Role-aware passage theme event argument extraction method and device

Abstract

The application provides a role-aware chapter theme event argument extraction method and device, and the method comprises the following steps: obtaining argument role information of a chapter theme event of an event type according to the event type; performing sentence segmentation and title extraction on a target article to obtain a sentence set and an event title; the argument role information, the event type, and the event title constitute event-related information; constructing an argument role-aware graph by using the event-related information and the sentence set, performing event-related sentence detection, and obtaining a chapter theme event-related sentence set; taking the chapter theme event-related sentence set as input, constructing a question for each argument role, predicting all candidate arguments in the chapter theme event-related sentence set, and screening out a target argument from the candidate arguments. The method improves the model effect while maintaining the flexibility of the model.

Description

Technical Field

[0001] This invention relates to the field of event extraction technology, and in particular to a method and apparatus for extracting chapter-themed event arguments based on role perception. Background Technology

[0002] Existing event extraction tasks can be categorized into sentence-level event extraction and document-level event extraction based on text granularity. Sentence-level event extraction focuses on identifying events that may be involved in a single sentence and determining the role of entities within that event. While sentence-level extraction considers sufficiently general event types, in real life, events are typically described through entire documents composed of multiple sentences. Compared to sentence-level event extraction, document-level event extraction has a more macro-level objective, focusing on events scattered throughout the entire document. Existing event extraction methods can be further divided into cross-sentence event extraction and document-theme event extraction, depending on the content being focused on. Cross-sentence event extraction is a natural extension of sentence-level event extraction: in real-world scenarios, all arguments of a sentence-level event generally do not appear in the same sentence as the trigger word; some arguments may also appear in other sentences, thus requiring the extraction of different arguments for the corresponding event from multiple different sentences. Document-theme events, on the other hand, focus on the main content described by the document, i.e., extracting the thematic events expressed by the document. While cross-sentence event extraction can extract events from arguments that may be scattered across different sentences in an entire document, its definition of events is essentially the same as that of sentence-level events. They only focus on a predefined event system and do not focus on the main content of the document. For any event that appears in the document, as long as it appears in the event system, it is extracted. Therefore, the events extracted by sentence-level event extraction methods and cross-sentence event extraction methods are relatively fragmented, and the defined argument roles are not fine-grained enough, so they cannot well summarize the main content of the document.

[0003] The extraction of topic events from current texts faces two challenges: argument sparsity and multiple references to argument roles. Argument sparsity and dispersion refer to the fact that the arguments of an event are scattered across multiple sentences and may only appear in a few sentences of the text, thus requiring a comprehensive understanding of the entire text for topic event extraction. Multiple references to argument roles refer to the fact that a single argument role may correspond to multiple argument references within the text. This invention primarily addresses these two problems.

[0004] To address the issue of sparsity and dispersion of arguments, some recent studies have proposed first extracting all candidate arguments from a text, then selecting target arguments, constructing a graph from sentences and candidate arguments, and finally using a graph attention neural network to identify event regions in the text and aggregate event information.

[0005] To address the problem of multiple references for argument roles, EEQA and PAIE have proposed specific solutions. EEQA employs a thresholding strategy, comparing the score of each text segment with a manually adjusted threshold. This means that significant time and computational resources are required to find a good threshold, often resulting in suboptimal outcomes. PAIE, to avoid thresholding multiple references for arguments with the same role, flexibly uses multiple slots for the same role in the cue learning template. The number of slots for a role is heuristically determined based on the maximum number of arguments for each role in the training dataset. However, both of these methods are cross-sentence event extraction methods, and many problems remain when extracting arguments for true text-level thematic events.

[0006] (1) Heuristic threshold adjustment depends on threshold selection.

[0007] The choice of threshold has a significant impact on model performance. Furthermore, models with the same architecture but different pre-trained language models will have completely different optimal thresholds even on the same dataset, let alone on different datasets.

[0008] (2) Setting the number of heuristic multireferences

[0009] Although PAIE does not require threshold adjustment because each time slot in the prompt can only predict at most one Analects text fragment, as mentioned above, the PAIE model needs to count the number of multiple references for different argument roles in different datasets in order to formulate a specific prompt template. Obviously, this approach is not generalizable.

[0010] (3) The learning method still falls short when faced with too many argument roles.

[0011] For an event, PAIE connects all argument roles as a template and feeds it into the BART decoder. This relies on the learning ability and text understanding ability of the BART model. PAIE is mainly aimed at event argument extraction datasets across sentences. The characteristic of this type of dataset is that the number of event argument roles is relatively small. Summary of the Invention

[0012] To address the aforementioned problems, this invention proposes a role-aware method and apparatus for extracting discourse theme event arguments, which better solves the problems of argument sparsity and dispersion and argument role multi-reference in the discourse theme event argument extraction task, improving the model performance while maintaining the model's flexibility.

