CN116069831B - Event relation mining method and related device - Google Patents
Event relation mining method and related device Download PDFInfo
- Publication number
- CN116069831B CN116069831B CN202310309890.3A CN202310309890A CN116069831B CN 116069831 B CN116069831 B CN 116069831B CN 202310309890 A CN202310309890 A CN 202310309890A CN 116069831 B CN116069831 B CN 116069831B
- Authority
- CN
- China
- Prior art keywords
- event
- graph
- layer
- document
- upper layer
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/24—Querying
- G06F16/245—Query processing
- G06F16/2458—Special types of queries, e.g. statistical queries, fuzzy queries or distributed queries
- G06F16/2465—Query processing support for facilitating data mining operations in structured databases
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Databases & Information Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Probability & Statistics with Applications (AREA)
- Fuzzy Systems (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The application discloses a mining method of event relation and a related device, wherein the method comprises the steps of respectively obtaining document characterization and event characterization of each target document in a plurality of target documents; constructing an upper layer diagram based on the document representation of each target document, and constructing a lower layer diagram based on the event representation of each target document; and determining event relations among event characterizations of each target document based on a pre-trained graph mining model, the upper layer graph and the lower layer graph. According to the method, an upper layer diagram and a lower layer diagram are formed based on the document representation and the event representation of the target documents, then the event relationship is determined through the diagram mining model, so that the event relationship mining process is converted into a diagram completion task, the event relationship among the events can be mined only by marking the event relationship among the target documents, and manpower and material resources divided by the event relationship marking are reduced, so that the cost for mining the event relationship can be reduced.
Description
Technical Field
The application relates to the technical field of natural language processing, in particular to an event relation mining method and a related device.
Background
In the internet age, the rapidly expanding amount of information puts great pressure on information collection and information arrangement, and how to extract valuable information from massive texts by using natural language processing technology becomes a concern. The event information is composite information containing time, place, person, action and other contents, and has great value in the fields of news mining, financial analysis, community management and the like.
Existing event information processing processes are generally focused on event extraction but rarely focused on interrelationships between events (i.e., event relationships). This is because, when predicting the event relationship, it is necessary to make a fine annotation of the event relationship between event information in advance, which requires a lot of manpower resources, and further increases the cost of event information processing.
There is thus a need for improvements and improvements in the art.
Disclosure of Invention
The application aims to solve the technical problem of providing an event relation mining method and a related device aiming at the defects of the prior art.
In order to solve the above technical problem, a first aspect of an embodiment of the present application provides a method for mining an event relationship, where the method includes:
Respectively acquiring document characterization and event characterization of each target document in a plurality of target documents;
constructing an upper layer graph based on the document representation of each target document and constructing a lower layer graph based on the event representation of each target document, wherein the upper layer graph takes the document representation as a document node and takes the event relationship among the target documents as an edge, and the lower layer graph takes the event representation as an event node and adopts a random initialization mode to form an edge;
determining event relations among event characterizations of each target document based on a pre-trained graph mining model, the upper graph and the lower graph, wherein the graph mining model comprises an upper graph variation self-encoder and a lower graph variation self-encoder.
The mining method for the event relation is characterized in that the obtaining document representation and event representation of each target document in a plurality of target documents specifically comprises the following steps:
for each target document, acquiring a word embedding vector of an event trigger word in the target document through a pre-training model;
the word embedding vector of the event trigger word is used as the event representation of the target document;
and taking the average value of each event representation as the document representation of the target document.
The mining method of the event relation, wherein the event relation is marked among part of the target documents in the plurality of target documents.
The mining method of the event relation, wherein the constructing the lower layer graph based on the event characterization of each target document specifically comprises the following steps:
for each target document, taking event representation of the target document as an event node, and connecting the event node in a random initialization mode to form an edge so as to form a lower sub-graph;
and for any two target documents marking event relationships, selecting part of event nodes from lower sub-graphs corresponding to the two target documents in a random initialization mode, and connecting the selected part of event nodes to form a lower graph.
The mining method of the event relationship, wherein the determining the event relationship among the event characterizations of each target document based on the pre-trained graph mining model, the upper layer graph and the lower layer graph specifically comprises:
acquiring event characterization features and a lower-layer adjacency matrix of the lower-layer graph, and determining lower-layer coding features based on a lower-layer graph variation self-encoder, the event characterization features and the lower-layer adjacency matrix;
acquiring initial document characterization features and an upper layer adjacency matrix of the upper layer graph, and splicing the lower layer coding features and the initial document characterization features to obtain document characterization features;
Determining a predictive upper layer adjacency matrix based on the upper layer graph variation self-encoder, the document characterization feature, and the upper layer adjacency matrix;
updating the lower-layer adjacency matrix based on the predicted upper-layer adjacency matrix, and determining the event relation among the event characterizations of each target document based on the updated lower-layer adjacency matrix, the event characterizations and the lower-layer graph variation self-encoder.
The mining method of the event relation, wherein updating the lower layer adjacency matrix based on the predicted upper layer adjacency matrix specifically comprises:
determining event relations among all target documents according to the prediction upper-layer adjacency matrix;
for two target documents with event relations, selecting part of event tokens by adopting a random initialization mode to connect the event tokens corresponding to the two target documents into edges so as to update a lower layer diagram;
and determining an updated lower-layer adjacency matrix according to the updated lower-layer graph.
The mining method of the event relationship, wherein the method further comprises the following steps:
acquiring the number of characterization pairs of the event characterization pairs with the event relationship between every two target documents;
and when the number of the characterization pairs is greater than or equal to a preset number threshold, judging that a time relationship exists between the two target documents.
