A method for generating multi-dimensional event profiles across chapters

Abstract

This invention discloses a method for generating multi-dimensional event profiles across multiple texts. For a given event type, it searches for relevant texts describing that type of event and segments each text into sentence blocks, identifying text blocks such as basic information, event sequence, causes and effects, and comments from various parties. Next, it extracts events from the text blocks describing basic information, obtaining basic elements such as event type, occurrence / end time, location, and actors. Then, it identifies sub-events from the text blocks describing the event sequence and sorts them chronologically to form an event timeline. Finally, it performs coreference resolution on events from different texts to form a complete event profile. This method can correlate and fuse event information distributed across multiple texts, extract complex elements such as causes and effects, and comments from various parties, and can discover sub-events such as the early developments, main processes, and subsequent actions of an event, enabling the analysis of various elements and the evolution of major events.

Description

Technical Field

[0001] This invention relates to a method for generating multi-dimensional event profiles, and more particularly to a method for generating multi-dimensional event profiles across chapters. Background Technology

[0002] The goal of generating multi-dimensional event profiles is to extract information from various dimensions of major events from text data of various sources, thereby supporting a comprehensive understanding. Its main dimensions include basic information, event timeline, causes and impacts, comments from various parties, and similar events, such as... Figure 2 As shown.

[0003] Without the ability to generate multi-dimensional profiles of events, users would need to discover the events they need to pay attention to from massive amounts of existing data and constantly changing real-time data, organize, understand, extract, and correlate textual data around the events, and sort out their development process. This process would take a lot of time.

[0004] Multi-dimensional event profiling primarily helps them solve three core problems: first, automatically identifying major events of interest from constantly changing text data; second, extracting the basic elements of an event and important sub-events in its development process, thereby helping users gain a comprehensive understanding of all aspects of the event; and third, associating and organizing various types of text for each event, providing a foundation for further reading, exploration, and data mining analysis.

[0005] The generation of multidimensional event profiles cannot be solved simply by using event extraction techniques. The main difficulties and problems it faces include: (1) Elements such as causes, effects, and comments from various parties are no longer simple entities, but rather sentences or even paragraphs of varying lengths. How to extract these complex elements with varying lengths? (2) The event timeline includes sub-events such as the main process of the event, the initial trends, and subsequent actions. How to discover these sub-events and sort out their sequential relationships? (3) The same event may be described by multiple documents. How to associate and integrate events distributed across multiple documents? Summary of the Invention

[0006] Purpose of the invention: The technical problem to be solved by the present invention is to provide a method for generating multi-dimensional event profiles across chapters, addressing the shortcomings of the existing technology.

[0007] To address the aforementioned technical problems, this invention discloses a method for generating multi-dimensional event profiles across chapters, comprising four stages: text preprocessing, extraction of basic event information, generation of event context, and cross-chapter event fusion.

[0008] Step 1, the text preprocessing stage, includes: accessing text data, classifying the text data for relevance to obtain relevant text describing the type of event of interest, and performing text structure analysis and fine-grained sentence segmentation on the relevant text to obtain text blocks describing different dimensions of information about the event, including: text blocks describing basic information, text blocks describing the course of the event, text blocks describing the causes and effects, text blocks describing comments from various parties, and text blocks describing similar events;

[0009] Step 2, the basic event information extraction stage, includes: extracting events from text blocks describing basic information to obtain the basic elements of the event, including the type of event, the time of occurrence or end, the location, and the subject of the action;

[0010] Step 3, the event context generation stage, includes: identifying sub-events from text blocks describing the course of events, and sorting them in chronological order to form an event context;

[0011] Step 4, the event cross-text fusion stage, includes: performing coreference discovery on events extracted from different text blocks to form a complete event profile.

[0012] Beneficial effects:

[0013] This invention utilizes techniques such as sequence statement annotation, joint extraction of event trigger words and arguments, event sequence relationship identification, and event coreference discovery to achieve the generation of multi-dimensional event profiles across texts. Compared with existing sentence-level and text-level event extraction techniques, its significant advantages are: (1) It can extract event elements with numerous data points and varying lengths. (2) It can organize the development trajectory of events. (3) It can perform cross-text fusion of events distributed across multiple texts. Attached Figure Description

[0014] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0015] Figure 1 This is a flowchart of the processing of the present invention.