[0013] To achieve the above objectives, the present invention provides a method for extracting chapter-based thematic event arguments based on role perception, comprising:

[0014] Obtain the argument role information of the chapter theme event of the known event type;

[0015] The target article is segmented into sentences and the titles are extracted to obtain a set of sentences and event titles. The argument role information, event type, and event title constitute event-related information.

[0016] Using the event-related information and the sentence set, an argument role perception graph is constructed, and event-related sentences are detected to obtain a set of sentences related to the chapter theme event;

[0017] Using the set of sentences related to the theme event of the passage as input, a question is constructed for each argument role, predicting all candidate arguments in the set of sentences related to the theme event of the passage, and then selecting the target argument from the candidate arguments.

[0018] Optionally, the step of constructing an argument role perception graph using the event-related information and the sentence segment set, and performing event-related sentence detection to obtain a set of sentences related to the chapter theme event, includes:

[0019] Encode each event-related information and each sentence in the sentence set;

[0020] Using the encoding of each clause and the encoding of each event-related information as nodes, an argument role perception graph is constructed. The argument role perception graph includes clause nodes, event-related information nodes, and the edges between the clause nodes and the event-related information nodes.

[0021] The argument role perception graph is modeled and iteratively interacted using a graph attention network to obtain the sentence feature representation of role perception;

[0022] The sentence feature representations of the character perception are binary classified to obtain a set of sentences related to the theme event of the chapter.

[0023] Optionally, encoding each event-related information and each sentence of the sentence set includes:

[0024] Each clause is segmented into words, and tags are added at the beginning and end of each clause.

[0025] For each event-related information, add tags between the event title, event type, and different argument role information;

[0026] The event-related information after adding tags, as well as the sentence set, are input into the BERT multilayer Transformer model to pre-encode the word sequence;

[0027] Self-attention text segment encoding is performed on the pre-encoded output sequence.

[0028] Optionally, the step of modeling and iteratively interacting the argument role-aware graph using a graph attention network to obtain a sentence feature representation of role awareness includes:

[0029] The graph attention layer is constructed by updating the representation of each node in the argument role perception graph using a graph attention network.

[0030] After each graph attention layer, a feedforward FFN layer consisting of two linear transformations is used;

[0031] The sentence nodes are updated using adjacent word nodes through the graph attention layer and FFN layer;

[0032] The updated clause nodes are used to obtain new event-related information node representations, and the clause nodes are further iterated and updated.

[0033] Optionally, the step of performing binary classification on the sentence feature representation of the role perception to obtain a set of sentences related to the chapter theme event includes:

[0034] The sentence features perceived by the role are used to predict the label of the sentence, thus obtaining the predicted label;

[0035] By performing binary classification using the predicted labels and the actual labels, a set of sentences related to the theme and events of the text is obtained.

[0036] Optionally, the construction problem for each argument role includes:

[0037] Add a generic semantic type to each argument role;

[0038] Using general semantic types, event types, and argument role information, generate questions for each argument role.

[0039] Optionally, predicting all candidate arguments in the set of sentences related to the chapter's topic event includes:

[0040] The generated question and the set of sentences related to the theme and event of the chapter are concatenated to obtain the input sequence;

[0041] The input sequence is pre-encoded using a BERT multilayer Transformer model;

[0042] The output representation of the last four graph attention layers of the pre-encoded BERT multi-layer Transformer model is split into a representation of the problem and a representation of the paragraph.

[0043] Using the representation of the problem and the representation of the paragraph, a standard decoding strategy is used to predict the start and end positions of the pre-encoded input sequence, and legal answer segments are selected from all predicted answer segments to obtain all candidate arguments.

[0044] Optionally, the step of filtering target arguments from the candidate arguments includes:

[0045] Using the representation of the question and the representation of the paragraph, calculate the number of answers to the question to obtain the number of arguments;

[0046] The target argument is obtained by using a non-maximum suppression algorithm to extract and select a corresponding number of arguments from all candidate arguments.

[0047] In another aspect, the present invention provides a role-aware text theme event argument extraction device, employing the aforementioned role-aware text theme event argument extraction method, including:

[0048] The event-related information extraction module is used to obtain the argument role information of the chapter theme event of a known event type based on the event type of the chapter theme event;

[0049] The target article is segmented into sentences and the titles are extracted to obtain a set of sentences and event titles. The argument role information, event type, and event title constitute event-related information.

