Event causal relationship identification method based on information enhancement

Abstract

The invention belongs to the technical field of natural language processing, and particularly relates to an event causal relationship identification method based on information enhancement. In order to solve the problems that an existing event causal relationship identification method rarely considers causal label information of events and rarely pays attention to interaction information between event pairs and the like, the event causal relationship identification method mainly comprises the following two modules: 1) an event causal label sensing module, and provides an event pair representation method of event causal label sensing; an event causal label prediction task is introduced to serve as an auxiliary task, the dependency relationship between events is mined, and whether the events are causal-related events or not is judged. And 2) an event pair interaction graph module, which proposes an event pair interaction relationship modeling method based on a graph neural network, constructs an interaction graph by taking event pairs as nodes, captures a dependency relationship among the nodes by using a graph attention mechanism, and further discriminates a causal relationship among events.

Description

Technical Field

[0001] The present invention belongs to the technical field of natural language processing, and in particular relates to an event causal relationship identification method based on information enhancement. Background Art

[0002] Event causal relationship recognition is an important research task in the field of natural language processing, which aims to identify the causal relationship between given events in the text. For example, given the sentence "On Friday, a magnitude 8.9 earthquake occurred in northeastern Japan, causing a large number of houses to collapse", the model needs to identify the causal event pair <earthquake, causing, houses to collapse>. Causality, as a semantic relationship, is of great significance for semantic understanding and paragraph analysis. In addition, the causal relationship identified from the text can be used as a kind of knowledge to apply to many natural language processing tasks, such as machine reading comprehension, relational reasoning, etc. Related research on event causal relationship recognition tasks can be divided into methods based on statistical feature semantic patterns and methods based on deep learning.

[0003] In early studies, grammatical rules or statistical features were usually used to solve the task of event causal relationship identification, including causal clues (such as "cause" and "therefore"), event co-occurrence patterns, lexical and syntactic features, etc. For example, Beamer et al. counted the sequential events in the corpus to infer whether there is a causal relationship between temporally adjacent events when there is less text information. Do et al. used distribution similarity and discourse connectives to capture the causal relationship between events. Riaz et al. identified the causal relationship between gerund pairs by mining the deep semantic information of nouns and verbs. Ning et al. designed a joint model to jointly identify causal and temporal relationships using the intrinsic constraints between these two types of relationships. Gao et al. used integer linear programming to capture the fine-grained and global features of document-level causal relationships, including event types, syntactic relationships, discourse relationships, and coreference information. Methods based on lexical and syntactic features rely on manually designed features, have weak generalization capabilities for events, and have poor domain transferability.

[0004] In recent years, with the rapid development of deep learning technology in the field of natural language processing, event causal relationship recognition methods based on deep learning have become a research hotspot. Due to the diverse forms of event expression, which can exist in various forms of explicit or implicit expressions, most current research works are based on language features and external knowledge to identify causal relationships. Liu et al. proposed a mention mask reasoning mechanism to mine event-independent context-specific patterns for reasoning, enhance the generalization ability of the model, and enrich the representation of events by introducing external knowledge. Based on the idea of ​​remote supervision, Zuo et al. proposed a knowledge fusion data enhancement framework for event causal relationship detection to alleviate the problem of data shortage. In 2021, Zuo et al. further proposed a learnable knowledge-guided data enhancement framework, which uses the duality between generation and recognition to generate sentences related to the event causal relationship recognition task through dual learning. In addition, Zuo et al. also proposed a self-supervised framework for learning context-specific causal patterns from external causal statements, and adopted a contrastive transfer strategy to integrate the learned context-specific causal patterns into the target model. In the same year, Cao et al. proposed a latent structure induction network to use external structural knowledge to complete event causal relationship tasks, introduced a description graph induction module to obtain and encode descriptive knowledge based on graph structure, and introduced a relationship graph induction module to summarize a more reasonable reasoning path for causal reasoning. Phu et al. further proposed a document-level event causal relationship recognition model to solve the limitation of previous methods that only focus on event mentions within sentences, and captured document-level event information by introducing chapter, syntax and similarity information. In addition, Gao et al. proposed a Chinese-Vietnamese news causal relationship recognition model based on type matrix transfer, which effectively combined the semantic features obtained by the Chinese-Vietnamese shared embedding layer based on adversarial training, and the tree-shaped LSTM to extract and calculate the causal relationship pairs of event sentence pairs classified by similarity semantic information.