[0016] Figure 2 This is a schematic diagram of a multi-dimensional portrait system.

[0017] Figure 3 This is a technical architecture diagram of the present invention.

[0018] Figure 4 This is a network structure diagram of the statement sequence classification model described in this invention.

[0019] Figure 5 This is a network structure diagram of the event element extraction model described in this invention. Detailed Implementation

[0020] like Figure 1 As shown, a method for generating multi-dimensional event profiles across chapters includes four stages: text preprocessing, extraction of basic event information, generation of event context, and fusion of events across chapters.

[0021] Step 1, the text preprocessing stage, includes: accessing text data, classifying the text data for relevance to obtain relevant text describing the type of event of interest, and performing text structure analysis and fine-grained sentence segmentation on the relevant text to obtain text blocks describing different dimensions of information about the event, including: text blocks describing basic information, text blocks describing the course of the event, text blocks describing the causes and effects, text blocks describing comments from various parties, and text blocks describing similar events;

[0022] The fine-grained sentence segmentation method is as follows: sentence segmentation is modeled as a sequence sentence classification problem, and the BERT model is used to perform context-dependent representation and classification of sentences in the relevant text. Specific methods include:

[0023] Step 1-1, Text Preprocessing: Given a series of statements S = <S1,…,S s >, where s represents the number of statements in a series of statements, which is then divided into multiple text segments of no more than 10 sentences. For each text segment... m ,…,S k > Execute steps 1-2 and 1-3 to obtain the sequence statement classification result of the text fragment; where S1 represents the first statement in a series of statements S, S n S represents the nth statement in a series of statements S. m This represents the m-th statement in a series of statements S. k This represents the k-th statement in a series of statements S;

[0024] Step 1-2, Input representation: For text fragments m ,…,S k >, append the [SEP] separator to the end of each statement to form a sequence containing multiple statements, and finally insert [CLS] at the beginning of the entire sequence to obtain the input representation;

[0025] Steps 1-3, Sentence Classification: Input the input representation described in Steps 1-2 into the BERT model to obtain the [SEP] encoding of each sentence, and input the result into a multilayer feedforward neural network (MLP) to output the sentence classification result;

[0026] Steps 1-4, Results Summary: Concatenate the classification results obtained from all text fragments after steps 1-3 in order to obtain the final classification result. ​​

[0027] Step 2, the basic event information extraction stage, includes: extracting events from text blocks describing basic information to obtain the basic elements of the event, including the type of event, the time of occurrence or end, the location, and the subject of the action;

[0028] The event extraction method employs a neural network structure containing a BERT-BiLSTM-CRF encoding layer, an attention layer, and an argument mapping layer to identify event trigger words and event types, and extract each argument and its role, specifically including:

[0029] BERT-BiLSTM-CRF encoding layer: The network structure of pre-trained language model BERT plus bidirectional LSTM plus conditional random field CRF is adopted. The identification of trigger words and event types is treated as a sequence labeling problem, and the BIO labeling scheme is used for identification.

[0030] Attention layer: Analyzes the correlation between each word in the text block describing basic information and the trigger word, and concatenates it with the output of the Bi-LSTM encoding layer to capture the semantic correlation of each word with respect to the trigger word;

[0031] Argument mapping layer: A classification model is built for each argument role to identify whether each word in the text block describing basic information is the start or end position of that argument.

[0032] The BERT-BiLSTM-CRF encoding layer includes: a pre-trained language model BERT, a Bi-LSTM encoding layer, and a CRF layer;

[0033] The pre-trained language model BERT is used to semantically encode the words in the input sentence; given sentence S i ={w1,w2,w3,…,w n-1 ,w n}, where w n Statement S i The nth word in the sequence is encoded as x = {x1, x2, x3, ..., xn}. n-1 ,x n}, where x n Indicator w n The encoding;

[0034] A Bi-LSTM encoding layer is used to encode global information for the statement, learning contextual information for each word; the statement x = {x1, x2, x3, ..., x n-1 ,x n After forward LSTM and backward LSTM encoding, the information outputs are respectively and in, x represents n The vector after forward LSTM encoding x represents n The vector after backward LSTM encoding is concatenated with the forward and backward encoding results to obtain the concatenated result. After passing through the Bi-LSTM encoding layer, the encoded result of the statement is represented as h = {h1, h2, h3, ..., h...} n-1 ,h n}, where h n This represents the vector after the nth character is encoded using Bi-LSTM.