The mining method of the event relationship, wherein the determining process of the loss function in the training process of the graph mining model is as follows:
determining a lower layer prediction map based on a lower layer prediction adjacency matrix output from an encoder by a lower layer map variation, mapping the lower layer prediction map to an upper layer map, and determining a first loss function based on the upper layer map and the upper layer map;
determining an upper layer prediction map based on an upper layer prediction adjacency matrix output from an encoder by an upper layer map variation, and determining a second loss function based on the upper layer prediction map and the upper layer map;
and determining a loss function corresponding to the graph mining model based on the first loss function and the second loss function.
The method for mining the event relationship, wherein the mapping the lower layer prediction graph into the upper layer mapping graph specifically includes:
dividing the lower-layer predictive graph into lower-layer sub-predictive graphs according to a target document in which an event node in the lower-layer predictive graph is located;
and for each two lower-layer sub-prediction graphs, acquiring the node pair number of event node pairs with edges in the two lower-layer sub-prediction graphs, and setting edges for document nodes of the two lower-layer sub-prediction graphs with the node pair number larger than a set number threshold value to obtain an upper-layer mapping graph.
A second aspect of an embodiment of the present application provides an event relation mining system, the system including:
the acquisition module is used for respectively acquiring document characterization and event characterization of each target document in the plurality of target documents;
the construction module is used for constructing an upper layer diagram based on the document representation of each target document and constructing a lower layer diagram based on the event representation of each target document, wherein the upper layer diagram takes the document representation as a document node and the event relationship among the target documents as an edge, and the lower layer diagram takes the event representation as an event node and forms an edge in a random initialization mode;
and the mining module is used for determining event relations among event characterizations of each target document based on a pre-trained graph mining model, the upper layer graph and the lower layer graph, wherein the graph mining model comprises an upper layer graph variation self-encoder and a lower layer graph variation self-encoder.
A third aspect of the embodiments of the present application provides a computer-readable storage medium storing one or more programs executable by one or more processors to implement steps in a method of mining an event relationship as described in any of the above.
A fourth aspect of an embodiment of the present application provides a terminal device, including: a processor, a memory, and a communication bus, the memory having stored thereon a computer readable program executable by the processor;
the communication bus realizes connection communication between the processor and the memory;
the processor, when executing the computer readable program, implements the steps in the mining method of event relationships as described in any of the above.
The beneficial effects are that: compared with the prior art, the application provides a mining method and a related device of event relation, wherein the method comprises the steps of respectively obtaining document characterization and event characterization of each target document in a plurality of target documents; constructing an upper layer diagram based on the document representation of each target document, and constructing a lower layer diagram based on the event representation of each target document; and determining event relations among event characterizations of each target document based on a pre-trained graph mining model, the upper layer graph and the lower layer graph. According to the method, an upper layer diagram and a lower layer diagram are formed based on the document representation and the event representation of the target documents, then the event relationship is determined through the diagram mining model, so that the event relationship mining process is converted into a diagram completion task, the event relationship among the events can be mined only by marking the event relationship among the target documents, and manpower and material resources divided by the event relationship marking are reduced, so that the cost for mining the event relationship can be reduced.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for the description of the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present application, and other drawings may be obtained according to these drawings without creative effort for a person of ordinary skill in the art.
Fig. 1 is a flowchart of a method for mining event relationships provided by the present application.
Fig. 2 is a schematic diagram of an upper layer diagram and a lower layer diagram in the method for mining event relationships provided by the present application.
Fig. 3 is a schematic diagram of the working principle of the mining model.
Fig. 4 is a schematic structural diagram of an event relation mining system provided by the present application.
Fig. 5 is a schematic structural diagram of a terminal device provided by the present application.
Detailed Description
The application provides a method and a related device for mining event relations, which are used for making the purposes, technical schemes and effects of the application clearer and more definite, and the application is further described in detail below by referring to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the application.
As used herein, the singular forms "a", "an", "the" and "the" are intended to include the plural forms as well, unless expressly stated otherwise, as understood by those skilled in the art. It will be further understood that the terms "comprises" and/or "comprising," when used in this specification, specify the presence of stated features, integers, steps, and/or operations, but do not preclude the presence or addition of one or more other features, integers, steps, operations, and/or groups thereof. It should be understood that "and/or" includes all or any element and all combination of one or more associated listed items.
It will be understood by those skilled in the art that all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs unless defined otherwise. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the prior art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.
It should be understood that the sequence number and the size of each step in this embodiment do not mean the sequence of execution, and the execution sequence of each process is determined by the function and the internal logic of each process, and should not be construed as limiting the implementation process of the embodiment of the present application.
Through research, in the Internet age, the rapid expansion of information volume brings great pressure to information collection and information arrangement, and how to extract valuable information from massive texts by using natural language processing technology becomes a concern of people. The event information is composite information containing time, place, person, action and other contents, and has great value in the fields of news mining, financial analysis, community management and the like.
Existing event information processing processes are generally focused on event extraction but rarely focused on interrelationships between events (i.e., event relationships). At present, only co-reference resolution exists in the event relationship concerned in the event information processing process, wherein the co-reference resolution refers to the relationship pointing to the same specific event among the identified events, and redundant event extraction can be reduced by paying attention to the co-reference resolution relationship. The common event common-finger model comprises a pipeline model and a joint model, wherein the pipeline model needs to detect event trigger words in advance, and then common-finger event pairs are identified according to the trigger words, so that the problem of error propagation of overall errors caused by error identification of the trigger words exists; the combined model extracts events and completes coreference resolution in the same model, so that the problem of error propagation is avoided, but the complexity of the model is increased, and the universality is reduced.
In addition, there are more relationships among the events besides coreference resolution, such as time sequence, causality, approximation, etc., but at present, no machine learning algorithm capable of exploring the relationships exists, because when the event relationship prediction is performed, the event relationship among the event information needs to be subtly marked in advance, and more complex event relationship brings more expensive marking cost, so that the cost of event information processing is greatly increased.