[0005] Existing methods have achieved good performance for event causal relationship recognition tasks, but they usually generate additional training data by integrating common sense knowledge or using data augmentation to improve the performance of event causal relationship recognition, and pay less attention to the characteristics of the data itself, mainly including the following two points:

[0006] (1) Event causal label information is rarely used.

[0007] In the event causal relationship recognition task, a sentence usually contains multiple events, but not all events are involved in causal relationships. For example, Figure 1 The sentence in (a) contains 7 events in total. The number of candidate event pairs that can be formed is as follows: Figure 1As shown in (b), only three of the event pairs have causal relationships, so the scope of causal relationship judgment is relatively large. When considering this task in the way humans think, they usually first preliminarily screen out which events may have causal relationships based on the context semantics and the relationship between events, and then further make a detailed judgment on the relationship between them. Therefore, if the causal relationship identification model can first screen out e3, e5, and e7 as causally related events in this way, then the event pairs consisting of causally related events are Figure 1 The dark area in (b) can narrow the scope of causal relationship recognition to a certain extent. Therefore, it is necessary to introduce event causal label information to assist in the task of event causal relationship recognition. However, as far as we know, existing methods rarely use this information.

[0008] (2) Less attention is paid to the interactive information between event pairs.

[0009] In the event causal relationship recognition task, there are usually associations between multiple causal event pairs in a sentence. Figure 1 In sentence (a), the distance between e3 and e7 is relatively far, so it is difficult for the model to directly determine the relationship between them. If the model directly determines that e7 can lead to e5, and e5 can lead to e3, two causal event pairs are formed.<e7,e5> and<e5,e3> , then according to the transitivity of causal relationship, we can further infer the causal relationship between e3 and e7<e7,e3> .

[0010] Therefore, it is necessary to introduce the interaction information between event pairs to mine the associations between them. Phu et al. used events as nodes and used toolkits to obtain information (such as syntactic dependency trees and coreferences) to construct event graphs. However, modeling with events as nodes cannot directly learn the interaction information between event pairs.

[0011] In summary, existing technologies rarely utilize event causal label information and the interactive information between event pairs. Therefore, a new method is needed to overcome this limitation in order to further improve the performance of event causal relationship identification. Summary of the invention

[0012] The defect overcome by the present invention is that the existing event causal relationship identification method rarely considers the causal label information of the event and pays less attention to the interactive information between event pairs. Therefore, in view of the above limitations, the present invention proposes an event causal relationship identification method based on information enhancement.

[0013] To achieve the above object, the present invention adopts the following technical solutions:

[0014] The present invention provides an event causal relationship identification method based on information enhancement, comprising the following steps:

[0015] Step 1: Given a sentence S = {w1,w2,...,w n}, add special marks before and after the sentence, as shown in formula (1):

[0016] I=([CLS],w1,w2,...,w n ,[SEP])(1)

[0017] Among them, [CLS] and [SEP] are special marks;

[0018] Step 2: Input the specially marked sentence into a pre-trained event causal relationship recognition model based on information enhancement, and output the predicted probability that the event pair is a causal event pair. The event causal relationship recognition model based on information enhancement includes: (1) encoding layer: based on the specially marked sentence, obtain the context-aware event representation H E (2) Event causal label perception module: By introducing the event causal label prediction task as an auxiliary task, the dependency relationship between events is mined and the event representation H is discriminated. E Whether the events in the prediction layer are causally related events, and obtain the causal labels of the event pairs: (3) Event pair interaction graph module: By constructing an interaction graph with event pairs as nodes, the graph attention mechanism is used to capture the dependencies between nodes and further identify the causal relationships between events; (4) Prediction layer: Calculate the probability that each node is a causal event pair.

[0019] Furthermore, the coding layer is specifically:

[0020] The specially marked sentence is input into the pre-trained language model BERT to obtain the contextual representation of each word, and the contextual representation of each word contained in each event in the sentence is added to obtain the context-aware event representation H. E , as shown in formulas (2) to (4):

[0021]

[0022] in, For the i-th word w i The context indicates that For the i-th event e i The context of m is the event e i The number of .