[0035] The CRF layer identifies trigger words and their types in a statement, which are then used as input for subsequent argument recognition. The identification of trigger words and their types is treated as a sequence labeling model, employing a BX, IX, O labeling scheme. In BX, B represents the start position of the trigger word, X represents the event type of the trigger word, I represents the middle or end position of the trigger word in IX, and O represents a non-trigger word result. Each word is labeled accordingly, thus achieving the identification of trigger words and their types, and outputting a trigger word set T = {t1, t2, t3, ..., t...}. m}, where t m This indicates the m-th trigger word.

[0036] The attention layer comprises: based on the trigger word set T, using an attention mechanism, analyzing the correlation between each word in the sentence and the trigger word, and concatenating it with the output of the Bi-LSTM encoding layer as subsequent input; the specific calculation process is as follows:

[0037] For the i-th trigger word t i Perform the following operations

[0038] q i =sum([h start :h end ])

[0039] Where [:] represents vector slicing, sum() represents vector addition, and start represents the trigger word t. i The beginning position, end indicates the trigger word t i The end position, q i Indicates trigger word t i In h, for the set of trigger words T, we obtain the set of trigger word representations Q = {q1,q2,q3,…,q}. m};

[0040] The representation q of the i-th trigger word i Using the attention mechanism, the following calculation is performed to obtain the correlation a between the j-th word and the i-th trigger word.ij :

[0041]

[0042] Then calculate the attention representation of the j-th word in the statement relative to the i-th trigger word: h j *a ij And concatenate it with the original representation:

[0043] h′ ij =[h j ,h j *a ij ]

[0044] Finally, obtain the hidden layer representation h′ of each word in the statement relative to the i-th trigger word. i ={h′ i1 ,h′ i2 ,h′ i3 ,…,h′ in}

[0045] The argument mapping layer includes: for the i-th trigger word t i Each argument role is modeled into a classification model to identify whether each word is the start or end position of that argument. The model is then fed into a fully connected layer. i Mapped to binary classification:

[0046] O head =sigmoid(h′) i W head +b head )

[0047] O tail =sigmoid(h′) i W tail +b tail )

[0048] Among them, W head and b head For the weight parameters to predict the starting position, W rail and b tail For the weight parameters of the predicted end position, O head This represents the probability vector for predicting the starting position of each argument for each word, O tail This represents the probability vector that each word is predicted to be the end position of an argument.

[0049] Step 3, the event context generation stage, includes: identifying sub-events from text blocks describing the course of events, and sorting them in chronological order to form an event context;

[0050] The method for determining the chronological order is as follows: when there is no explicit time description in the sub-events, the chronological order is determined by comprehensively using time sequence connectors, event type pairs, and the transitivity of time sequence relationships. Specifically, this includes:

[0051] Temporal Connectors: Construct a mapping table between temporal connectors and temporal relationships, and constrain event pairs that have temporal relationships;

[0052] Event type pairs: Construct a list of event type pairs and constrain the temporal relationship between event pairs that conform to a specific type;

[0053] Transitivity of temporal relationships: The temporal relationships of events have the property of transitive closure. If relation(e0,e1) and relation(e1,e2) are known, then relation(e0,e2) is derived. Here, relation(e0,e1) indicates that events e0 and e1 have a temporal relationship, relation(e1,e2) indicates that events e1 and e2 have a temporal relationship, and relation(e0,e2) indicates that events e0 and e2 have a temporal relationship.

[0054] Step 4, the event cross-text fusion stage, includes: performing coreference discovery on events extracted from different text blocks to form a complete event profile.