In order to solve the above problems, in the embodiment of the present application, document representations and event representations of each of a plurality of target documents are acquired respectively; constructing an upper layer diagram based on the document representation of each target document, and constructing a lower layer diagram based on the event representation of each target document; and determining event relations among event characterizations of each target document based on a pre-trained graph mining model, the upper layer graph and the lower layer graph. According to the method, an upper layer diagram and a lower layer diagram are formed based on the document representation and the event representation of the target documents, then the event relationship is determined through the diagram mining model, so that the event relationship mining process is converted into a diagram completion task, the event relationship among the events can be mined only by marking the event relationship among the target documents, and manpower and material resources divided by the event relationship marking are reduced, so that the cost for mining the event relationship can be reduced.
The application will be further described by the description of embodiments with reference to the accompanying drawings.
The embodiment provides a method for mining event relationships, as shown in fig. 1, the method includes:
s10, respectively acquiring document characterization and event characterization of each target document in a plurality of target documents.
Specifically, the target documents may be news manuscripts, financial reports, social files, and the like, wherein the target documents may be chinese documents or foreign documents, but the language types of the target documents are the same, for example, the target documents are chinese documents, or the target documents are english documents. In one particular implementation, each of the plurality of target documents is an English document.
The document is characterized by being a text embedded vector of the target document, the event is characterized by being a word embedded vector of event trigger words in the target document, and the number of the event tokens is the same as the number of the event trigger words in the target document, namely, when one event trigger word exists in the target document, the target document corresponds to one event token, and when a plurality of event trigger words exist in the target document, the target document corresponds to a plurality of event tokens. In addition, the document characterization is determined based on the event characterization of the target document, for example, the document characterization is obtained by taking the average of the event characterization of the target document, or may be obtained by weighting the event characterization of the target document, or may be the event characterization with the largest occurrence number in the event characterization of the target document, or the like. Furthermore, it should be noted that the language type of the trigger word is the same as that of the target document, for example, when the target document is a chinese document, the trigger word is a chinese word, and when the target document is an english document, the trigger word is an english word.
In one implementation manner, the obtaining the document representation and the event representation of each target document in the plurality of target documents specifically includes:
s11, for each target document, acquiring a word embedding vector of an event trigger word in the target document through a pre-training model;
s12, using a word embedding vector of an event trigger word as an event representation of the target document;
s13, taking the mean value of each event representation as the document representation of the target document.
Specifically, the pre-training model is used for acquiring a word representation of each word in the target document (i.e. a word embedding vector of each word), and then extracting the word representation of the event trigger word as an event representation, wherein the pre-training model can adopt an existing natural voice processing model, such as a BERT model and the like. For example, the target document is "A meets B" at the corner of the park, the word representation of each word in "A meets B" at the corner of the park is obtained through the pre-training model, and then the event trigger word "meet" is extracted as the event representation.
After the event characterization of the target document (namely, the word embedding vector of each event trigger word in the target document) is obtained, all the obtained event characterizations are obtained and averaged to obtain the text characterizations. The target document can be correspondingly provided with one event token or at least two event tokens, and when the target document corresponds to one event token, the event token is used as the document token; when the target document corresponds to a plurality of event characterizations, taking the mean value of the plurality of event characterizations as the document characterization. For example, the event tokens of the target document include event token 1, event token 2, and event token 3, and then the mean of event token 1, event token 2, and event token 3 is taken as the document token.
S20, constructing an upper layer diagram based on the document representation of each target document, and constructing a lower layer diagram based on the event representation of each target document.
Specifically, the document representation is used as a document node in the upper layer diagram, the event relationship among all the target documents is used as an edge, that is, the document node in the upper layer diagram is used as the document representation of the target document, and the edge is the event relationship among the target documents. The event relationship among the target documents is marked in advance, and the event relationship can be a coreference resolution relationship, a time sequence relationship, a causal relationship or an approximate relationship and the like.
Further, some of the target documents may be marked with event relationships, or all of the target documents may be marked with event relationships, etc. In this embodiment, event relationships are marked between some target documents in the plurality of target documents, that is, whether there is an association tag between every two target documents, where the association tag is used to reflect that there is an event relationship between two documents, so that by taking the event relationship as an edge, the edge of the upper layer graph is accurate. For example, as shown in the upper layer diagram of fig. 2, an event relationship is marked between a target document a and a target document B, and an event relationship is marked between a target document C and a target document D, so that an edge is connected between a document node a corresponding to the target document a and a document node B corresponding to the target document B, and an edge is connected between a document node C corresponding to the target document C and a document node D corresponding to the target document D.
The lower layer graph takes event characterization as event nodes, connection edges among the event characterization determined in a random initialization mode are edges, in other words, the event nodes in the lower layer graph are event characterization, and the edges are formed in a random initialization mode. In addition, the target document can comprise a plurality of event representations, so that edges can be formed among the event representations corresponding to the target document, edges can be formed among the event representations of different target documents, and the forming of the edges among the event representations in a random initialization mode comprises forming the edges among the event representations corresponding to the target document by the random initialization mode, and forming the edges among the event representations of different target documents by the random initialization mode. According to the embodiment, the event characterization is used as the node of the lower layer graph, and the information loss of the mean value is avoided, so that the lower layer graph has more accurate standards.
Based on the above, the construction of the lower layer graph based on the event characterization of each target document specifically comprises:
s21, regarding each target document, taking event representation of the target document as an event node, and connecting the event nodes in a random initialization mode to form edges so as to form a lower sub-graph;
S22, selecting partial event nodes by adopting a random initialization mode for the target documents of any two marked event relations, and connecting the lower sub-graphs corresponding to the two target documents into edges to form a lower graph.