[0023] Furthermore, the event causal label perception module is specifically:

[0024] A sentence usually contains multiple events, but not all events are causally related events. The present invention designs an event causal label perception module to mine the dependencies between events and help the model pay more attention to identifying causal relationships from event pairs consisting of causally related events.

[0025] Generally speaking, events with causal relationships have strong semantic correlation, so the event representation H E Finally, the Transformer mechanism is used to capture the dependencies between events and obtain the updated event representation As shown in formula (5):

[0026]

[0027] in, For event e i Updated event representation;

[0028] The event causal label prediction task is introduced as an auxiliary task. By building a binary classifier, we can predict each event e i Probability of being a causally related event As shown in formula (6):

[0029]

[0030] in, and are learnable parameters;

[0031] Considering that if both events in an event pair are causally related events, the probability that the event pair has a causal relationship is relatively high, therefore, the present invention adopts the probability Filter out causally related events and causally unrelated events, and set the causal label embedding l to distinguish them. The event pairs consisting of causally related events are marked as cr, otherwise they are marked as cu, as shown in formula (7):

[0032]

[0033] Among them, l ij For event pairs (e i ,e j )’s causal label embedding representation.

[0034] Furthermore, the event pair interaction diagram module is specifically:

[0035] In order to capture the interaction information between event pairs, the present invention designs an event pair interaction graph module, constructs an interaction graph with event pairs as nodes, and adopts a graph attention mechanism to fuse the information of neighbor nodes.

[0036] Construction of event pair interaction graph: For the nodes of the interaction graph, events are combined in pairs to form a candidate causal event pair set EP as nodes, as shown in formula (8):

[0037] EP={ep ij}(0<i,j≤|E|,i≠j)(8);

[0038] Where E is a given event set;

[0039] Each event pair is represented by ep ij It consists of four parts, including the representation of two events and The relative position of two events in the sentence is embedded in r ij , and the causal label embedding representation l obtained from the event causal label perception module ij , as shown in formula (9):

[0040]

[0041] in, and is the context representation of two events, r ij Embedding representation for the relative position of two events in the sentence;

[0042] For the edges of the interaction graph, each event pair is considered to have a strong association with the event pair in the same row or column, such as Figure 1 As shown in (b), the event<e7,e2> The relationship can be achieved through the event pair of the same row<e7,e5> And the event pairs in the same column<e5,e2> Therefore, the present invention connects each node with the event pair in the same row or column, and adds a self-loop edge to fuse the information of the node itself;

[0043] Event pair interaction graph update: Considering that different neighbor nodes have different importance for each event pair, the present invention uses the graph attention network (GAT) to learn the interaction information between nodes. The graph attention mechanism propagates information between nodes by stacking multiple graph attention layers. In the tth graph attention layer, the representation of each node Update using formulas (10) to (12):

[0044]

[0045] in, and is the node representation of the t-1th graph attention layer, for and The joint statement Represents node ep ij The neighbor nodes of is the attention weight, indicating the node and The degree of aggregation between w t , W t , and b t are learnable parameters;

[0046] After learning the relationship between nodes by stacking T layers of graph attention, we get the updated node representation

[0047]

[0048] Furthermore, the prediction layer is specifically:

[0049] By Perform binary classification and predict the probability p of a causal event pair ij , as shown in formula (13):

[0050]

[0051] in, and are learnable parameters.

[0052] Furthermore, the method also includes training an event causal relationship recognition model based on information enhancement:

[0053] During the training process, the cross entropy function is used as the loss function for the prediction of causal event pairs, as shown in formula (14):

[0054]

[0055] Where D represents the training set, s represents the sentence in D, and E s represents the set of events contained in s, y ij Represents an event pair (e i ,e j )’s true label;

[0056] At the same time, auxiliary supervision is set for the event causal label prediction task, as shown in formula (15):

[0057]

[0058] in, Indicates event e i The probability of being predicted as a causally related event, which can be automatically obtained through the labels of the event pair without manual annotation. Indicates event e iThe true label of

[0059] The final loss of the model is obtained by weighted addition of the loss of causal event pair prediction and the loss of event causal label prediction, as shown in formula (16):

[0060] L=L ep +λ e L e (16)

[0061] Among them, λ e ∈(0,1) is the weight coefficient.