[0055] The time correlation Rel T The calculation method for (e1, e2) is as follows:

[0056]

[0057] Where time(e1) represents the time of event e1, time(e2) represents the time of event e2, and window T It is a preset time window constant;

[0058] The location relevance Rel P The specific calculation formula for (e1, e2) is as follows:

[0059]

[0060] Where P(e1) represents the location region of event e1, P(e2) represents the location region of event e2, dis() is the distance between the center points of the regions, and window P It is a distance constant;

[0061] The participant relevance Rel A The specific calculation formula for (e1, e2) is as follows:

[0062]

[0063] Where actor(e1) represents the set of participants in event e1, and actor(e2) represents the set of participants in event e2;

[0064] The type correlation is measured by the distance between two events e1 and e2 in the classification system. The strength of the association is strongest when they are of the same type, denoted as Rel. C (e1,e2).

[0065] The overall relevance is obtained by weighting and summing the relevance of events in terms of time, location, participants, and type, specifically as follows:

[0066] Rel(e1,e2)=αRel T (e1,e2)+βRel P (e1,e2)+γRel A (e1,e2)+δRel C (e1,e2).

[0067] Example:

[0068] The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0069] This embodiment provides a method for generating multi-dimensional profiles of events across chapters. This method can be applied to obtain multi-dimensional profiles of major events from text data from various sources, thereby supporting a comprehensive understanding of the causes, development process, impact, and external reactions of events.

[0070] This application provides a method for generating multi-dimensional event profiles across chapters, as shown in the embodiments below. Figure 3 As shown, it includes four stages: text preprocessing, extraction of basic event information, generation of event context, and cross-chapter fusion of events;

[0071] The text preprocessing stage is used to access text data, classify the text data for relevance, obtain relevant text describing the type of event of interest, and perform text structure analysis and fine-grained sentence segmentation on the relevant text to obtain text blocks describing different dimensions of information about the event.

[0072] The event basic information extraction stage is used to extract events from text blocks describing basic information, and obtain basic elements such as event type, occurrence / end time, location, and subject of action;

[0073] The event timeline generation stage is used to identify sub-events from text blocks describing the course of events and sort them in chronological order to form an event timeline.

[0074] The event cross-text fusion stage is used to discover common references to events in different texts and form a complete event profile.

[0075] In this embodiment, the fine-grained sentence segmentation method in the text preprocessing stage, such as... Figure 4 As shown, sentence segmentation is modeled as a sequence sentence classification problem. A special input representation is proposed, which can use the BERT model (reference: 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of NAACL-HLT 2019, pages 4171–4186) to perform context-dependent representation and classification of sentences, specifically including:

[0076] Step 1-1 (Long Text Processing): Because BERT only supports 512 characters of input, limiting the model's ability to process large documents, document segmentation is performed. Considering that the meaning of a sentence usually depends more on the one or two sentences before and after it than on other, more distant sentences, direct segmentation, except for edge sentences which may be slightly affected, will not significantly impact the completeness of the meaning of other sentences. The document is segmented according to the number of sentences (e.g., 10 sentences per input), ensuring the total number of tokens does not exceed 512. Then, each segmented text fragment is input into Step 2-2 for sequence classification.

[0077] Step 1-2 (Input Representation): A special input representation is proposed, given a series of statements S = <S1,…,S s >, append the [SEP] delimiter to the first statement, and repeat this process for each statement to form a sequence containing multiple statements. Finally, insert [CLS] at the beginning of the entire sequence.

[0078] Steps 1-3 (Sentence Classification): Input the above representation into the BERT model to obtain the [SEP] encoding for each sentence. The [SEP] learns the information of the sentence itself and its contextual relationships with other sentences, thus fully capturing its semantics for the classification task. Therefore, a multilayer feedforward neural network (MLP) is used on the [SEP] of each sentence (Reference: Multilayer perceptron, fuzzy sets, classification [J]. IEEE Transactions on Neural Networks. 1992.), and the output is the classification result corresponding to that sentence.