Specifically, the lower-layer subgraph is determined based on target documents, each target document corresponds to one lower-layer subgraph, wherein the lower-layer subgraph takes event representations of the corresponding target documents as nodes, takes a connecting edge formed by adopting a random initialization mode between the event representations corresponding to the target documents as edges, and can select part of event representations corresponding to the target documents to randomly form the connecting edge, or randomly form the connecting edge between all the event representations, wherein the random forming of the connecting edge refers to the random selection of two event representations, and forms the connecting edge between the two event representations. For example, as shown in FIG. 2, lower level sub-graph 1 includes event token x1, event token x2, and event token x3, with a connecting edge formed between event token x1 and event token x 2.
Further, after obtaining the lower sub-graph corresponding to each target document, obtaining each target document pair marking the event relationship, and then forming edges between the lower sub-graph pairs corresponding to each obtained target document pair, namely forming edges between event characterizations included in the lower sub-graph pairs, wherein the edges formed between the event characterizations included in the lower sub-graph pairs can be generated in a random initialization mode, namely randomly selecting part of the event characterizations in one lower sub-graph in the lower sub-graph pair, randomly selecting part of the event characterizations in the other lower sub-graph, and then forming edges between the randomly selected part of the event characterizations, for example, selecting a first event characterizations in the lower sub-graph and selecting a second event characterizations in the first lower sub-graph, and then forming edges between the first event characterizations and the second event characterizations.
In addition, in practical application, when generating edges for the lower layer graph, the edges can be generated for the lower layer graph by generating connecting edges for the same event characterization, besides adopting a random initialization mode. For example, as shown in fig. 2, the event tokens corresponding to the target document a include the event token of "promoting" formation and the event token of "promoting" formation, and the event token corresponding to the target document b also includes the event token of "developing" formation and the event token of "developing" formation, so that when determining the edge between the lower-layer sub-graph corresponding to the target document a and the lower-layer sub-graph corresponding to the target document b, an edge may be formed between the event token of "promoting" formation and the event token of "promoting" formation, and an edge may be formed between the event token of "developing" formation and the event token of "developing" formation.
S30, determining event relations among event characterizations of all target documents based on a pre-trained graph mining model, the upper layer graph and the lower layer graph, wherein the graph mining model comprises an upper layer graph variable self-encoder and a lower layer graph variable self-encoder.
Specifically, the graph mining model is trained in advance and is used for mining event relations among event characterizations of all target documents based on an upper layer graph and a lower layer graph, wherein the graph mining model comprises an upper layer graph variable self-encoder and a lower layer graph variable self-encoder, the upper layer graph variable self-encoder is used for carrying out information mining on the upper layer graph, the lower layer graph variable self-encoder is used for carrying out information mining on the lower layer graph, and in the mining process of the upper layer graph variable self-encoder and the lower layer graph variable self-encoder, the upper layer graph and the lower layer graph can be subjected to information interaction, so that the lower layer graph variable self-encoder can learn accurate information of edges in the upper layer graph, the upper layer graph variable self-encoder can learn accurate information of nodes in the lower layer graph, and the graph mining model can quickly acquire relations among the target documents (namely, a predicted upper layer adjacent matrix corresponding to the upper layer graph) and event characterizations (namely, a predicted lower layer adjacent matrix corresponding to the lower layer graph). Therefore, the event relationship of the event level can be obtained by marking the event relationship of the document level, so that the workload of marking the event relationship among event characterizations in the lower layer graph can be reduced, and the cost of marking the relationship among documents in large-scale data is reduced.
In one implementation, as shown in fig. 3, the determining, based on the pre-trained graph mining model, the upper layer graph and the lower layer graph, an event relationship between event characterizations of each target document specifically includes:
s31, acquiring event characterization features and a lower-layer adjacency matrix of the lower-layer graph, and determining lower-layer coding features based on a lower-layer graph variation self-encoder, the event characterization features and the lower-layer adjacency matrix;
s32, acquiring initial document characterization features and an upper layer adjacency matrix of the upper layer graph, and splicing the lower layer coding features and the initial document characterization features to obtain document characterization features;
s33, determining a prediction upper layer adjacency matrix based on the upper layer graph variation self-encoder, the document characterization feature and the upper layer adjacency matrix;
s34, updating the lower-layer adjacent matrix based on the predicted upper-layer adjacent matrix, and determining event relations among event characterizations of all target documents based on the updated lower-layer adjacent matrix, the event characterizations and the lower-layer graph variation self-encoder.
Specifically, the picture variation self-encoder comprises an encoder and a decoder, so that the upper layer picture variation self-encoder comprises an upper layer encoder and an upper layer decoder, and the lower layer picture variation self-encoder comprises a lower layer encoder and a lower layer decoder, wherein the upper layer picture variation self-encoder is used for determining a prediction upper layer adjacent matrix corresponding to the upper layer picture, and the lower layer picture variation self-encoder is used for determining a prediction lower layer adjacent matrix corresponding to the lower layer picture.
In step S31, the lower layer graph is transformed from the input items of the lower layer encoders in the encoders into a lower layer adjacency matrix and event characterization features determined based on the lower layer graph, wherein the lower layer adjacency matrix is used for reflecting feature information of edges in the lower layer graph, and the event characterization features are used for reflecting feature information of nodes in the lower layer graph. The lower layer encoder is configured to convolve the lower layer adjacency matrix and the event characterization feature to form a lower layer encoded feature, wherein the lower layer encoded feature carries information learned by the lower layer encoder from the lower layer adjacency matrix and the event characterization feature. The input of the lower layer decoder is a lower layer coding feature, and a predicted lower layer adjacency matrix is determined based on the lower layer coding feature, and the lower layer decoder reconstructs a lower layer graph based on the lower layer decoder learning information from the lower layer adjacency matrix and the event characterization feature to obtain the predicted lower layer adjacency matrix. In this way, the information is learned by the upper layer encoder, then the lower layer diagram is reconstructed by the lower layer decoder based on the learned information, the edge characteristics can be encoded into the node characteristics, and the similarity of the node characteristics with event relations is improved.
