO of "Apple", etc., and finally form these entities and relationships into an initial recognition graph;

[0072] Obtain inference rules and perform reasoning on the initial recognition graph. Causal reasoning shows that "Apple" producing "iPhone" may lead to an increase in stock prices. Logical reasoning shows that "Tim Cook" is also the producer of "iPhone". Probabilistic reasoning shows that based on the historical sales data of "Apple" and "iPhone", the new iPhone has achieved great sales success. After reasoning, the first knowledge graph is obtained.

[0073] In the embodiment of the present invention, the key entity is "iPhone", and the event is obtained, and the second knowledge graph is obtained by combining the time and place of the event. The event with a higher importance score is selected as the initial summary, such as "Apple releases a new iPhone" and "The new iPhone has achieved great sales success", which are respectively used as the first event and the second event in the initial summary in chronological order;

[0074] The pre-trained editing network is used to generate the final summary corresponding to the initial summary. Taking the initial summary "The new iPhone has achieved huge sales success" as an example, the final summary generated is "Apple has launched a new iPhone, which has received a warm market response and achieved great sales success." The generated final summary is more humanistic. It should be noted that during the optimization process, the events do not interfere with each other, that is, the order of the final summary events remains unchanged, and only adjustments are made within each event sentence.

[0075] To sum up, in response to the existing problems, this invention proposes an automatic summarization method for multi-source documents, which obtains information from multi-source documents to establish an easy-to-understand initial knowledge graph from rich data sources; and performs automatic reasoning and event information supplement on the initial knowledge graph. Automatic reasoning includes causal reasoning, logical reasoning and probabilistic reasoning, which can deduce new conclusions based on known facts and rules in multi-source documents. Event information supplement is driven by key entities, around which related events in multi-source files can be connected in series. Automatic reasoning and event supplement enrich the information contained in the initial knowledge graph to avoid missing important information in multi-source documents; important events are screened and sorted by time as the initial summary. While considering the importance score, the order of events is used to improve the logic; and the expression of the initial summary is optimized by editing the network to enhance the readability of the generated final summary.

[0076] Device Embodiment

[0077] According to an embodiment of the present invention, a multi-source document automatic summarization device is provided. Figure 2 Schematic diagram of a multi-source document automatic summarization device according to an embodiment of the present invention. Figure 2 As shown, the multi-source document automatic summarization device according to an embodiment of the present invention specifically includes:

[0078] The initial construction module 20 is used to extract information from multi-source documents and construct an initial knowledge graph, specifically for:

[0079] Use the named entity recognition (NER) model to extract entities from multi-source documents, use the BERT model to identify the relationships between entities, form an initial recognition graph in the form of a graph based on the identified entities and relationships, and use the minimum spanning tree algorithm to determine the connection weights between entities in the initial knowledge graph.

[0080] The automatic reasoning module 22 is used to obtain reasoning rules based on the initial knowledge graph, perform automatic reasoning according to the reasoning rules, and add or annotate the reasoning results to the initial knowledge graph to obtain a first knowledge graph, specifically for:

[0081] In the initial knowledge graph, causal reasoning, DPLL logical reasoning and Bayesian probabilistic reasoning are performed in sequence according to the inference rules.

[0082] The event association module 24 is used to take the entity for which automatic summary is to be generated as the key entity, extract the events corresponding to the key entity from the multi-source documents, construct the relationship between the extracted event and the key entity, and add the constructed relationship to the first knowledge graph to obtain the second knowledge graph, which is specifically used for:

[0083] Use a neural network model to extract events corresponding to key entities.

[0084] The initial summary generation module 26 is used to traverse all events associated with the key entity in the second knowledge graph, calculate the importance scores of all events, and select some events as initial summaries according to the importance scores, specifically for:

[0085] The importance score of the event relative to the key entity is calculated using Formula 1:

[0086] s(e) = w1×I(e) + w2×T(e) + w3×D(e) Formula 1;

[0087] Among them, s(e) represents the importance score of event e, I(e) represents the number of entities involved in event e, T(e) represents the time corresponding to event e, D(e) represents the distribution of events in multi-source documents, and w1, w2 and w3 represent weight parameters.