[0079] Steps 1-4 (Sentence Classification): Concatenate the classification results of all text fragments to obtain the final classification result;

[0080] In this embodiment, the event basic information extraction stage employs an event extraction method, such as... Figure 5 The diagram shows a neural network structure containing a BERT-BiLSTM-CRF encoding layer, an attention layer, and an argument mapping layer. It can identify event trigger words and event types, and extract each argument and its role, specifically including:

[0081] (1) BERT-BiLSTM-CRF encoding layer: The network structure of pre-trained language model BERT + bidirectional LSTM + conditional random field CRF is adopted. The recognition of trigger words and event types is regarded as a sequence labeling problem, and the BIO labeling scheme is used for recognition.

[0082] The pre-trained language model BERT performs character-level semantic encoding on the input statement. Given a statement S i ={w1,w2,w3,…,w n-1 ,w n The encoded output is x = {x1, x2, x3, ..., x}. n-1 ,x n}

[0083] The Bi-LSTM encodes global information for the statement, learning contextual information for each word. The statement x = {x1, x2, x3, ..., x...} n-1 ,x n After forward LSTM and backward LSTM encoding (see: Bidirectional LSTM-CRF models for sequence tagging[J].arXiv preprint arXiv:1508.01991,2015.), the information outputs are as follows: and The forward and backward encoding results are concatenated to obtain Therefore, after passing through the Bi-LSTM encoding layer, the encoded result of the statement can be represented as h = {h1, h2, h3, ..., h...} n-1 ,h n}

[0084] The CRF layer identifies trigger words and their types in the statement, which are then used as input for subsequent argument recognition. The identification of trigger words and their types is treated as a sequence labeling model, employing a BX, IX, O labeling scheme (BIO labeling scheme). In BX, B represents the start position of the trigger word, X represents the event type of the trigger word, I represents the middle or end position of the trigger word in IX, and O represents a non-trigger word result. Each word is then labeled accordingly to achieve the identification of trigger words and their types, outputting a trigger word set T = {t1, t2, t3, ..., t...}.m}

[0085] (2) Attention Layer: Based on the trigger word set T, an attention mechanism is used to analyze the correlation between each word in the sentence and the trigger word, and then concatenates it with the output of the Bi-LSTM encoding layer as subsequent input. The specific calculation process is as follows:

[0086] For each trigger word t i Perform the following operations

[0087] q i =sum([h start :h end ])

[0088] Where [:] represents vector slicing, sum() represents vector addition, and start represents the trigger word t. i The beginning position, end indicates the trigger word t i The end position, q i Indicates trigger word t i The representation in h. Therefore, for the set of trigger words T, we can obtain Q = {q1,q2,q3,…,q}. m}

[0089] The representation q of the i-th trigger word i Using the attention mechanism, the following calculation is performed to obtain the correlation a between the j-th word and the i-th trigger word. ij :

[0090]

[0091] Then calculate the attention representation of the j-th word in the statement relative to the i-th trigger word: h j *a ij And concatenate it with the original representation:

[0092] h′ ij =[h j ,h j *a ij ]

[0093] Finally, obtain the hidden layer representation h′ of each word in the statement relative to the i-th trigger word. i ={h′ i1 ,h′ i2 ,h′ i3 ,…,h′ in}

[0094] (3) Argument Mapping Layer: A classification model is built for each argument role (such as time, location, participant) to identify whether each word is the start or end position of the argument. The model is input into h′ through a fully connected layer.i Mapped to binary classification,

[0095] O head =sigmoid(h′) i W head +b head )

[0096] O tail =sigmoid(h′) i W tail +b tail )

[0097] Among them W head b head For the weight parameters to predict the starting position, W tail b tail The weight parameters are used to predict the ending position.

[0098] In this embodiment, the event basic information extraction stage uses the loss function of the event extraction method for calculation. The loss function of the CRF is added to the loss function of the argument mapping layer to obtain the final loss function. The loss function is defined as follows:

[0099] loss = P(y|s) + binary_loss

[0100] Where P(y|s) is the loss function of CRF, and binary_loss is the loss function of the argument mapping layer, which is a binary cross entropy (BCE) loss function, as shown in the following formula.

[0101]

[0102] Among them, y i For the distribution of true label values, The distribution of predicted label values.