In step S32, the initial document characterization feature is determined based on the node characteristics in the upper layer graph, the upper layer adjacency matrix is determined based on the edge characteristics in the upper layer graph, and the document characterization feature is obtained by splicing the initial document characterization feature with the lower layer coding feature, wherein the document characterization feature carries the node characteristics of the upper layer graph and also carries interaction information learned from the edge characteristics of the lower layer graph and the node characteristics of the lower layer graph through the lower layer encoder. In this way, the document characterization features are used as input items of the upper layer encoder, so that the upper layer encoder can learn from the side features of the upper layer diagram, the node features of the upper layer diagram and the interaction information learned by the lower layer encoder from the side features of the lower layer diagram and the node features of the lower layer diagram, and the accuracy of the coding vector learned by the upper layer encoder is improved.
When the initial document characterization feature and the lower layer coding feature are spliced, the sequence of the initial document characterization feature-the lower layer coding feature or the sequence of the lower layer coding feature-the initial document characterization feature can be adopted for splicing. In the embodiment, the initial document characterization feature and the lower layer coding feature are spliced in the sequence of the initial document characterization feature and the lower layer coding feature, and in the process of multiple iterations, the initial document characterization feature is kept unchanged, and the lower layer coding feature is continuously updated to introduce new information for the document characterization feature, so that an upper layer encoder can learn the node characterization of the precision in the lower layer diagram, and the matching property of the document characterization feature input into the upper layer encoder and the initial document characterization can be ensured. For example, the initial document characterization feature is D dimension, the lower layer encoding feature is D dimension, then the document characterization feature is 2D dimension, wherein the front D dimension is the initial document characterization feature, the rear D is the lower layer encoding feature, and in the iterative process, the initial document characterization feature of the front D dimension remains unchanged, and the lower layer encoding feature of the rear D dimension is continuously updated.
In step S33, the upper layer graph variation self-encoder convolves the upper layer adjacency matrix with the document characterization feature through the upper layer encoder to learn the edge feature in the upper layer graph, the node feature in the upper layer graph and the interaction information transferred from the lower layer to the upper layer, and then the upper layer decoder reconstructs the upper layer graph based on the upper layer coding features learned by the upper layer encoder to obtain the upper layer prediction adjacency matrix.
In step S34, the self-encoder determines that the upper-layer adjacency matrix is predicted based on the upper-layer graph variation, and updates the lower-layer adjacency matrix, that is, transmits side information between document nodes in the upper-layer graph to the lower-layer graph, so that the lower-layer graph learns accurate side information of the upper-layer graph to enrich the information of the lower-layer graph.
In one implementation, the updating the lower layer adjacency matrix based on the predicted upper layer adjacency matrix specifically includes:
determining event relations among all target documents according to the prediction upper-layer adjacency matrix;
for two target documents with event relations, selecting part of event tokens by adopting a random initialization mode to connect the event tokens corresponding to the two target documents into edges so as to update a lower layer diagram;
and determining an updated lower-layer adjacency matrix according to the updated lower-layer graph.
Specifically, the prediction upper layer adjacency matrix carries edge characteristics, so that edges among document nodes in the upper layer graph can be predicted based on the prediction upper layer adjacency matrix, and event relations among target documents can be determined. Selecting part of event representations corresponding to two target documents by adopting a random initialization mode to form edges, namely forming edges between two lower sub-images corresponding to the two target documents by adopting a random initialization mode, wherein the forming of the edges between the two lower sub-images by adopting the random initialization mode is that part of event representations are selected from the two lower sub-images, and connecting edges are randomly formed between the selected part of event representations.
Further, in practical application, in the interaction process of the upper layer graph and the lower layer graph, other processes may be used for interaction, for example, firstly, determining a predicted upper layer adjacency matrix based on the document characterization feature and the upper layer adjacency matrix of the upper layer graph, wherein the document characterization feature comprises an initial document characterization feature and a preset lower layer coding feature (for example, each feature element of the preset lower layer coding feature is 0, etc.), secondly, updating the lower layer adjacency matrix of the lower layer graph based on the predicted upper layer adjacency matrix, determining the lower layer coding feature based on the updated lower layer adjacency matrix and the event characterization feature, and updating the preset lower layer coding feature in the document characterization feature by the lower layer coding feature; then, determining a predicted upper layer adjacency matrix again based on the updated document characterization features and the upper layer adjacency matrix; and finally updating the lower layer adjacent matrix based on the redetermined predicted upper layer adjacent matrix, and determining the lower layer predicted adjacent matrix based on the updated lower layer adjacent matrix and the event characterization features.
In one implementation, the training process of the graph mining model is the same as the use process of the graph mining model, and is a process of acquiring an upper layer graph and a lower layer graph and then interacting the upper layer graph and the lower layer graph. However, in the training process of the graph mining model, a loss function corresponding to the graph mining model needs to be determined, and the graph mining model is reversely learned based on the loss function. The determining process of the loss function in the training process of the graph mining model comprises the following steps:
Determining a lower layer prediction map based on a lower layer prediction adjacency matrix output from an encoder by a lower layer map variation, mapping the lower layer prediction map to an upper layer map, and determining a first loss function based on the upper layer map and the upper layer map;
determining an upper layer prediction map based on an upper layer prediction adjacency matrix output from an encoder by an upper layer map variation, and determining a second loss function based on the upper layer prediction map and the upper layer map;
and determining a loss function corresponding to the graph mining model based on the first loss function and the second loss function.
Specifically, the document nodes of the upper mapping graph are document representations of target documents, and the edges are event relationships among the target documents, wherein the edges in the upper mapping graph are determined based on the lower prediction graph, for example, when edges exist between event nodes between two lower subgraphs corresponding to two target documents, the edges can be set for two document nodes corresponding to the two target documents; or when edges exist between all event nodes between two lower-layer subgraphs corresponding to two target documents, the edges can be set for the two document nodes corresponding to the two target documents; alternatively, when an edge exists between part of event nodes between two lower-level subgraphs corresponding to two target documents, an edge may be set for two document nodes corresponding to two target documents.