[0062] Compared with the prior art, the present invention has the following advantages:

[0063] The present invention mainly includes the following two modules: 1) Event causal label perception module, which proposes an event pair representation method based on event causal label perception. By introducing the event causal label prediction task as an auxiliary task, it mines the dependency relationship between events and determines whether the event is a causally related event. 2) Event pair interaction graph module, which proposes an event pair interaction relationship modeling method based on graph neural network. By constructing an interaction graph with event pairs as nodes, the graph attention mechanism is used to capture the dependency relationship between nodes, and further determine the causal relationship between events. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 For example analysis;

[0065] Figure 2 This is a framework diagram of the event causal relationship identification model based on information enhancement. DETAILED DESCRIPTION

[0066] In order to further illustrate the technical solution of the present invention, the present invention is further described below through embodiments.

[0067] The given sentences in this embodiment are shown in Table 1:

[0068] Table 1 Example analysis

[0069]

[0070] A method for identifying event causal relationships based on information enhancement in this embodiment includes the following steps:

[0071] Step 1: Given a sentence S = {w1,w2,...,w n}, add special marks before and after the sentence, as shown in formula (1):

[0072] I=([CLS],w1,w2,...,w n ,[SEP])(1)

[0073] Among them, [CLS] and [SEP] are special marks;

[0074] Specifically, given a sentence S = "According to relevant news reports...normal connection", first add special tags [CLS] and [SEP] before and after the sentence as BERT input I = [CLS] According to relevant news reports...normal connection [SEP], as shown in formula (1);

[0075] Step 2, input the specially marked sentence into a pre-trained event causal relationship recognition model based on information enhancement, and output the predicted probability that the event pair is a causal event pair. The event causal relationship recognition model based on information enhancement is as follows: Figure 1 As shown in Figure 1, it includes: (1) Encoding layer: Based on the sentences with special tags, context-aware event representation H is obtained. E (2) Event causal label perception module: By introducing the event causal label prediction task as an auxiliary task, the dependency relationship between events is mined and the event representation H is discriminated. E Whether the events in the prediction layer are causally related events, and obtain the causal labels of the event pairs: (3) Event pair interaction graph module: By constructing an interaction graph with event pairs as nodes, the graph attention mechanism is used to capture the dependencies between nodes and further identify the causal relationships between events; (4) Prediction layer: Calculate the probability that each node is a causal event pair.

[0076] (1) Coding layer:

[0077] The specially marked sentence is input into the pre-trained language model BERT to obtain the contextual representation of each word, and the contextual representation of each word contained in each event in the sentence is added to obtain the context-aware event representation H. E , as shown in formulas (2) to (4):

[0078]

[0079] in, For the i-th word w i The context indicates that For the i-th event e i The context of m is the event e i The number of .

[0080] (2) Event Causal Label Perception Module:

[0081] Get event representation H E Finally, the Transformer mechanism is used to capture the dependencies between events and obtain the updated event representation As shown in formula (5):

[0082]

[0083] in, For event e i Updated event representation;

[0084] The event causal label prediction task is introduced as an auxiliary task. By building a binary classifier, we can predict each event e i Probability of being a causally related event As shown in formula (6):

[0085]

[0086] in, and are learnable parameters;

[0087] According to probability We filter out causally related events “[rush, plan, connect]” and causally unrelated events “[report, use, implement, maintain]”, and set the causal label embedding l to distinguish them. We mark the event pairs consisting of causally related events as cr, otherwise as cu, as shown in formula (7):

[0088]

[0089] Among them, l ij For event pairs (e i ,e j )’s causal label embedding representation.

[0090] (3) Event-to-interaction graph module:

[0091] Construction of event pair interaction graph: For the nodes of the interaction graph, events are combined in pairs to form a candidate causal event pair set EP as nodes, as shown in formula (8):

[0092] EP={ep ij}(0<i,j≤|E|,i≠j)(8);

[0093] Where E is a given event set;

[0094] Each event pair is represented by ep ij It consists of four parts, as shown in formula (9):

[0095]

[0096] in, and is the context representation of two events, r ij Embedding representation for the relative position of two events in the sentence;

[0097] For the edges of the interaction graph, each node is connected to the event pair in the same row or column, and self-loop edges are added to integrate the information of the own node, such as Figure 1 As shown in (b);