[0103] In this embodiment, the event sequence determination method used in the event context generation stage, when there is no explicit time description in the event, comprehensively uses time sequence connectors, event type pairs, and the transitivity of time sequence relationships to determine the sequence, specifically including:

[0104] (1) Temporal conjunctions: Construct a mapping table of temporal conjunctions such as "after", "cause", "before", "because" and temporal relationships (e.g., A causes B, indicating that A is before B), and constrain event pairs that have temporal relationships.

[0105] (2) Event type pairs: Construct a list of event type pairs such as “arrest → interrogation” and “earthquake → injury”, and constrain the temporal relationship between event pairs that conform to certain specific types.

[0106] (3) Transitivity of temporal relations: The temporal relations of events have the property of transitive closure. If relation(e0,e1) and relation(e1,e2) are known, then relation(e0,e2) can be deduced.

[0107] In this embodiment, the event coreference discovery method in the cross-chapter fusion stage comprehensively judges the similarity of two events from multiple dimensions such as time, location, participants, and type, specifically including:

[0108] The temporal correlation mentioned here refers to the fact that events often have a continuous process from beginning to end, and this process can occur at any point in time. Therefore, for two co-referenced events, their times are not exactly the same. To determine whether two events are similar in the temporal dimension, the interval between the two times needs to be used as a measure. If the difference between the times of the two events is within a certain range, then the two events are considered to be temporally correlated. The specific calculation formula is as follows:

[0109]

[0110] Where time(e) i () indicates the time of the event, window T It is a settable time window constant, such as 3 days.

[0111] Location relevance measures the degree of connection between events at different locations. One method is to assess the geographical hierarchy of locations; if names intersect, they are considered related. Another method is to measure the relevance of locations that do not have an inclusion relationship, by observing their distance. If the distance is within a certain range, the two events are considered geographically related. The strength of the correlation is related to the distance between locations. The closer the locations are, the stronger the correlation. The specific calculation formula is as follows:

[0112]

[0113] Where P(e) i The `dis()` method represents the location area of ​​the event, and `window` represents the distance between the center points of that area. P It is a distance constant.

[0114] The participant relevance is defined as follows: events often involve multiple participants. If two events involve the same participants or have an inclusion relationship, then the two events are considered related in terms of participants. The strength of the association is determined by the number of intersections. The specific calculation formula is as follows:

[0115]

[0116] Where actor(e) i ) represents the set of participants in an event.

[0117] The correlation between two events is defined as follows: if two events are of the same or similar types, they are considered correlated in type. This correlation can be measured by the distance between the two events in a classification system; the strength of the correlation is strongest when they are of the same type, denoted as Rel. C (e1,e2).

[0118] The overall relevance is obtained by weighting and summing the relevance of events in four aspects: time, location, participants, and type. Specifically:

[0119] Rel(e1,e2)=αRel T (e1,e2)+βRel P (e1,e2)+γRel A (e1,e2)+δRel C (e1,e2).

[0120] In its specific implementation, this application provides a computer storage medium and a corresponding data processing unit. The computer storage medium is capable of storing a computer program, which, when executed by the data processing unit, can run the inventive content of the cross-chapter event multi-dimensional profile generation method provided by this invention, as well as some or all of the steps in various embodiments. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0121] Those skilled in the art will clearly understand that the technical solutions in the embodiments of the present invention can be implemented using computer programs and their corresponding general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of computer programs, i.e., software products. These computer program software products can be stored in a storage medium and include several instructions to cause a device containing a data processing unit (which may be a personal computer, server, microcontroller, MUU, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.

[0122] This invention provides a method for generating multi-dimensional event profiles across chapters. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for generating multi-dimensional event profiles across chapters, characterized in that, The method comprises the following steps: The method comprises the following steps: Step 1, the text preprocessing stage, comprising: accessing text data, and classifying the text data according to relevance to obtain relevant text describing events of interest, and performing discourse structure analysis and fine-grained sentence segmentation on the relevant text to obtain text blocks describing different dimension information of events, including text blocks describing basic information, text blocks describing event progress, text blocks describing reasons and influences, text blocks describing comments of parties, and text blocks describing similar events; Step 2, the event basic information extraction stage, comprising: performing event extraction on the text blocks describing basic information to obtain basic elements of events, including the type, occurrence or end time, location and behavior subject of the event; Step 3, the event context generation stage, comprising: identifying sub-events from the text blocks describing event progress, and sorting the sub-events according to time sequence to form an event context; Step 4, the event cross-paragraph fusion stage, comprising: performing coreference discovery on the events extracted from different text blocks to form a complete event image; In step 2, the event extraction method is a neural network structure comprising a BERT-BiLSTM-CRF encoding layer, an attention layer and a mention mapping layer, which is used to identify event trigger words and event types, and extract each argument and its role, and specifically comprises: The BERT-BiLSTM-CRF encoding layer uses a network structure of a pre-trained language model BERT, a bidirectional LSTM and a conditional random field CRF, identifies the trigger words and event types as a sequence labeling problem, and uses a BIO labeling scheme for identification; The attention layer analyzes the relevance between each word in the sentence in the text block describing basic information and the trigger word, and concatenates the output of the Bi-LSTM encoding layer to capture the semantic relevance of the trigger word for each word; The mention mapping layer models a classification model for each argument role to identify whether each word in the sentence in the text block describing basic information is the start position or end position of the argument; In step 4, the coreference discovery method refers to the similarity of two events determined according to the overall correlation degree by the time correlation degree, the location correlation degree, the participant correlation degree and the type correlation degree.

2. The method of claim 1, wherein, The fine-grained sentence segmentation method in step 1 is to model the sentence segmentation as a sequence sentence classification problem, and use a BERT model to represent and classify the context of the sentences in the relevant text, and the specific method comprises: Step 1-1, text preprocessing: given a series of sentences S = <S1, …, S s >, s represents the number of sentences in the series of sentences, which is cut into multiple text segments of no more than 10 sentences, and the sequence sentence classification result of each text segment <S m , …, S k > is obtained by executing steps 1-2 and 1-3; wherein S1 represents the first sentence in the series of sentences S, S n n represents the nth sentence in the series of sentences S, S m m represents the mth sentence in the series of sentences S, and S k k represents the kth sentence in the series of sentences S. Step 1-2, input representation: for a text segment <S m ,…,S k >, concatenate a [SEP] separator after each sentence to form a sequence containing multiple sentences, and finally insert [CLS] at the beginning of the entire sequence to obtain the input representation; Step 1-3, sentence classification: input the input representation in step 1-2 into the BERT model to obtain the [SEP] encoding of each sentence, and input the result into a multi-layer feedforward neural network MLP to output the classification result of the sentence; Step 1-4, result summary: concatenate the classification results of all text segments obtained after step 1-3 in order to obtain the final classification result.

3. The method of claim 2, wherein, The BERT-BiLSTM-CRF encoding layer described in step 2 includes a pre-trained language model BERT, a Bi-LSTM encoding layer, and a CRF layer. The pre-trained language model BERT is used to semantically encode the words in the input sentence; given sentence S i ={w1,w2,w3,…,w n-1 ,w n }, where w n Statement S i The nth word in the sequence is encoded as x = {x1, x2, x3, ..., xn}. n-1 ,x n }, where x n Indicator w n The encoding; A Bi-LSTM encoding layer is used to encode global information for the statement, learning contextual information for each word; the statement x = {x1, x2, x3, ..., x...} n-1 ,x n After forward LSTM and backward LSTM encoding, the information outputs are respectively and in, x represents n The vector after forward LSTM encoding x represents n The vector after backward LSTM encoding is concatenated with the forward and backward encoding results to obtain the concatenated result. After passing through the Bi-LSTM encoding layer, the encoded result of the statement is represented as h = {h1, h2, h3, ..., h...} n-1 ,h n }, where h n This represents the vector after the nth character is encoded using Bi-LSTM. The CRF layer identifies the trigger word and its type in the sentence, which is used for subsequent argument recognition input; the identification of the trigger word and its type is regarded as a sequence labeling model, and a labeling scheme of B-X, I-X and O is adopted, wherein B in B-X represents the start position of the trigger word, X represents the event type of the trigger word, I in I-X represents the middle position or end position of the trigger word, and O represents the non-trigger word result; each word is labeled with a corresponding label, the identification of the trigger word and its type is realized, and a trigger word set T = {t1, t2, t3, …, t m} is output, wherein t m m represents the mth trigger word.

4. The method of claim 3, wherein, The attention layer described in step 2 includes using attention mechanism to analyze the relevance between each word in the sentence and the trigger words based on the trigger word set T, and concatenating the output of the Bi-LSTM encoding layer as the subsequent input; the specific calculation process is as follows: For the ith trigger word t i the following operations are performed q i = sum([h start : h end ]) Wherein, [:] represents vector intercept, sum() represents vector addition, start represents the starting position of trigger word t i , end represents the ending position of trigger word t i , q i represents the representation of trigger word t i in h, for the trigger word set T, the trigger word representation set Q = {q1, q2, q3, …, q m} is obtained. for the i-th trigger word q i , the following calculation is made to obtain the relevance a ij of the j-th word to the i-th trigger word using an attention mechanism: The attention representation of the jth word in the sentence with respect to the ith trigger word is then computed: h j a ij and concatenated with the original representation: h' ij = [h j ,h j *a ij ] Finally, the hidden layer representation h' of each word in the sentence with respect to the i-th trigger word is obtained i = {h' i1 , h' i2 , h' i3 ,..., h' in}.

5. The method of claim 4, wherein, The argument mapping layer described in step 2 includes: for the i-th trigger word t i , respectively, model the classification model for each argument role, identify whether each word is the start position or end position of the argument, and map the model input h′ i to binary classification through a fully connected layer: O head = sigmoid(h′ i W head +b head ) O tail = sigmoid(h′ i W tail +b tail ) where W head and b head are weight parameters for predicting the start position, W tail and b tail are weight parameters for predicting the end position, O head denotes a probability vector that each word prediction is a start position of an argument, and O tail denotes a probability vector that each word prediction is an end position of an argument.

6. The method of claim 5, wherein, The time sequence judgment method described in step 3 is to use the time sequence conjunction word, the event type pair and the time sequence relationship transitivity to judge the sequence when there is no explicit time description in the sub-event, which specifically includes: Time sequence conjunction word: construct a time sequence conjunction word and time sequence relationship mapping table to constrain the event pairs with time sequence relationship; Event type pair: construct an event type pair list to constrain the time sequence relationship between the event pairs of specific types; Time sequence relationship transitivity: the event time sequence relationship has the property of transitive closure, if relation(e0,e1) and relation(e1,e2) are known, then relation(e0,e2) is derived; wherein, relation(e0,e1) represents that events e0 and e1 have a time sequence relationship, relation(e1,e2) represents that events e1 and e2 have a time sequence relationship, and relation(e0,e2) represents that events e0 and e2 have a time sequence relationship.

7. The method of claim 6, wherein, The time correlation Rel described in step 4 T The (el, e2) is calculated as follows: wherein time(e1) represents the time of event e1, time(e2) represents the time of event e2, window T is a pre-set time window constant; The location relevance Rel P The specific calculation formula of (e1, e2) is as follows: where P(e1) represents the location area of event e1, P(e2) represents the location area of event e2, dis() is the distance between the center points of the areas, and window is a distance constant. P is a distance constant. The participant relevance Rel A The specific calculation formula of (e1, e2) is as follows: Wherein, actor(e1) represents the set of participants in event e1, and actor(e2) represents the set of participants in event e2; The type relevance is measured by the distance of two events e1 and e2 on the taxonomy, the strength of the association is strongest for the same type, denoted Rel C (e1,e2).

8. The method of claim 7, wherein, The overall correlation degree described in step 4 is obtained by weighting and summing the time correlation degree, the location correlation degree, the participant correlation degree and the type correlation degree of the event, which is specifically: Rel(e1, e2) = aRel T (e1, e2) + βRel P (e1, e2) + γRel A (e1, e2) + δRel C (e1, e2).

Images