In one implementation, the mapping the lower layer prediction map to the upper layer map specifically includes:
dividing the lower-layer predictive graph into lower-layer sub-predictive graphs according to a target document in which an event node in the lower-layer predictive graph is located;
and for each two lower-layer sub-prediction graphs, acquiring the node pair number of event node pairs with edges in the two lower-layer sub-prediction graphs, and setting edges for document nodes of the two lower-layer sub-prediction graphs with the node pair number larger than a set number threshold value to obtain an upper-layer mapping graph.
Specifically, the lower-layer sub-prediction graph corresponds to a target document, the event nodes in the lower-layer sub-prediction graph correspond to the event characterizations of the target document one by one, and each event in the target document is characterized as an event node in the lower-layer sub-prediction graph. That is, the number of lower-layer sub-predictive pictures divided by the lower-layer predictive pictures is the same as the number of the plurality of target documents, and each lower-layer sub-predictive picture corresponds to each target document one by one.
After the lower-layer sub-prediction graph is obtained by dividing, the number of node pairs of event node pairs with edges of every two lower-layer sub-prediction graphs is obtained, for example, the number of node pairs corresponding to the lower-layer prediction sub-graph a and the lower-layer prediction sub-graph b is 1 when an edge is arranged between the event node 1 in the lower-layer prediction sub-graph a and the event node 2 in the lower-layer prediction sub-graph b. Setting a number threshold value to be preset, wherein the number threshold value is a basis for setting edges between document nodes or not, and when the number of node pairs is larger than the number threshold value, the edges are set for the document nodes of the two lower-layer sub-predictive graphs; conversely, when the number of node pairs is less than or equal to the set number threshold, no edge is set for the document nodes of the two lower-layer subprediction graphs.
In one implementation, after the event relationships between the event characterizations of the target documents, the event relationships between the target documents may also be determined based on the event relationships between the event characterizations of the target documents. Based thereon, the method further comprises:
acquiring the number of characterization pairs of the event characterization pairs with the event relationship between every two target documents;
and when the number of the characterization pairs is greater than or equal to a preset number threshold, judging that a time relationship exists between the two target documents.
Specifically, the preset quantity threshold is preset and is the basis for judging whether an event relationship exists between the target documents, wherein when the quantity of the characterization pairs is larger than the preset quantity threshold, the event relationship exists between the two target documents is indicated; otherwise, when the number of the characterization pairs is smaller than or equal to a preset number threshold, the fact that no event relationship exists between the two target documents is indicated. Thus, the event relation among the event characterizations in the target document can be determined, and the event relation among the target documents can be determined.
In summary, the present embodiment provides a method for mining an event relationship, where the method includes respectively obtaining a document representation and an event representation of each target document in a plurality of target documents; constructing an upper layer diagram based on the document representation of each target document, and constructing a lower layer diagram based on the event representation of each target document; and determining event relations among event characterizations of each target document based on a pre-trained graph mining model, the upper layer graph and the lower layer graph. According to the method, an upper layer diagram and a lower layer diagram are formed based on the document representation and the event representation of the target documents, then the event relationship is determined through the diagram mining model, so that the event relationship mining process is converted into a diagram completion task, the event relationship among the events can be mined only by marking the event relationship among the target documents, and manpower and material resources divided by the event relationship marking are reduced, so that the cost for mining the event relationship can be reduced.
Based on the above-mentioned mining method of event relationships, this embodiment provides a mining system of event relationships, as shown in fig. 4, where the system includes:
an obtaining module 100, configured to obtain a document representation and an event representation of each target document in a plurality of target documents respectively;
the construction module 200 is configured to construct an upper layer graph based on document representations of each target document, and construct a lower layer graph based on event representations of each target document, wherein the upper layer graph uses the document representations as document nodes and event relationships among the target documents as edges, and the lower layer graph uses the event representations as event nodes and adopts a random initialization mode to form edges;
the mining module 300 is configured to determine an event relationship between event characterizations of each target document based on a pre-trained graph mining model, the upper layer graph and the lower layer graph, where the graph mining model includes an upper layer graph variation self-encoder and a lower layer graph variation self-encoder.
Based on the above-described event relation mining method, the present embodiment provides a computer-readable storage medium storing one or more programs executable by one or more processors to implement the steps in the event relation mining method as described in the above-described embodiment.
Based on the mining method of the event relationship, the application also provides a terminal device, as shown in fig. 5, which comprises at least one processor (processor) 20; a display screen 21; and a memory (memory) 22, which may also include a communication interface (Communications Interface) 23 and a bus 24. Wherein the processor 20, the display 21, the memory 22 and the communication interface 23 may communicate with each other via a bus 24. The display screen 21 is configured to display a user guidance interface preset in the initial setting mode. The communication interface 23 may transmit information. The processor 20 may invoke logic instructions in the memory 22 to perform the methods of the embodiments described above.
Further, the logic instructions in the memory 22 described above may be implemented in the form of software functional units and stored in a computer readable storage medium when sold or used as a stand alone product.
The memory 22, as a computer readable storage medium, may be configured to store a software program, a computer executable program, such as program instructions or modules corresponding to the methods in the embodiments of the present disclosure. The processor 20 performs functional applications and data processing, i.e. implements the methods of the embodiments described above, by running software programs, instructions or modules stored in the memory 22.
The memory 22 may include a storage program area that may store an operating system, at least one application program required for functions, and a storage data area; the storage data area may store data created according to the use of the terminal device, etc. In addition, the memory 22 may include high-speed random access memory, and may also include nonvolatile memory. For example, a plurality of media capable of storing program codes such as a usb disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (Random Access Memory, RAM), a magnetic disk, or an optical disk, or a transitory storage medium may be used.