[0098] Event-to-interaction graph update: The graph attention mechanism is used to learn the interaction information between nodes. The graph attention mechanism propagates information between nodes by stacking multiple graph attention layers. In the tth graph attention layer, the representation of each node Update using formulas (10) to (12):

[0099]

[0100] in, and is the node representation of the t-1th graph attention layer, for and The joint statement Represents node ep ij The neighbor nodes of is the attention weight, indicating the node and The degree of aggregation between w t , W t , and b t are learnable parameters;

[0101] After learning the relationship between nodes by stacking T layers of graph attention, we get the updated node representation

[0102]

[0103] (4) Prediction layer:

[0104] By Perform binary classification and predict the probability p of a causal event pair ij , as shown in formula (13):

[0105]

[0106] in, and are learnable parameters.

[0107] In addition, the method of this embodiment also includes training of an event causal relationship recognition model based on information enhancement:

[0108] During the training process, the cross entropy function is used as the loss function for the prediction of causal event pairs, as shown in formula (14):

[0109]

[0110] Where D represents the training set, s represents the sentence in D, and E s represents the set of events contained in s, y ij Represents an event pair (e i ,e j )’s true label;

[0111] At the same time, auxiliary supervision is set for the event causal label prediction task, as shown in formula (15):

[0112]

[0113] in, Indicates event e i The probability of being predicted to be a causally related event, Indicates event e i The true label of

[0114] The final loss of the model is obtained by weighted addition of the loss of causal event pair prediction and the loss of event causal label prediction, as shown in formula (16):

[0115] L=L ep +λ e L e (16)

[0116] Among them, λ e ∈(0,1) is the weight coefficient.

[0117] Example 2

[0118] This example verifies the effectiveness of the proposed model on two public datasets, EventStoryLine and Causal-TimeBank. Similar to the previous methods, the present invention performs 5-fold cross validation and 10-fold cross validation on the two datasets, and uses precision (P), recall (R) and F1 value as evaluation indicators.

[0119] The implementation of this model is based on the PyTorch framework, the operating environment is Ubuntu 22.04, and the computing device is NVIDIATITAN RTX. The detailed experimental configuration and hyperparameter settings are shown in Table 1.

[0120] Table 2 Model parameter configuration table

[0121]

[0122] This embodiment is compared with a series of existing event causality identification methods (DD, VR-C, OP, Seq, ILP, KMMG, KnowDis, CauSeRL, LSIN, LearnDA and RichGCN), and the experimental results are shown in Table 3.

[0123] Table 3 shows that the proposed method ECLEP outperforms all baselines on both datasets. Specifically, on the EventStoryLine dataset, compared with the current best method RichGCN, ECLEP improves the F1 value by 1.9%; on the Causal-TimeBank dataset, compared with the current best method CauSeRL, ECLEP improves the F1 value by 3.1%, which shows that the proposed method is effective for event causal relationship identification tasks. In addition, we observe that baseline methods usually need to use external knowledge or toolkits to improve performance, while our method achieves the best performance by mining the inherent characteristics of the data.

[0124] Table 3 Experimental results

[0125]

[0126] In order to further verify the effectiveness of each module designed by the present invention, this embodiment conducts an ablation experiment on the EventStoryLine dataset, and the results are shown in Table 4. Specifically, the present invention conducts the following experiments:

[0127] 1) -ECL and -EPI represent the removal of event causal label information and event pair interaction information, respectively. Removing any one of the information will reduce the performance of event causal relationship recognition, which shows that the two types of information introduced by the present invention are effective.

[0128] 2)-L e It means removing the auxiliary supervision for the event causal label prediction task. Removing this supervision will reduce the overall performance because it can help the model learn the representation of event causal labels more fully.

[0129] 3) -pos means removing the relative position embedding of two events in the sentence. The experimental results show that relative position information plays a certain role in causal relationship recognition, because the probability of causal relationship between events with closer distance is relatively higher than that between events with farther distance.

[0130] Table 4 Ablation experiment results

[0131]

[0132] The above shows and describes the main features and advantages of the present invention. It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes that fall within the meaning and scope of the equivalent elements of the claims are included in the present invention.

[0133] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A method for identifying event causal relationships based on information enhancement, characterized in that: The following steps are involved: Step 1: Given a sentence S = {w1,w2,...,w n }, add special marks before and after the sentence, as shown in formula (1): I=([CLS],w1,w2,...,w n ,[SEP])(1) Among them, [CLS] and [SEP] are special marks; Step 2: Input the specially marked sentence into a pre-trained event causal relationship recognition model based on information enhancement, and output the predicted probability that the event pair is a causal event pair. The event causal relationship recognition model based on information enhancement includes: (1) encoding layer: based on the specially marked sentence, obtain the context-aware event representation H E (2) Event causal label perception module: By introducing the event causal label prediction task as an auxiliary task, the dependency relationship between events is mined and the event representation H is discriminated. E Whether the events in the prediction layer are causally related events, and obtain the causal labels of the event pairs: (3) Event pair interaction graph module: By constructing an interaction graph with event pairs as nodes, the graph attention mechanism is used to capture the dependencies between nodes and further identify the causal relationships between events; (4) Prediction layer: Calculate the probability that each node is a causal event pair.

2. The event causal relationship identification method based on information enhancement according to claim 1 is characterized in that: The coding layer is specifically: The specially marked sentence is input into the pre-trained language model BERT to obtain the contextual representation of each word, and the contextual representation of each word contained in each event in the sentence is added to obtain the context-aware event representation H. E , as shown in formulas (2) to (4): in, For the i-th word w i The context indicates that For the i-th event e i The context of m is the event e i The number of .

3. The event causal relationship identification method based on information enhancement according to claim 2 is characterized in that: The event causal label perception module is specifically: Get event representation H E Finally, the Transformer mechanism is used to capture the dependencies between events and obtain the updated event representation As shown in formula (5): in, For event e i Updated event representation; The event causal label prediction task is introduced as an auxiliary task. By building a binary classifier, we can predict each event e i Probability of being a causally related event As shown in formula (6): in, and are learnable parameters; According to probability Filter out causally related events and causally unrelated events, and set the causal label embedding l to distinguish them. The event pairs consisting of causally related events are marked as cr, otherwise they are marked as cu, as shown in formula (7): Among them, l ij For event pairs (e i ,e j )’s causal label embedding representation.

4. The event causal relationship identification method based on information enhancement according to claim 3 is characterized in that: The event pair interaction diagram module is specifically: Construction of event pair interaction graph: For the nodes of the interaction graph, events are combined in pairs to form a candidate causal event pair set EP as nodes, as shown in formula (8): EP={ep ij }(0<i,j≤|E|,i≠j)(8); Where E is a given event set; Each event pair is represented by ep ij It consists of four parts, as shown in formula (9): in, and is the context representation of two events, r ij Embedding representation for the relative position of two events in the sentence; For the edges of the interaction graph, each node is connected to the event pair in the same row or column, and self-loop edges are added to integrate the information of the own node; Events update the interaction graph: The graph attention mechanism is used to learn the interaction information between nodes. The graph attention mechanism propagates information between nodes by stacking multiple graph attention layers. In the tth graph attention layer, the representation of each node Update using formulas (10) to (12): in, and is the node representation of the t-1th graph attention layer, for and The joint statement Represents node ep ij The neighbor nodes of is the attention weight, indicating the node and The degree of aggregation between w t , W t , and b t are learnable parameters; After learning the relationship between nodes by stacking T layers of graph attention, we get the updated node representation 5. The event causal relationship identification method based on information enhancement according to claim 4 is characterized in that: The prediction layer is specifically: By Perform binary classification and predict the probability p of a causal event pair ij , as shown in formula (13): in, and are learnable parameters.

6. The event causal relationship identification method based on information enhancement according to any one of claims 1 to 5, characterized in that: The method also includes training an event causal relationship recognition model based on information enhancement: During the training process, the cross entropy function is used as the loss function for the prediction of causal event pairs, as shown in formula (14): Where D represents the training set, s represents the sentence in D, and E s represents the set of events contained in s, y ij Represents an event pair (e i ,e j )’s true label; At the same time, auxiliary supervision is set for the event causal label prediction task, as shown in formula (15): in, Indicates event e i The probability of being predicted to be a causally related event, Indicates event e i The true label of The final loss of the model is obtained by weighted addition of the loss of causal event pair prediction and the loss of event causal label prediction, as shown in formula (16): L=L ep +λ e L e (16) Among them, λ e ∈(0,1) is the weight coefficient.

Images