[0050] The chapter theme event-related sentence detection module is used to construct an argument role perception graph using the event-related information and the sentence set, and to detect event-related sentences to obtain a set of chapter theme event-related sentences;

[0051] The chapter theme event argument extraction module is used to take the set of sentences related to the chapter theme event as input, construct a question for each argument role, predict all candidate arguments in the set of sentences related to the chapter theme event, and select the target argument from the candidate arguments.

[0052] Optionally, the chapter-theme event-related sentence detection module further includes:

[0053] The text information encoding module is used to encode each event-related information and each sentence of the sentence set;

[0054] The argument role perception graph module is used to construct an argument role perception graph using the encoding of each clause and the encoding of each event-related information as nodes. The argument role perception graph includes clause nodes, event-related information nodes, and the edges between the clause nodes and the event-related information nodes.

[0055] The argument role perception graph is modeled and iteratively interacted using a graph attention network to obtain the sentence feature representation of role perception;

[0056] The sentence classification module is used to perform binary classification on the sentence feature representations perceived by the role to obtain a set of sentences related to the theme event of the chapter.

[0057] As can be seen from the above solutions, the advantages of the present invention are:

[0058] The role-aware text-topic event argument extraction method provided by this invention has practical advantages. First, addressing the sparsity and dispersion problem of text-topic event arguments, it comprehensively considers the semantic information of the text-topic event itself and extracts event-related information based on practical needs. Then, it utilizes a graph attention neural network to consider the relationship between different nodes such as event-related information and sentence sets from a global perspective, and captures the high-order interactions between sentence nodes and event-related information nodes to detect event-related sentences, obtaining a set of text-topic event-related sentences. This effectively removes event-irrelevant interference information, thereby effectively reducing the interference of irrelevant information on the model and improving the overall performance of the text-topic event extraction model.

[0059] Furthermore, addressing the issue of multiple referentials of argument roles in text-specific events, this invention, based on practical needs and comprehensively considering the semantic information of the text-specific events themselves, uses a set of sentences related to the text-specific events as input. It constructs a question for each argument role, learns the characteristics of semantic features at different levels of the text, predicts all candidate arguments in the set of sentences related to the text-specific events, and then selects the target argument from the candidate arguments, thus improving the argument extraction performance of the text-specific event extraction model. In addition, this invention uses a non-heuristic method, making it more generalizable. Attached Figure Description

[0060] Figure 1 A flowchart illustrating the chapter-theme event argument extraction method for role perception provided in an embodiment of the present invention is shown.

[0061] Figure 2 It shows Figure 1 The detailed flowchart of step S3;

[0062] Figure 3 It shows Figure 2 Overall structure diagram of the section on detecting sentences related to the theme and events of a passage;

[0063] Figure 4 This is a schematic diagram of residual connections in a graph attention network.

[0064] Figure 5 It shows Figure 1 The detailed flowchart of step S4;

[0065] Figure 6 It shows Figure 5 Structure diagram of the method for extracting thematic event arguments from a text;

[0066] Figure 7 The diagram shows the architecture of the chapter-theme event argument extraction device 20 for role perception;

[0067] in,

[0068] 20-Role perception chapter theme event argument extraction device;

[0069] 21-Event-related information extraction module;

[0070] 22-Chapter-themed event-related sentence detection module;

[0071] 221 - Text Information Encoding Module;

[0072] 222 - Argument Role Perception Graph Module;

[0073] 223 - Sentence Classification Module;

[0074] 23 - Chapter Theme Event Argument Extraction Module;

[0075] 231 - Problem Generation Module;

[0076] 232-Context Encoding Module;

[0077] 233-Argument Prediction Module;

[0078] 234 - Multi-argument extraction module. Detailed Implementation

[0079] To make the above features and effects of the present invention clearer and easier to understand, specific embodiments are described below, and detailed descriptions are provided in conjunction with the accompanying drawings.

[0080] To overcome the argument sparsity and dispersion problem in the extraction of thematic events from the Analects, this invention introduces a thematic event-related sentence detection task. Specifically, it is achieved by constructing a thematic event-related sentence detection method based on a role-aware graph. First, event-related information and the sentences in the text are constructed into a role-aware heterogeneous graph. Then, an improved graph attention network is used to iteratively interact with the event information nodes and sentence nodes, enabling the sentences to learn the global information related to the event, thereby obtaining the role-aware sentence representation. Finally, the role-aware sentence representation is binary classified to obtain a set of event-related sentences, which serve as the input for subsequent argument extraction.