In addition, the specific processes that the storage medium and the plurality of instruction processors in the terminal device load and execute are described in detail in the above method, and are not stated here.
Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present application, and are not limiting; although the application has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present application.
Claims (10)
1. A method of mining event relationships, the method comprising:
respectively acquiring document characterization and event characterization of each target document in a plurality of target documents;
constructing an upper layer graph based on the document representation of each target document and constructing a lower layer graph based on the event representation of each target document, wherein the upper layer graph takes the document representation as a document node and takes the event relationship among the target documents as an edge, and the lower layer graph takes the event representation as an event node and adopts a random initialization mode to form an edge;
determining event relations among event characterizations of each target document based on a pre-trained graph mining model, the upper graph and the lower graph, wherein the graph mining model comprises an upper graph variation self-encoder and a lower graph variation self-encoder;
the determining the event relation among the event characterizations of each target document based on the pre-trained graph mining model, the upper layer graph and the lower layer graph specifically comprises:
acquiring event characterization features and a lower adjacency matrix of the lower layer graph, and determining lower layer coding features based on a lower layer graph variation self-encoder, the event characterization features and the lower layer adjacency matrix, wherein the lower layer adjacency matrix is used for reflecting feature information of edges in the lower layer graph, and the event characterization features are used for reflecting feature information of nodes in the lower layer graph;
Acquiring initial document characterization features and an upper layer adjacency matrix of the upper layer graph, and splicing the lower layer coding features with the initial document characterization features to obtain document characterization features, wherein the initial document characterization features are determined based on node features in the upper layer graph, and the upper layer adjacency matrix is determined based on edge features in the upper layer graph;
determining a predictive upper layer adjacency matrix based on the upper layer graph variation self-encoder, the document characterization feature, and the upper layer adjacency matrix;
updating the lower-layer adjacency matrix based on the predicted upper-layer adjacency matrix, and determining event relations among event characterizations of all target documents based on the updated lower-layer adjacency matrix, the event characterizations and the lower-layer graph variation self-encoder;
the updating the lower layer adjacency matrix based on the predictive upper layer adjacency matrix specifically comprises:
determining event relations among all target documents according to the prediction upper-layer adjacency matrix;
for two target documents with event relations, selecting part of event tokens by adopting a random initialization mode to connect the event tokens corresponding to the two target documents into edges so as to update a lower layer diagram;
And determining an updated lower-layer adjacency matrix according to the updated lower-layer graph.
2. The method for mining event relationships according to claim 1, wherein the obtaining the document representation and the event representation of each target document of the plurality of target documents specifically includes:
for each target document, acquiring a word embedding vector of an event trigger word in the target document through a pre-training model;
the word embedding vector of the event trigger word is used as the event representation of the target document;
and taking the average value of each event representation as the document representation of the target document.
3. The method of mining event relationships according to claim 1, wherein event relationships are marked between some of the target documents.
4. A method of mining event relationships according to claim 1 or 3, wherein the constructing a lower layer graph based on event characterization of each target document specifically comprises:
for each target document, taking event representation of the target document as an event node, and connecting the event node in a random initialization mode to form an edge so as to form a lower sub-graph;
and for any two target documents marking event relationships, selecting part of event nodes from lower sub-graphs corresponding to the two target documents in a random initialization mode, and connecting the selected part of event nodes to form a lower graph.
5. The method of mining event relationships according to claim 1, further comprising:
acquiring the number of characterization pairs of the event characterization pairs with the event relationship between every two target documents;
and when the number of the characterization pairs is greater than or equal to a preset number threshold, determining that an event relationship exists between the two target documents.
6. The mining method of event relationships according to claim 1, wherein the determining of the loss function in the training of the graph mining model is:
determining a lower layer prediction map based on a lower layer prediction adjacency matrix output from an encoder by a lower layer map variation, mapping the lower layer prediction map to an upper layer map, and determining a first loss function based on the upper layer map and the upper layer map;
determining an upper layer prediction map based on an upper layer prediction adjacency matrix output from an encoder by an upper layer map variation, and determining a second loss function based on the upper layer prediction map and the upper layer map;
and determining a loss function corresponding to the graph mining model based on the first loss function and the second loss function.
7. The method of mining event relationships according to claim 6, wherein the mapping the lower-layer prediction graph to an upper-layer mapping graph specifically comprises:
Dividing the lower-layer predictive graph into lower-layer sub-predictive graphs according to a target document in which an event node in the lower-layer predictive graph is located;
and for each two lower-layer sub-prediction graphs, acquiring the node pair number of event node pairs with edges in the two lower-layer sub-prediction graphs, and setting edges for document nodes of the two lower-layer sub-prediction graphs with the node pair number larger than a set number threshold value to obtain an upper-layer mapping graph.
8. A mining system for event relationships, the system comprising:
the acquisition module is used for respectively acquiring document characterization and event characterization of each target document in the plurality of target documents;
the construction module is used for constructing an upper layer diagram based on the document representation of each target document and constructing a lower layer diagram based on the event representation of each target document, wherein the upper layer diagram takes the document representation as a document node and the event relationship among the target documents as an edge, and the lower layer diagram takes the event representation as an event node and forms an edge in a random initialization mode;
the mining module is used for determining event relations among event characterizations of all target documents based on a pre-trained graph mining model, the upper layer graph and the lower layer graph, wherein the graph mining model comprises an upper layer graph variation self-encoder and a lower layer graph variation self-encoder;
The determining the event relation among the event characterizations of each target document based on the pre-trained graph mining model, the upper layer graph and the lower layer graph specifically comprises:
acquiring event characterization features and a lower adjacency matrix of the lower layer graph, and determining lower layer coding features based on a lower layer graph variation self-encoder, the event characterization features and the lower layer adjacency matrix, wherein the lower layer adjacency matrix is used for reflecting feature information of edges in the lower layer graph, and the event characterization features are used for reflecting feature information of nodes in the lower layer graph;
acquiring initial document characterization features and an upper layer adjacency matrix of the upper layer graph, and splicing the lower layer coding features with the initial document characterization features to obtain document characterization features, wherein the initial document characterization features are determined based on node features in the upper layer graph, and the upper layer adjacency matrix is determined based on edge features in the upper layer graph;
determining a predictive upper layer adjacency matrix based on the upper layer graph variation self-encoder, the document characterization feature, and the upper layer adjacency matrix;
updating the lower-layer adjacency matrix based on the predicted upper-layer adjacency matrix, and determining event relations among event characterizations of all target documents based on the updated lower-layer adjacency matrix, the event characterizations and the lower-layer graph variation self-encoder;
The updating the lower layer adjacency matrix based on the predictive upper layer adjacency matrix specifically comprises:
determining event relations among all target documents according to the prediction upper-layer adjacency matrix;
for two target documents with event relations, selecting part of event tokens by adopting a random initialization mode to connect the event tokens corresponding to the two target documents into edges so as to update a lower layer diagram;
and determining an updated lower-layer adjacency matrix according to the updated lower-layer graph.