[0088] Based on the importance scores from high to low, a preset number of events are screened as candidate events, and the candidate events are sorted in chronological order from far to near to obtain an initial summary.

[0089] The final summary generation module 28 is used to input the initial summary into the pre-trained editing network to obtain the final summary, specifically for:

[0090] Generate the final summary through Seq2Seq editing network;

[0091] Among them, the Seq2Seq editing network includes an encoder and a decoder. The encoder obtains the sequence of initial summaries, converts the sequence into a vector, receives the vector and the partial summary generated by the previous time step through the decoder, generates the word corresponding to the current time step, adds the word to the partial summary as a new partial summary and inputs it into the decoder, and iterates repeatedly until the final summary is generated.

[0092] To sum up, in response to the existing problems, this invention proposes an automatic summarization device for multi-source documents, which obtains information from multi-source documents to establish an easy-to-understand initial knowledge graph from rich data sources; and performs automatic reasoning and event information supplement on the initial knowledge graph. Automatic reasoning includes causal reasoning, logical reasoning and probabilistic reasoning, which can deduce new conclusions based on known facts and rules in multi-source documents. Event information supplement is driven by key entities, around which related events in multi-source files can be connected in series. Automatic reasoning and event supplement enrich the information contained in the initial knowledge graph to avoid missing important information in multi-source documents; important events are screened and sorted by time as the initial summary. While considering the importance score, the events are sorted in the order of occurrence to improve the logic; and the expression of the initial summary is optimized by editing the network to enhance the readability of the generated final summary.

[0093] Electronic device embodiment

[0094] Figure 3 3 is a schematic diagram of an electronic device according to an embodiment of the present invention. The electronic device 300 may include at least one processor 310 and a memory 320. The processor 310 may execute instructions stored in the memory 320. The processor 310 is connected to the memory 320 through a data bus. In addition to the memory 320, the processor 310 may also be connected to an input device 330, an output device 340, and a communication device 350 through a data bus.

[0095] The processor 310 may be any conventional processor, such as a commercially available CPU. The processor may also include a graphics processor (Graphic Process Unit, GPU), a field programmable gate array (Field Programmable Gate Array, FPGA), a system on chip (System on Chip, SOC), an application specific integrated circuit (Application Specific Integrated Circuit, ASIC) or a combination thereof.

[0096] The memory 320 may be implemented by any type of volatile or nonvolatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0097] In the embodiment of the present disclosure, executable instructions are stored in the memory 320, and the processor 310 can read the executable instructions from the memory 320 and execute the instructions to implement all or part of the steps of any multi-source document automatic summarization method in the above exemplary embodiments.

[0098] Computer Readable Storage Medium Embodiments

[0099] In addition to the above-mentioned methods and apparatus, an exemplary embodiment of the present disclosure may also be a computer program product or a computer-readable storage medium storing the computer program product. The computer product includes computer program instructions, and the computer program instructions can be executed by a processor to implement all or part of the steps described in any of the multi-source document automatic summarization methods in the above-mentioned exemplary embodiments.

[0100] The computer program product may be written in any combination of one or more programming languages ​​to write program codes for performing the operations of the embodiments of the present application, including object-oriented programming languages ​​such as Java, C++, etc., and also conventional procedural programming languages ​​such as "C" language or similar programming languages ​​and scripting languages ​​(e.g., Python). The program code may be executed entirely on the user computing device, partially on the user computing device, as an independent software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0101] Computer readable storage media can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can include, for example, but is not limited to, a system, device or device of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination of the above. More specific examples of readable storage media include: a static random access memory (SRAM) with one or more wires electrically connected, an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk or an optical disk, or any suitable combination of the above.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-source document automatic summarization method, characterized in that: include: Extract information from multi-source documents and build an initial knowledge graph; Obtaining an inference rule based on the initial knowledge graph, performing automatic inference according to the inference rule, and adding or annotating the inference result to the initial knowledge graph to obtain a first knowledge graph; The entity for which automatic summary is to be generated is taken as a key entity, events corresponding to the key entity are extracted from the multi-source documents, a relationship between the extracted event and the key entity is constructed, and the constructed relationship is added to the first knowledge graph to obtain a second knowledge graph; Traversing all events associated with the key entity in the second knowledge graph, calculating importance scores of all events, and selecting some events as initial summaries according to the importance scores; The initial summary is input into a pre-trained editing network to obtain a final summary.

2. The method according to claim 1, characterized in that The extracting of information from multi-source documents and constructing an initial knowledge graph specifically includes: Use the named entity recognition (NER) model to extract entities from multi-source documents, use the BERT model to identify the relationships between entities, form an initial recognition graph in the form of a graph based on the identified entities and relationships, and use the minimum spanning tree algorithm to determine the connection weights between entities in the initial knowledge graph.

3. The method according to claim 1, characterized in that The automatic reasoning according to the inference rules specifically includes: in the initial knowledge graph, performing causal reasoning, DPLL logical reasoning and Bayesian probabilistic reasoning in sequence according to the inference rules.

4. The method according to claim 1, characterized in that: The extracting the events corresponding to the key entities from the multi-source documents specifically includes: extracting the events corresponding to the key entities using a neural network model.

5. The method according to claim 1, characterized in that The calculation of the importance scores of all events specifically includes: The importance score of the event relative to the key entity is calculated using Formula 1: s(e) = w1×I(e) + w2×T(e) + w3×D(e) Formula 1; Among them, s(e) represents the importance score of event e, I(e) represents the number of entities involved in event e, T(e) represents the time corresponding to event e, D(e) represents the distribution of events in multi-source documents, and w1, w2 and w3 represent weight parameters.

6. The method according to claim 1, characterized in that The selecting of some events as the initial summary according to the importance score specifically includes: Based on the importance scores from high to low, a preset number of events are screened as candidate events, and the candidate events are sorted in chronological order from far to near to obtain an initial summary.

7. The method according to claim 1, characterized in that The step of inputting the initial summary into a pre-trained editing network to obtain a final summary specifically includes: Generate the final summary through Seq2Seq editing network; The Seq2Seq editing network includes an encoder and a decoder. The encoder obtains a sequence of initial summaries, converts the sequence into a vector, receives the vector and a partial summary generated at a previous time step, generates a word corresponding to a current time step, adds the word to the partial summary and inputs it into the decoder as a new partial summary, and iterates repeatedly until a final summary is generated.

8. A multi-source document automatic summarization device, characterized in that: include: The initial building module is used to extract information from multi-source documents and build an initial knowledge graph; An automatic reasoning module, used to obtain reasoning rules based on the initial knowledge graph, perform automatic reasoning according to the reasoning rules, and add or annotate the reasoning results to the initial knowledge graph to obtain a first knowledge graph; An event association module, used to take the entity for which automatic summary is to be generated as a key entity, extract events corresponding to the key entity from the multi-source documents, construct a relationship between the extracted event and the key entity, and add the constructed relationship to the first knowledge graph to obtain a second knowledge graph; An initial summary generation module, used to traverse all events associated with the key entity in the second knowledge graph, calculate importance scores of all events, and select some events as initial summaries according to the importance scores; The final summary generation module is used to input the initial summary into a pre-trained editing network to obtain a final summary.

9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, the steps of the multi-source document automatic summarization method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores an implementation program for information transmission, and when the program is executed by a processor, the steps of the multi-source document automatic summarization method according to any one of claims 1 to 7 are implemented.

Images