[0081] Furthermore, to overcome the problem of multiple references to argument roles in the extraction of thematic events from the Analects, this invention transforms the extraction of multiple references to argument roles into a multi-answer question-and-answer task for event argument extraction. First, the number of text segments to be extracted is predicted and modeled as a classification problem. After obtaining the number of answers, to extract a specific number of non-overlapping text segments, this invention employs a non-maximum suppression algorithm to sort the predicted answers according to their scores, finally obtaining the target answer. The invention will be described in detail below.

[0082] like Figure 1 As shown, Figure 1 A flowchart illustrating the chapter-theme event argument extraction method for role perception provided in an embodiment of the present invention is shown below:

[0083] A method for extracting chapter-based thematic event arguments based on role perception, including:

[0084] S1. Obtain the argument role information of the chapter theme event based on the known event type of the chapter theme event.

[0085] S2. Segment the target article and extract the title to obtain a set of sentences and event titles. The argument role information, event type, and event title constitute event-related information.

[0086] S3. Using the event-related information and the sentence set, construct an argument role perception graph, perform event-related sentence detection, and obtain a set of sentences related to the chapter theme event.

[0087] In specific implementations, such as Figure 2 As shown, Figure 2 A detailed flowchart of step S3 is shown. Figure 3 This diagram illustrates the overall structure of the sentence detection section related to the main event of this passage. Specifically, it includes the following steps:

[0088] S31. Encode each event-related information and each sentence of the sentence set.

[0089] Specifically, each clause is segmented into words, and tags are added at the beginning and end of each clause. Simultaneously, for each event-related information item, tags are added between the event title, event type, and different argument roles. In practice, the WordPieceTokenizer module in the Transformers library can be used to segment each clause, and then special tags [CLS] and [SEP] consistent with BERT and the training task are added to the beginning and end of each clause. When encoding event-related information, tags [SEP] and [CLS] need to be added to each semantic unit, i.e., between the event title, event type, and different argument roles.

[0090] Then, the event-related information after adding labels and the sentence set are input into the BERT multilayer Transformer model to pre-encode the word sequence, that is:

[0091]

[0092] Among them, w i,j w represents the j-th token of the i-th semantic unit. i,0 and The tag character representing the i-th semantic unit <cls>and <sep>h i,j This represents the hidden layer representation of the corresponding token, i.e., the pre-encoded output sequence.

[0093] Then, the pre-encoded output sequence is subjected to self-attention text segment encoding, i.e.:

[0094] For the i-th semantic unit, this semantic unit is composed of l i Composed of 10 words, the output sequence pre-encoded by BERT is The representation of this semantic unit can then be obtained using the following method:

[0095] α i,j =W2·ReLU(W1h) ij +b1)+b2

[0096]

[0097]

[0098] Where, α i,j It is the attention score of the j-th word in the i-th semantic unit, a ij It is the normalized attention score of the j-th word. It is a weighted sum of the hidden states output by BERT, where b1, b2, and W1 are all learnable parameters. The above process can be simply represented as...

[0099] S32. Using the encoding of each clause and the encoding of each event-related information as nodes, construct an argument role perception graph. The argument role perception graph includes clause nodes, event-related information nodes, and the edges between the clause nodes and the event-related information nodes.

[0100] Specifically, let G = {V, E} denote any graph, where V represents the set of nodes and E represents the set of edges; formally, the role-aware heterogeneous graph can be defined as V = V S ∪V E E = {e} 1,1 ,…,e N,N }, where V s ={s1,s2,…,s m } represents m clause nodes, V R ={R1,R2,R... n } represents n event entity nodes. i,j The edge representing the i-th sentence and the j-th event is used to connect sentence-event related information.

[0101] S33. The argument role perception graph is modeled and iteratively interacted using a graph attention network to obtain the sentence feature representation of role perception.

[0102] Specifically, the process includes two parts: graph attention network modeling and iterative interaction.

[0103] For the graph attention network modeling part, h is first used i Let i ∈ {1,…,(m+n)} be the representation of the input nodes. A graph attention network is used to update the representation of each node in the argument role perception graph, constructing a graph attention layer. This can be represented as:

[0104] z i,j =LeaklyReLU(W a [W e h i W e h j ])

[0105]

[0106]

[0107] Among them, W a W e W v All are learnable parameters, α ij For node h i and h j The attention weights between them, multi-head attention can be represented as:

[0108]

[0109] Then, residual connections are used to avoid gradient vanishing after multiple iterations. A schematic diagram of residual connections is shown below. Figure 4 As shown, therefore, the final output can be represented as:

[0110] h′ i =u i +h i

[0111] Simultaneously, after each graph attention layer, a feedforward FFN layer consisting of two linear transformations is used.

[0112] For the iterative interaction part, messages are passed between event-related information and sentence nodes. Specifically, after initialization, the sentence nodes are updated with adjacent word nodes through the graph attention layer and the FFN layer, i.e.:

[0113]

[0114]

[0115] in, express For attention queries, For attention keys and values.

[0116] Then, the updated sentence nodes are used to obtain new event-related information node representations, and the sentence nodes are further iterated and updated, that is, the sentence feature representations of role perception are updated.

[0117] Each iteration includes an update process from sentence to event-related information (e←s) and from event-related information to sentence (s←e).

[0118] For the t-th iteration, the process can be represented as:

[0119] Clause segmentation to event-related information:

[0120] Updated event-related information nodes represent:

[0121] Event-related information to clauses:

[0122] Updated sentence feature representation:

[0123] S34. Perform binary classification on the updated sentence feature representations of character perception to obtain a set of sentences related to the chapter's theme and events. Specifically:

[0124] First, the sentence features of the character perception are represented. Predict the label of the sentence to obtain the predicted label, i.e.:

[0125]

[0126] in, Let σ(·) represent the predicted label, and let σ(·) represent the Logistic function. It is the predicted label probability between 0 and 1, i.e., the predicted label. Sentence features representing role perception, W y and b y All are learnable parameters;

[0127] Then, binary classification is performed using the predicted labels and the true labels to obtain a set of sentences related to the article's theme and events. The constructed loss function is:

[0128]

[0129] in, For the true label of the sentence, This is the loss function.

[0130] S4. Using the set of sentences related to the chapter's theme event as input, construct a question for each argument role, predict all candidate arguments in the set of sentences related to the chapter's theme event, and select the target argument from the candidate arguments. For example... Figure 5 As shown, Figure 5 A detailed flowchart of step S4 is shown. Figure 6 This diagram illustrates the overall structure of the thematic event argument extraction for this chapter. Specifically, it includes:

[0131] S41. Add a general semantic type to each argument role, and use the general semantic type, event type, and argument role information to generate a question for each argument role.

[0132] In practice, such as Figure 6 As shown, the semantic information of argument roles can be enriched by incorporating definitions from the Oxford Dictionary. First, a general semantic type for each argument role is determined, such as person, place, etc., to introduce the interrogative word "WH," for example, "who" for people, "where" for places, and "what" for other situations. Then, to allow the model to better learn event-related information, the event type is also incorporated into the question; therefore, the question can be represented as:<WH_word> is the<role information> in <eventtype>?

[0133] Among them, role information refers to the information on the role of arguments. For technical terms or ambiguous words in the original argument roles, the Oxford Dictionary is used to explain them; and the original argument roles that do not conform to conventional language expression are also converted accordingly.

[0134] S42. Concatenate the generated question and the set of sentences related to the chapter's theme event to obtain the input sequence;

[0135] The input sequence is pre-encoded using the BERT multilayer Transformer model to obtain the distributed semantic representation of each word after BERT encoding;

[0136] Then, the output representation of the last four Transformer graph attention layers of the pre-encoded BERT multi-layer Transformer model is split into a representation of the question and a representation of the paragraph. Specifically,

[0137] First, after determining the question, the generated question and the text paragraphs containing sentences related to the chapter's theme and events are concatenated and represented as the input sequence:

[0138] [CLS] <questions>[SEP] <paragraph>[SEP]

[0139] Then, the output of each Transformer layer in the BERT multi-layer Transformer model is represented as H. i Then, a series of L pre-trained transformer blocks (the BERT-large used in this invention has 24 layers) project the input embedding into the context representation, as shown below:

[0140]

[0141] Then, take the output representation of the last four transformer layers of the pre-encoded BERT multi-layer Transformer model (H L-3 ,…,H L The terms are represented as M0, M1, M2, and M3, respectively, and broken down into representations of the question and paragraphs; for example, M2 is divided into a question representation Q2 and a paragraph representation P2 using [SEP]. The model then calculates two vectors summarizing the question and the article information, respectively. and

[0142] α Q =softmaxt(W Q Q2)

[0143]

[0144] Calculated in the same way

[0145] Finally, calculate the probability that the question has an answer. Calculate a probability distribution that represents the choice between whether the question has an answer:

[0146]

[0147] Here, h CLS It is the first vector in the last layer context representation M3. FFN represents a feedforward network consisting of two linear projections with Gelu activations, with a layer normalization in between.

[0148] S43. Using the representation of the question and the representation of the paragraph, a standard decoding strategy is used to predict the start and end positions of the pre-encoded input sequence. Valid answer segments are then selected from all predicted answer segments to obtain all candidate arguments. Specifically:

[0149] We compute problem representations at three different levels. Specifically, this is done by computing three vectors, namely... They summarize the problem information between different levels of problem representation:

[0150]

[0151] in and Calculated from Q0 and Q1 respectively.

[0152] Then, the probabilities of the start and end of the answer segment are calculated from the input sequence as follows:

[0153]

[0154]

[0155] Probability of starting segment:

[0156] Probability of ending segment:

[0157] in, The outer product of vector h and each token representation M.

[0158] Finally, the loss is calculated. Specifically, the training is performed by minimizing the cross-entropy between the start and end indices of the target span.

[0159]

[0160] in, Indicates the start of index loss, This indicates the end of the index loss.

[0161] During testing, valid answer fragments are selected from all predicted answer fragments by finding the metric (s,e), thus obtaining all candidate arguments, i.e.:

[0162]

[0163] S44. Using the representation of the question and the representation of the paragraph, calculate the number of answers to the question to obtain the number of arguments;

[0164] The target argument is obtained by using a non-maximum suppression algorithm to extract and select a corresponding number of arguments from all candidate arguments. Specifically,

[0165] First, predict the number of answers and model it as a classification problem, which is done by calculating the probability distribution of the number of answers:

[0166]

[0167] Among them, h CLS This is the first vector representing M3 in the last layer.

[0168] Then, multiple arguments are extracted using a non-maximum suppression algorithm. To extract a specific number of non-overlapping answers, this invention employs a non-maximum suppression (NMS) algorithm for redundancy pruning. Specifically, a top-K set of answers S is first proposed based on the descending order of answer scores, calculated as the score of the text segment (kl). It also predicts from P span The number of answers t extracted, and a new set initialized. Next, the segment with the highest answer score will be s. i Add to collection And remove it from S. Also delete anything related to s. i Any remaining overlapping fragments j The degree of overlap is evaluated using the text-level F1 function. This process is repeated for the remaining answers in S until S is empty or... The size reaches t.

[0169] During training, there are two objective functions: one is the far-supervised loss, which maximizes the probability of all matching answer text fragments, and the other is the classification loss, which maximizes the probability of the number of answers. During testing, the model first selects the answer type and then executes a specific prediction strategy. During testing, it obtains text fragments of candidate answers (i.e., all candidate arguments), then predicts the number of answers (i.e., the number of argument references), and finally selects candidate arguments with the corresponding number of argument references from all candidate arguments as the target arguments.

[0170] In summary, the role-aware text-topic event argument extraction method provided by this invention has practical advantages. First, addressing the sparsity and dispersion problem of text-topic event arguments, it comprehensively considers the semantic information of the text-topic event itself from a practical perspective to extract event-related information. Then, it utilizes a graph attention neural network to consider the relationships between different nodes, such as event-related information and sentence sets, from a global perspective, capturing high-order interactions between sentence nodes and event-related information nodes to detect event-related sentences, thus obtaining a set of text-topic event-related sentences. This effectively removes event-irrelevant interference information, thereby significantly reducing the interference of irrelevant information on the model and improving the overall performance of the text-topic event extraction model.

[0171] Furthermore, addressing the issue of multiple referentials of argument roles in text-specific events, this invention, based on practical needs and comprehensively considering the semantic information of the text-specific events themselves, uses a set of sentences related to the text-specific events as input. It constructs a question for each argument role, learns the characteristics of semantic features at different levels of the text, predicts all candidate arguments in the set of sentences related to the text-specific events, and then selects the target argument from the candidate arguments, thus improving the argument extraction performance of the text-specific event extraction model. In addition, this invention uses a non-heuristic method, making it more generalizable.

[0172] Furthermore, the above embodiments of the present invention can be applied to terminal devices that implement the chapter theme event argument extraction method for role perception. These terminal devices may include personal terminals and host computer terminals, etc., and the embodiments of the present invention do not limit this. The terminal can support operating systems such as Windows, Android, iOS, and Windows Phone.

[0173] A role-aware text theme event argument extraction device 200, applied to a role-aware text theme event argument extraction method, can be used in personal terminals and host computer terminal devices. It can achieve [the following] through [methods]: Figure 1 , Figure 2 , Figure 5 The role-aware chapter theme event argument extraction method shown in this application embodiment, and the role-aware chapter theme event argument extraction device provided in this application embodiment, can realize each process of the above-mentioned role-aware chapter theme event argument extraction method. Figure 7 The diagram shows the architecture of the chapter-theme event argument extraction device 20 perceived by the character.

[0174] A role-aware chapter-theme event argument extraction device 20, comprising at least:

[0175] The event-related information extraction module 21 is used to obtain the argument role information of the known chapter theme event based on the event type of the known chapter theme event;

[0176] The target article is segmented into sentences and the titles are extracted to obtain a set of sentences and event titles. The argument role information, event type, and event title constitute event-related information.

[0177] The chapter theme event-related sentence detection module 22 is used to construct an argument role perception graph using the event-related information and the sentence set, and to detect event-related sentences to obtain a set of chapter theme event-related sentences.

[0178] The chapter theme event argument extraction module 23 is used to take the set of sentences related to the chapter theme event as input, construct a question for each argument role, predict all candidate arguments in the set of sentences related to the chapter theme event, and select the target argument from the candidate arguments.

[0179] In addition, for the chapter-themed event-related sentence detection module 22, such as Figure 3 As shown, it mainly includes:

[0180] The text information encoding module 221 is used to encode each event-related information and each sentence of the sentence set;

[0181] Argument role perception graph module 222 is used to construct an argument role perception graph using the encoding of each clause and the encoding of each event-related information as nodes. The argument role perception graph includes clause nodes, event-related information nodes, and the edges between the clause nodes and the event-related information nodes.

[0182] The argument role perception graph is modeled and iteratively interacted using a graph attention network to obtain the sentence feature representation of role perception;

[0183] The sentence classification module 223 is used to perform binary classification on the sentence feature representation of the role perception to obtain a set of sentences related to the theme event of the chapter.

[0184] For the chapter topic event argument extraction module 23, such as Figure 6 As shown, it mainly includes:

[0185] The question generation module 231 is used to add a general semantic type to each argument role, and to generate a question for each argument role using the general semantic type, event type, and argument role information.

[0186] Context encoding module 232 is used to concatenate the generated question and the set of sentences related to the chapter's theme event to obtain an input sequence;

[0187] The input sequence is pre-encoded using a BERT multilayer Transformer model;

[0188] The output representation of the last four graph attention layers of the pre-encoded BERT multi-layer Transformer model is split into a representation of the problem and a representation of the paragraph.

[0189] Argument prediction module 233 is used to predict the start and end positions of the pre-encoded input sequence using the representation of the question and the representation of the paragraph, and to filter out the legal answer segments from all predicted answer segments to obtain all candidate arguments.

[0190] The multi-argument extraction module 234 is used to calculate the number of answers to the question and the number of arguments by using the representation of the question and the representation of the paragraph; and to extract and select the corresponding number of arguments from all candidate arguments by using a non-maximum suppression algorithm to obtain the target argument.

[0191] Furthermore, it should be understood that the division of the above-described functional modules in the role-aware chapter topic event argument extraction device 20 according to the embodiments of this application is only an example. In actual applications, the above functions can be assigned to different functional modules as needed. That is, the role-aware chapter topic event argument extraction device 20 can be divided into functional modules different from the modules illustrated above to complete all or part of the functions described above.

[0192] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one…" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be applied, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0193] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.< / paragraph> < / questions> < / eventtype> < / sep> < / cls>

Claims

1. A method for extracting chapter-based thematic event arguments based on role perception, characterized in that, include: Obtain the argument role information of the chapter theme event of the known event type; The target article is segmented into sentences and the titles are extracted to obtain a set of sentences and event titles. The argument role information, event type, and event title constitute event-related information. Using the event-related information and the sentence segment set to construct an argument role perception graph, event-related sentence detection is performed to obtain a set of sentences related to the chapter theme event, including: The BERT multi-layer Transformer model is used to encode each event-related information and each sentence in the sentence set, including: Each clause is segmented into words, and tags are added at the beginning and end of the clause; for each event-related information, tags are added between the event title, event type, and different argument roles. The event-related information after adding tags and the sentence set are input into the BERT multilayer Transformer model to pre-encode the word sequence; the pre-encoded output sequence is then subjected to self-attention text fragment encoding. Using the encoding of each clause and the encoding of each event-related information as nodes, an argument role perception graph is constructed. The argument role perception graph includes clause nodes, event-related information nodes, and the edges between the clause nodes and the event-related information nodes. The argument role perception graph is modeled and iteratively interacted using a graph attention network to obtain a sentence feature representation of role perception, including: The graph attention network is used to update the representation of each node in the argument role perception graph, and a graph attention layer is constructed. After each graph attention layer, a feedforward FFN layer consisting of two linear transformations is used. The sentence nodes are updated with adjacent word nodes through the graph attention layer and the FFN layer. The updated sentence nodes are used to obtain new event-related information node representations, and the sentence nodes are iteratively updated further. The sentence feature representations of the character perception are binary classified to obtain a set of sentences related to the chapter theme event; The set of sentences related to the theme event of the text is taken as input. A general semantic type is added to each argument role. Using the general semantic type, event type, and argument role information, a question for each argument role is generated. All candidate arguments in the set of sentences related to the theme event of the text are predicted. The target argument is selected from the candidate arguments. The prediction of all candidate arguments in the set of sentences related to the theme event of the text includes: The generated question and the set of sentences related to the chapter's theme event are concatenated to obtain the input sequence. The input sequence is pre-encoded using a BERT multilayer Transformer model. The output representation of the last four graph attention layers of the pre-encoded BERT multilayer Transformer model is split into a question representation and a paragraph representation. Using the question representation and the paragraph representation, a standard decoding strategy is used to predict the start and end positions of the pre-encoded input sequence. Valid answer fragments are selected from all predicted answer fragments to obtain all candidate arguments.

2. The method according to claim 1, characterized in that, The sentence feature representation of the character perception is binary classified to obtain a set of sentences related to the chapter theme event, including: The sentence features perceived by the role are used to predict the label of the sentence, thus obtaining the predicted label; By performing binary classification using the predicted labels and the actual labels, a set of sentences related to the theme and events of the text is obtained.

3. The method according to claim 1, characterized in that, The step of selecting target arguments from the candidate arguments includes: Using the representation of the question and the representation of the paragraph, calculate the number of answers to the question to obtain the number of arguments; The target argument is obtained by using a non-maximum suppression algorithm to extract and select a corresponding number of arguments from all candidate arguments.

4. A device for extracting chapter-themed event arguments based on role perception, characterized in that, The method for extracting chapter-theme event arguments based on role perception as described in any one of claims 1-3 includes: The event-related information extraction module is used to obtain the argument role information of the chapter theme event of a known event type based on the event type of the chapter theme event; The target article is segmented into sentences and the titles are extracted to obtain a set of sentences and event titles. The argument role information, event type, and event title constitute event-related information. The chapter-theme event-related sentence detection module is used to construct an argument role perception graph using the event-related information and the sentence segment set, and to detect event-related sentences to obtain a set of chapter-theme event-related sentences, including: Encoding each event-related information and each sentence in the sentence set includes: Each clause is segmented into words, and tags are added to the beginning and end of the clause. For each event-related information, tags are added between the event title, event type, and different argument roles. The event-related information with added tags and the sentence set are input into the BERT multilayer Transformer model to pre-encode the word sequence. The pre-encoded output sequence is then subjected to self-attention text fragment encoding. Using the encoding of each clause and the encoding of each event-related information as nodes, an argument role perception graph is constructed. The argument role perception graph includes clause nodes, event-related information nodes, and the edges between the clause nodes and the event-related information nodes. The argument role perception graph is modeled and iteratively interacted using a graph attention network to obtain a sentence feature representation of role perception, including: The graph attention network is used to update the representation of each node in the argument role perception graph, and a graph attention layer is constructed. After each graph attention layer, a feedforward FFN layer consisting of two linear transformations is used. The sentence nodes are updated with adjacent word nodes through the graph attention layer and the FFN layer. The updated sentence nodes are used to obtain new event-related information node representations, and the sentence nodes are iteratively updated further. The sentence feature representations of the character perception are binary classified to obtain a set of sentences related to the chapter theme event; The chapter theme event argument extraction module is used to take the set of sentences related to the chapter theme event as input, add a general semantic type to each argument role, generate a question for each argument role using the general semantic type, event type, and argument role information, predict all candidate arguments in the set of sentences related to the chapter theme event, and select the target argument from the candidate arguments; The prediction of all candidate arguments in the set of sentences related to the theme event of the text includes: The generated question and the set of sentences related to the chapter's theme event are concatenated to obtain the input sequence. The input sequence is pre-encoded using a BERT multilayer Transformer model. The output representation of the last four graph attention layers of the pre-encoded BERT multilayer Transformer model is split into a question representation and a paragraph representation. Using the question representation and the paragraph representation, a standard decoding strategy is used to predict the start and end positions of the pre-encoded input sequence. Valid answer fragments are selected from all predicted answer fragments to obtain all candidate arguments.

Images