9. A computer readable storage medium storing one or more programs executable by one or more processors to implement the steps in the method of mining event relationships of any of claims 1-7.
10. A terminal device, comprising: a processor, a memory, and a communication bus; the memory has stored thereon a computer readable program executable by the processor;
the communication bus realizes connection communication between the processor and the memory;
the processor, when executing the computer readable program, implements the steps in the method of mining event relationships as claimed in any one of claims 1 to 7.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310309890.3A CN116069831B (en) | 2023-03-28 | 2023-03-28 | Event relation mining method and related device |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310309890.3A CN116069831B (en) | 2023-03-28 | 2023-03-28 | Event relation mining method and related device |
Publications (2)
Publication Number | Publication Date |
---|---|
CN116069831A CN116069831A (en) | 2023-05-05 |
CN116069831B true CN116069831B (en) | 2023-08-15 |
Family
ID=86170070
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202310309890.3A Active CN116069831B (en) | 2023-03-28 | 2023-03-28 | Event relation mining method and related device |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN116069831B (en) |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10839013B1 (en) * | 2018-05-10 | 2020-11-17 | Facebook, Inc. | Generating a graphical representation of relationships among a set of articles and information associated with the set of articles |
CN113032585A (en) * | 2021-05-31 | 2021-06-25 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Document-level entity relation extraction method based on document structure and external knowledge |
CN113590824A (en) * | 2021-07-30 | 2021-11-02 | 平安科技(深圳)有限公司 | Method and device for constructing causal graph and related equipment |
CN115510184A (en) * | 2022-08-25 | 2022-12-23 | 昆明理工大学 | News core event detection method integrating document graph and event graph |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10303999B2 (en) * | 2011-02-22 | 2019-05-28 | Refinitiv Us Organization Llc | Machine learning-based relationship association and related discovery and search engines |
-
2023
- 2023-03-28 CN CN202310309890.3A patent/CN116069831B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10839013B1 (en) * | 2018-05-10 | 2020-11-17 | Facebook, Inc. | Generating a graphical representation of relationships among a set of articles and information associated with the set of articles |
CN113032585A (en) * | 2021-05-31 | 2021-06-25 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Document-level entity relation extraction method based on document structure and external knowledge |
CN113590824A (en) * | 2021-07-30 | 2021-11-02 | 平安科技(深圳)有限公司 | Method and device for constructing causal graph and related equipment |
CN115510184A (en) * | 2022-08-25 | 2022-12-23 | 昆明理工大学 | News core event detection method integrating document graph and event graph |
Non-Patent Citations (1)
Title |
---|
Cross-document Misinformation Detection based on Event Graph Reasoning;Xueqing Wu etc.;Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies;第543-558页 * |
Also Published As
Publication number | Publication date |
---|---|
CN116069831A (en) | 2023-05-05 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109960810B (en) | Entity alignment method and device | |
CN112084383A (en) | Information recommendation method, device and equipment based on knowledge graph and storage medium | |
US20220100963A1 (en) | Event extraction from documents with co-reference | |
US20220100772A1 (en) | Context-sensitive linking of entities to private databases | |
CN108664512B (en) | Text object classification method and device | |
US20220100967A1 (en) | Lifecycle management for customized natural language processing | |
US20180365594A1 (en) | Systems and methods for generative learning | |
CN112949476B (en) | Text relation detection method, device and storage medium based on graph convolution neural network | |
CN113779225B (en) | Training method of entity link model, entity link method and device | |
CN115438215A (en) | Image-text bidirectional search and matching model training method, device, equipment and medium | |
CN116152833B (en) | Training method of form restoration model based on image and form restoration method | |
CN113704420A (en) | Method and device for identifying role in text, electronic equipment and storage medium | |
CN114612921A (en) | Form recognition method and device, electronic equipment and computer readable medium | |
CN116304341A (en) | Fraud discrimination method and system based on user network big data | |
JP2023062150A (en) | Character recognition model training, character recognition method, apparatus, equipment, and medium | |
CN114548296A (en) | Graph convolution recommendation method based on self-adaptive framework and related device | |
CN112950291B (en) | Model deviation optimization method, device, equipment and computer readable medium | |
EP4222635A1 (en) | Lifecycle management for customized natural language processing | |
US20220180044A1 (en) | Automatic delineation and extraction of tabular data in portable document format using graph neural networks | |
CN113656797A (en) | Behavior feature extraction method and behavior feature extraction device | |
CN117423124A (en) | Table data processing method, device, equipment and medium based on table image | |
CN116069831B (en) | Event relation mining method and related device | |
CN111967253A (en) | Entity disambiguation method and device, computer equipment and storage medium | |
CN116756281A (en) | Knowledge question-answering method, device, equipment and medium | |
CN115186738B (en) | Model training method, device and storage medium |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |