An event detection and extraction method and system based on entity and trigger word fusion recognition

Abstract

The present invention discloses an event detection and extraction method and system based on entity and trigger word fusion recognition, the method comprising: extracting trigger words from event text data; mapping event text to vector space based on a word embedding model, and constructing an event text classifier to measure similarity between all words with verb part of speech and trigger words in the event text; constructing an event detection model to obtain event categories in event text data; identifying key element entities of events; constructing a dependency syntax tree to represent the syntactic dependency of sentences; obtaining the score of the dependency syntax tree; inputting the preprocessed word vector and the part of speech vector representation obtained through supervised learning into a role labeling model for fusion to obtain a vector representation with part of speech features, and obtaining the final semantic role label of the word through conditional random field processing after training. The present invention realizes accurate event detection based on entities and drives high-quality extraction of events.

Description

Technical Field

[0001] The present invention relates to the technical field of event detection and extraction, and in particular to an event detection and extraction method and system based on entity and trigger word fusion recognition. Background Art

[0002] The rise of social media has changed the way people obtain information. It is particularly important to extract structured event information with important semantics and value from massive social networks, and aggregate this event information according to the structure of event trigger words, event types, event participants (arguments), and argument roles to support decision-making, trend analysis, and knowledge discovery. Major public events have the characteristics of high timeliness, high diffusion, high impact, and cross-modality. For the topic events contained in massive cross-modal news and social media, studying the rapid detection of events and the extraction of key elements, and establishing a representation model for events to provide technical support for subsequent downstream tasks has become a hot topic in this field, with broad application prospects and value.

[0003] However, the information in mainstream social networks is currently diverse and has a wide range of sources, which leads to incomplete feature extraction. At the same time, it faces challenges such as data sparsity and diversity, context dependency, and real-time and dynamic nature. Early event extraction methods mainly relied on manually designed pattern matching and rule-based methods, which usually require a lot of manual intervention and domain knowledge and cannot adapt to the current event extraction task requirements. Especially in the context of large language models, some semantic interference will cause the performance of text-based detection models to drop significantly, so there is an urgent need to develop more stable detection methods. Summary of the invention

[0004] In view of the problems existing in the detection and extraction tasks of social network information events, the present invention provides an event detection and extraction method and system based on the fusion recognition of entities and trigger words. By jointly extracting entities and trigger words existing in sentences, rich semantic representation of events in sentences is achieved, thereby improving the accurate extraction of cross-modal data based on event element features.

[0005] The present invention discloses an event detection and extraction method based on entity and trigger word fusion recognition, which comprises:

[0006] Step 1: Preprocess the event text data and extract the trigger words in the event text data;

[0007] Step 2: Map the event text to the vector space based on the word embedding model, and construct an event text classifier to measure the similarity between all the words with verb part of speech and the trigger words in the event text, so as to realize the classification and detection of event sentences and event categories;

[0008] Step 3: Build an event detection model to obtain event categories in event text data;

[0009] Step 4: Based on the acquired event categories, use named entity recognition technology to identify the key element entities of the event;

[0010] Step 5: Construct a dependency syntax tree to represent the syntactic dependencies of the sentence;

[0011] Step 6: Convert the dependency syntax tree to a dependency graph, and then convert the dependency graph to a dependency subgraph. The score of the dependency syntax tree is equal to the sum of the scores of each subordinate dependency subgraph.

[0012] Step 7: Use the network layer of the long short-term memory network, conditional random fields and post-processing to form a role labeling model; input the corpus into the role labeling model for training to obtain the initial semantic role label, add the learned part-of-speech feature information, and use the part-of-speech tag as the target vector to obtain the vector representation of the part-of-speech tag through supervised learning; input the preprocessed word vector and the part-of-speech vector representation learned through supervised learning into the role labeling model for fusion to obtain the vector representation with part-of-speech features, input the obtained vector representation into the role labeling model for training, and after training, process it through the conditional random field to obtain the final semantic role label of the word.

[0013] Furthermore, the step 1 comprises:

[0014] Search for words before the pronoun in the event text, analyze the contextual relationship between the pronoun and the antecedent, and combine the pronoun and the antecedent into a pronoun chain based on the contextual relationship. If multiple pronouns point to the same antecedent, they will be linked.

[0015] Furthermore, before step 1, the method further includes:

[0016] Based on the event category labels in the event extraction dataset, event-related data is collected from social media, and an event classification dataset is formed through data annotation. The event classification dataset is divided into training set, validation set and test set; the training set is used to train the proposed event detection and classification model; the test set is used to test the accuracy of the event detection and classification model; the validation set is used to determine the network structure and adjust the model's hyperparameters to obtain the current optimal model.

[0017] Furthermore, the step 2 comprises:

[0018] A word embedding model is used to map different words into vector space, and the distance between word vectors is calculated to measure the similarity between words. When the similarity between verbs in the event text and trigger words of multiple categories exceeds the specified threshold, the category with the highest trigger word similarity is selected as the candidate event category.

[0019] Furthermore, the step 3 comprises:

[0020] When analyzing event text data, each text that has been segmented is considered as a sequence, and the sequence is recorded as , also called observation sequence; by analyzing the text data to extract relevant features, and feature labeling each element in the sequence; the labeling result of the output sequence corresponding to the input sequence is recorded as , and each random variable All satisfy the Markov property, then the conditional probability As the output the probability distribution of the labeling results.

[0021] The constructed event detection model is trained through the time dataset to obtain the event categories in the event text data.

[0022] Furthermore, the conditional probability The calculation formula is:

[0023] (1)

[0024] (2)

[0025] In the formula, and is the characteristic function; and is the corresponding weight; is the normalization factor, and the sum is performed on all possible output sequences. First, all the label numbers Sum all the characteristic functions sum; is a characteristic function defined on the edge, called the transfer feature, which depends on the current and previous positions; It is a characteristic function defined on the node, called the state characteristic, which depends on the current position; and They all depend on the position and are local characteristic functions.

[0026] Furthermore, the step 4 comprises:

[0027] A semantic dictionary is introduced to annotate and classify the semantic roles in the event text, and entity disambiguation is combined to realize named entity recognition of the event text to identify the key element entities of the event; entity disambiguation is used to process the identified entity information.

[0028] Furthermore, the step 5 comprises:

[0029] Step 51: parse the domain event data using dependency syntactic analysis technology and generate a syntactic analysis tree to obtain the dependency syntactic relationship of the sentence, match the syntactic dependency pattern according to the key branches of the syntactic dependency tree, that is, the subject-predicate relationship or the verb-object relationship in the sentence, extract and fill the slots of the corresponding event element tuple according to the matched syntactic dependency pattern, and use the subject, predicate, object, time, and place as the sentence trunk of the meta-event tuple, and use the remaining syntactic relationship words as additional attributes of the core words, and finally obtain a single or multiple meta-event tuples; the remaining syntactic relationship words include adverbials and attributives;

[0030] Step 52: Filling semantic roles for the extracted entity elements, wherein the semantic roles are used to identify event elements in the event sentence, and then filling the event elements into corresponding semantic roles according to the event template to achieve the extraction of semantic roles;

[0031] Step 53: Extract key elements of the event to form an event quintuple, which includes event type, event sentence, time, place and person;

[0032] Step 54: Convert the sentence into a dependency syntax tree, and the dependency syntax tree is represented by a directed complete graph, which represents the structure of syntactic dependency by finding a maximum spanning tree.

[0033] Furthermore, the step 7 comprises:

[0034] Step 71: In the process of role labeling model training, firstly, the corpus is vectorized and preprocessed to generate word vectors and use them as network layer inputs of the semantic role label learning model, and the feature representation of the input word vector sequence is learned through the network layer;

[0035] Step 72: using the output value calculated by the LSTM network in the semantic role label learning model, that is, the corpus feature, as the input data of the conditional random field in the semantic role label learning model, and obtaining the initial semantic role label after the conditional random field and post-processing; using IOBES to label the semantic role;

[0036] Step 73: vectorize the corpus to obtain word vectors, input the word vectors into the part-of-speech learning model, use the part-of-speech tag as the target vector, and obtain the vector representation of the part-of-speech tag through supervised learning;

[0037] Step 74: Input the word vector obtained by one-hot encoding vectorization and the part-of-speech vector representation obtained by supervised learning into the semantic role label learning model for fusion to obtain a vector representation with part-of-speech features, and input the fused vector representation into the role labeling model for training. After training, the final semantic role label of the word is obtained through conditional random field processing.

[0038] The present invention also discloses an event detection and extraction system based on entity and trigger word fusion recognition, which implements any of the above-mentioned event detection and extraction methods based on entity and trigger word fusion recognition, and comprises:

[0039] An extraction module is used to pre-process the event text data and extract trigger words from the event text data;

[0040] The measurement module is used to map the event text to the vector space based on the word embedding model, and construct an event text classifier to measure the similarity between all the words with verb parts of speech and trigger words in the event text, so as to realize the classification and detection of event sentences and event categories;

[0041] A building module is used to build an event detection model to obtain event categories in event text data;

[0042] The recognition module is used to identify key element entities of the event based on the acquired event category by using named entity recognition technology;

[0043] The representation module is used to construct a dependency syntactic tree to represent the syntactic dependencies of a sentence;

[0044] A conversion module, used to convert the dependency syntax tree into a dependency graph, and then convert the dependency graph into a dependency subgraph, the score of the dependency syntax tree is equal to the sum of the scores of each subordinate dependency subgraph;

[0045] The semantic role labeling module is used to form a role labeling model using a long short-term memory network layer, a conditional random field and post-processing; the corpus is input into the role labeling model for training to obtain the initial semantic role label, the learned part-of-speech feature information is added, and the part-of-speech tag is used as the target vector to obtain the vector representation of the part-of-speech tag through supervised learning; the word vector obtained by one-hot encoding vectorization and the part-of-speech vector representation obtained by supervised learning are input into the role labeling model for fusion to obtain a vector representation with part-of-speech features, the obtained vector representation is input into the role labeling model for training, and after training, the final semantic role label of the word is obtained through conditional random field processing.

[0046] Due to the adoption of the above technical solution, the present invention has the following advantages:

[0047] 1. The present invention combines the two subtasks of network event detection and extraction in a social network environment, realizes accurate detection and high-quality extraction of network events based on dependency syntactic analysis, and overcomes the problems of weak stability and low accuracy of existing methods.

[0048] 2. For the topic events contained in massive cross-modal news and social media, the present invention completes the classification and identification of events by recognizing entity words in sentences, and further forms a new strategy for event extraction based on dependency syntactic analysis.

[0049] 3. In order to solve the problem of incomplete extraction of event elements due to the diverse types and wide sources of event data, the present invention adopts feature attributes such as the part of speech of words and adjacent words, whether they are entities, syntactic dependencies, etc., and adopts a conditional random field model to identify event trigger words. The trigger words are classified by a classifier to realize the detection and classification of major events.

[0050] 4. For the multiple key elements of events in domestic and foreign news and social media content, based on the recognition of entity information such as event time, people, and place through named entity recognition technology, the strategy of expanding event key sentences with trigger words is implemented using dependency syntax and semantic role filling methods, thereby realizing the extraction of the entire event.

[0051] 5. Experimental results on real data sets show that the present invention is stable and effective in the task of identifying and extracting social network events, especially under extreme conditions such as lack of data volume and missing data labels, and can still achieve stable detection results. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0053] Figure 1 A flow chart of an event detection and extraction method based on entity and trigger word fusion recognition according to an embodiment of the present invention;

[0054] Figure 2 A schematic diagram of an event detection process based on trigger word recognition according to an embodiment of the present invention;

[0055] Figure 3 A schematic diagram of a flow chart of an event extraction method based on dependency syntax and semantic role filling according to an embodiment of the present invention;

[0056] Figure 4The figure is a flowchart of a semantic role labeling algorithm based on LSTM-CRF of part-of-speech features according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] The present invention is further described in conjunction with the accompanying drawings and embodiments, and the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by ordinary technicians in this field should fall within the scope of protection of the embodiments of the present invention.

[0058] See also Figure 1 The present invention provides an embodiment of an event detection and extraction method based on entity and trigger word fusion recognition, which includes:

[0059] Step 1: Preprocess the event text data and extract the trigger words in the event text data;

[0060] In one embodiment of the present application, event text data is preprocessed using strategies such as event sentence decomposition, reference resolution, and data format conversion. The goal of reference resolution is to determine the specific object that a noun or pronoun in the text refers to. The regular expression method is mainly used to define a set of rules using regular expressions to identify different references that may represent the same entity in the text. For example, if there are two noun phrases "president" and "he" in the text, define a regular expression to identify these two phrases and infer that they refer to the same person. The specific steps are:

[0061] Find antecedents: Search for words before the pronoun in the text, which may be the objects referred to by the pronoun; Context analysis: Analyze the contextual relationship between the pronoun and possible antecedents, including their distance, semantic content, sentence structure, etc.; Construct a pronoun chain: Based on the determined contextual relationship, combine them into a pronoun chain. If multiple pronouns point to the same antecedent, link them together; Evaluation and optimization: Improve the performance of pronoun resolution.

[0062] Step 2: Map the event text to the vector space based on the word embedding model, and construct an event text classifier to measure the similarity between all the words with verb part of speech and the trigger words in the event text, so as to realize the classification and detection of event sentences and event categories;

[0063] In one embodiment of the present application, event trigger words are important components of event descriptions, and generally, event trigger words are verbs. Therefore, the present invention uses a conditional random field model to construct a feature extractor based on typical feature attributes such as the part of speech of words and adjacent words in the event text, whether they are entities, and syntactic dependencies, so as to identify trigger words in event text data;

[0064] Based on the word embedding model, the event text is mapped to the vector space, and an event text classifier is constructed to measure the similarity between all verbs and trigger words in the event text, so as to realize the classification and detection of event sentences and event categories. Among them, Word2vec is used to map different words to the vector space, and the distance between word vectors is calculated to measure the similarity between words. When the similarity between verbs in the event text and trigger words of multiple categories exceeds a specific threshold, the category with the highest trigger word similarity is selected as the candidate event category. The event detection process based on trigger word recognition is as follows: Figure 2 shown.

[0065] Step 3: Build an event detection model to obtain event categories in event text data;

[0066] In one embodiment of the present application, when analyzing text data, each text that has been segmented is first regarded as a sequence, and the sequence is recorded as , also called observation sequence. The text data is analyzed to extract relevant features, and each element in the sequence is annotated. The annotation result of the output sequence corresponding to the input sequence is recorded as , and each random variable All satisfy the Markov property, then the conditional probability As the probability distribution of the output labeling result. Its calculation formula is:

[0067] (1)

[0068] (2)

[0069] In the formula, and is the characteristic function, and is the corresponding weight. is the normalization factor, and the sum is performed on all possible output sequences. First, all the labeled sequences Sum all the characteristic functions Sum. is a characteristic function defined on the edge, called the transfer feature, which depends on the current and previous positions. It is a characteristic function defined on the node, called state characteristic, which depends on the current position. and They all depend on the position and are local characteristic functions. Usually, the characteristic function and The value is 1 or 0. The value is 1 when the characteristic condition is met, otherwise it is 0.

[0070] Based on the event detection model constructed above, the model training is performed on the training set of the own data set to obtain the event category probability in the text data. The difference between the actual value and the theoretical value is calculated on the test set of the own data set to test the accuracy of the event detection and classification model.

[0071] Step 4: Based on the acquired event categories, use named entity recognition technology to identify the key element entities of the event;

[0072] In one embodiment of the present application, a semantic dictionary is introduced to annotate and classify the semantic roles in the event text, and entity disambiguation is combined to realize named entity recognition of the event text, and identify the key element entities of the event, including entity information such as time, name, place name, and organization name; generally, due to the high accuracy of time entity extraction, place names and personal names may have problems such as unclear reference and ambiguity, and the entity disambiguation strategy is used to optimize the identified entity information;

[0073] Step 5: Construct a dependency syntax tree to represent the syntactic dependencies of the sentence;

[0074] In one embodiment of the present application, the domain event data is parsed using dependency syntactic analysis technology and a syntactic analysis tree is generated to obtain the dependency syntactic relationship of the sentence. The syntactic dependency pattern is matched according to the key branches of the syntactic dependency tree, that is, the subject-predicate relationship (SBV) or the verb-object relationship (VOB) in the sentence. According to the matched syntactic dependency pattern, the slots of the corresponding event element tuples are extracted and filled, and the subject, predicate, object, time, and place are used as the sentence trunk of the meta-event tuple, and the remaining syntactic relationship words such as adverbials and attributives are used as additional attributes of the core words, and finally a single or multiple meta-event tuples are obtained;

[0075] Filling semantic roles for the extracted entity elements, where semantic roles mainly identify event elements in event sentences, and then filling event elements into corresponding semantic roles according to event templates to achieve semantic role extraction;

[0076] The key elements of an event are extracted to form a five-tuple of events, including event type, event sentence, time, location, and person. The event extraction technology process based on dependency syntax and semantic role filling is as follows: Figure 3 As shown;

[0077] The sentence is converted into a dependency syntactic tree, and the tree can be represented by a directed complete graph, and the structure of syntactic dependency is represented by finding a maximum spanning tree in this directed complete graph. That is, the process of building the dependency structure can be converted into a process of finding the maximum spanning tree, and the final result obtained through the search is a global optimal solution. The dependency analysis method based on graph decoding regards dependency analysis as a search problem of finding the maximum spanning tree in a directed complete graph. The score of the dependency tree found is ultimately determined by the cumulative score of its subtrees. For a complete sentence that complies with grammatical rules, Included Vocabulary: ,in The root node ROOT is manually annotated, according to the sentence The combination of all dependency trees generated is , the searched maximum spanning tree can be expressed as follows:

[0078] (3)

[0079] in, is a set of dependency trees Any dependency tree in is the scoring function in the scoring model. The set of all calculated dependency trees The dependency syntax tree with the highest score.

[0080] Step 6: Convert the dependency syntax tree to a dependency graph, and then convert the dependency graph to a dependency subgraph. The score of the dependency syntax tree is equal to the sum of the scores of each subordinate dependency subgraph.

[0081] In one embodiment of the present application, a complete dependency tree is converted into a dependency graph, and then the dependency graph is converted into a dependency subgraph. The score of the final dependency tree is equal to the sum of the scores of each subordinate subgraph, that is,

[0082] (4)

[0083] In the above formula, It is a sentence The generated dependency graph is finally transformed into a dependency subgraph. is any dependency tree. The advantage of converting the score of the dependency tree into the score of the dependency subgraph is that it can simplify the process of searching for the global optimal solution. The simplest subgraph decomposition method is the first-order decomposition method. The key point of the method is to convert the dependency graph into a dependency subgraph, and then convert each dependency subgraph into multiple independent dependency edges, so the final score of the dependency tree is equal to the sum of the scores of each dependency edge, that is,

[0084] (5)

[0085] In the above formula, As the core word, As a modifier, There is a directed dependency edge between the first two, and this equation represents the dependency relationship formed by the above three. In general, dynamic programming is usually used to obtain the optimal dependency tree.

[0086] Step 7: Use the network layer of the Long Short-Term Memory (LSTM), conditional random fields, and post-processing to form a role labeling model; input the corpus into the role labeling model for training to obtain the initial semantic role label, add the learned part-of-speech feature information, and use the part-of-speech tag as the target vector to obtain the vector representation of the part-of-speech tag through supervised learning; input the word vector obtained by one-hot encoding vectorization and the part-of-speech vector representation obtained by supervised learning into the role labeling model for fusion to obtain a vector representation with part-of-speech features, input the obtained vector representation into the role labeling model for training, and after training, process it through the conditional random field to obtain the final semantic role label of the word.

[0087] In one embodiment of the present application, see Figure 4 In the process of model training, the corpus must first be vectorized and preprocessed to generate word vectors and use them as the network layer input of the semantic role label learning model SRL-LSTM, through which the feature representation of the input sequence is learned. t represents the input data at the current time t, { , ,⋯, } represents the vector representation obtained after preprocessing the input data at the current time t, n represents the dimension of the vector, the network output value y_t at time t is used to represent, h1, h2, ... hH represent the hidden layer parameters, H represents the number of nodes in the hidden layer, and the input layer, hidden layer, and output layer constitute the LSTM network. a_p^pos represents the vector containing part-of-speech features obtained after processing by the POS-BiLSTM network layer, a_p^srl represents the vector containing semantic features obtained after processing by the SRL-BiLSTM network layer, and Loss(1) and Loss(2) are loss functions. The output value calculated by LSTM, that is, the corpus feature, is used as the input data of the CRF layer, and the initial semantic role label is obtained after the CRF and post-processing steps. Among them, in the actual annotation process, there may be multiple words with the same semantic role label. In this case, if the traditional annotation strategy is used, the semantic role cannot be accurately labeled, so the IOBES annotation strategy is used to improve the annotation accuracy. Among them, B represents the beginning of a compound word, I represents the middle of a compound word, E represents the end of a compound word, S represents a simple word, and O represents other words other than the semantic role.

[0088] The word vectors are then input into the part-of-speech learning model POS-LSTM by introducing part-of-speech feature information, and the vector representation of the part-of-speech tag is obtained through supervised learning with the part-of-speech tag as the target vector. The preprocessed word vector and the part-of-speech vector representation learned through supervised learning are input into the semantic role label learning model SRL-LSTM for fusion to obtain a vector representation with part-of-speech features. The obtained new vector is input into the model for training, and after training, it is processed by CRF to obtain the final semantic role label of the word.

[0089] In one embodiment of the present application, before step 1, the method further includes:

[0090] Based on the event category labels in the event extraction dataset, event-related data is collected from social media, and an event classification dataset is formed through data annotation. The event classification dataset is divided into training set, validation set and test set; the training set is used to train the proposed event detection and classification model; the test set is used to test the accuracy of the event detection and classification model; the validation set is used to determine the network structure and adjust the model's hyperparameters to obtain the current optimal model.

[0091] Optionally, based on the event category labels in the ACE2005 event extraction dataset, event-related data is collected from social media such as Facebook and Twitter, and a self-owned event classification dataset is formed through data annotation, and the dataset is divided into a training set, a validation set, and a test set. Among them, the training set contains relatively clean data and is only used to train the proposed event detection and classification model; the test set contains more complex real event data, which is used to test the accuracy of the event detection and classification model; the data type of the validation set is similar to that of the test set, which is used to determine the network structure and adjust the hyperparameters of the model to obtain the current optimal model. The ACE2005 dataset is a mixed genre text data in English, Arabic, and Chinese, and is annotated with entities, relations, and events. This dataset can be used for three research target tasks: event detection and representation, entity detection and tracking, and relationship detection and representation. It contains 8 major categories of events (such as conflict, justice, transactions, business, etc.), with a total of 33 sub-categories of events.

[0092] Experimental verification:

[0093] The model is trained on the training set of the ACE2005 event extraction dataset, which can extract the key elements in the event text and form a five-tuple of elements, namely [event type; event sentence; time; location; person / organization]. The accuracy of the model is tested by calculating the difference between the actual value and the theoretical value on the test set. Figure 2 , Figure 3 and Figure 4 The experimental architecture demonstrates the excellent performance of the present invention in event recognition and extraction tasks.

[0094] The method of the present invention is applicable to social network environments, and combines the two subtasks of network event detection and extraction, thereby achieving accurate entity-based network event detection and leading to high-quality event extraction, thereby being widely used in application scenarios such as Weibo, Twitter, and news websites.

[0095] The present invention also provides an embodiment of an event detection and extraction system based on entity and trigger word fusion recognition, which implements the event detection and extraction method based on entity and trigger word fusion recognition described in the above embodiment, and includes:

[0096] An extraction module is used to pre-process the event text data and extract trigger words from the event text data;

[0097] The measurement module is used to map the event text to the vector space based on the word embedding model, and construct an event text classifier to measure the similarity between all the words with verb parts of speech and trigger words in the event text, so as to realize the classification and detection of event sentences and event categories;

[0098] A building module is used to build an event detection model to obtain event categories in event text data;

[0099] The recognition module is used to identify key element entities of the event based on the acquired event category by using named entity recognition technology;

[0100] The representation module is used to construct a dependency syntactic tree to represent the syntactic dependencies of a sentence;

[0101] A conversion module, used to convert the dependency syntax tree into a dependency graph, and then convert the dependency graph into a dependency subgraph, the score of the dependency syntax tree is equal to the sum of the scores of each subordinate dependency subgraph;

[0102] The semantic role labeling module is used to form a role labeling model using the network layer of the long short-term memory network, conditional random fields and post-processing; the corpus is input into the role labeling model for training to obtain the initial semantic role label, the learned part-of-speech feature information is added, and the vector representation of the part-of-speech tag is obtained through supervised learning with the part-of-speech tag as the target vector; the preprocessed word vector and the part-of-speech vector representation learned through supervised learning are input into the role labeling model for fusion to obtain a vector representation with part-of-speech features, and the obtained vector representation is input into the role labeling model for training. After training, the final semantic role label of the word is obtained through conditional random field processing.

[0103] The present application also provides a more specific embodiment of an event detection and extraction system based on entity and trigger word fusion recognition, which includes: an event detection module based on trigger word recognition and an event extraction module based on dependency syntax and semantic role;

[0104] The event detection module based on trigger word recognition is used to:

[0105] Event text data is preprocessed using strategies such as event sentence decomposition, coreference resolution, and data format conversion, and a set of rules are defined using regular expressions to identify different references that may represent the same entity in the text, thereby achieving coreference resolution, that is, determining the specific object pointed to by the noun or pronoun in the text. Secondly, based on typical feature attributes such as the part of speech of words and adjacent words in the event text, whether they are entities, and syntactic dependencies, a feature extractor is constructed using a conditional random field model to identify trigger words in event text data, where event trigger words are an important component of event descriptions. In general, event trigger words are verbs. Finally, based on the word embedding model, the event text is mapped to a vector space, and an event text classifier is constructed to measure the similarity between all words with verb parts of speech and trigger words in the event text, thereby achieving classification and detection of event sentences and event categories.

[0106] The event extraction module based on dependency syntax and semantic roles is used for:

[0107] On the basis of using event detection and classification models to obtain event category information, a semantic dictionary is further introduced to annotate and classify the semantic roles in event texts, and named entity recognition technology is used to identify the key element entities of events. Secondly, the dependency syntactic analysis technology is used to parse the domain event data and generate a syntactic analysis tree to obtain the dependency syntactic relationship of the sentence. The syntactic dependency pattern is matched according to the key branches of the syntactic dependency tree, such as the subject-predicate relationship (SBV) or the verb-object relationship (VOB). According to the matched syntactic dependency pattern, the slots of the corresponding event element tuple are extracted and filled, and the subject, predicate, object, time, and place are used as the sentence trunk of the meta-event tuple, while the remaining adverbials, attributives and other syntactic relationship words are used as additional attributes of the core words. Finally, a single or multiple meta-event tuples are obtained; then, the semantic roles are filled for the extracted entity elements, where the semantic roles are mainly to identify the event elements in the event sentence, and then the event elements are filled into the corresponding semantic roles according to the event template to realize the extraction of the semantic roles; finally, the extraction of the key elements of the event is completed, and the event quintuple is formed, including event type, event sentence, time, place, and person.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An event detection and extraction method based on entity and trigger word fusion recognition, characterized in that: include: Step 1: Preprocess the event text data and extract the trigger words in the event text data; Step 2: Map the event text to the vector space based on the word embedding model, and construct an event text classifier to measure the similarity between all the words with verb part of speech and the trigger words in the event text, so as to realize the classification and detection of event sentences and event categories; Step 3: Build an event detection model to obtain event categories in event text data; Step 4: Based on the acquired event categories, use named entity recognition technology to identify the key element entities of the event; Step 5: Construct a dependency syntax tree to represent the syntactic dependencies of the sentence; Step 6: Convert the dependency syntax tree to a dependency graph, and then convert the dependency graph to a dependency subgraph. The score of the dependency syntax tree is equal to the sum of the scores of each subordinate dependency subgraph. Step 7: Use the long short-term memory network layer, conditional random fields and post-processing to form a role labeling model; The corpus is input into the role labeling model for training to obtain the initial semantic role label, and the part-of-speech feature information is added. The part-of-speech label is used as the target vector to obtain the vector representation of the part-of-speech label through supervised learning; the word vector obtained by one-hot encoding vectorization and the part-of-speech vector representation obtained by supervised learning are input into the role labeling model for fusion to obtain the vector representation with part-of-speech features, and the obtained vector representation is input into the role labeling model for training. After training, the final semantic role label of the word is obtained through conditional random field processing; The step 3 comprises: When analyzing event text data, each text that has been segmented is considered as a sequence, and the sequence is recorded as , also called the observation sequence, is the i-th sequence; the text data is analyzed to extract relevant features, and each element in the sequence is annotated; the annotated result of the output sequence corresponding to the input sequence is recorded as , and each random variable All satisfy the Markov property, then the conditional probability As the probability distribution of the output labeling results; The constructed event detection model is trained through the time dataset to obtain the event categories in the event text data; Conditional Probability The calculation formula is: （1） （2） In the formula, and is the characteristic function; and is the corresponding weight; is the normalization factor, and the sum is performed on all possible output sequences. First, all the label numbers Sum all the characteristic functions sum; is a characteristic function defined on the edge, called the transfer feature, which depends on the current and previous positions; It is a characteristic function defined on the node, called the state characteristic, which depends on the current position; and They all depend on the position and are local characteristic functions; The step 5 comprises: Step 51: parse the domain event data using dependency syntactic analysis technology and generate a syntactic analysis tree to obtain the dependency syntactic relationship of the sentence, match the syntactic dependency pattern according to the key branches of the syntactic dependency tree, that is, the subject-predicate relationship or the verb-object relationship in the sentence, extract and fill the slots of the corresponding event element tuple according to the matched syntactic dependency pattern, and use the subject, predicate, object, time, and place as the sentence trunk of the meta-event tuple, and use the remaining syntactic relationship words as additional attributes of the core words, and finally obtain a single or multiple meta-event tuples; the remaining syntactic relationship words include adverbials and attributives; Step 52: Filling semantic roles for the extracted entity elements, wherein the semantic roles are used to identify event elements in the event sentence, and then filling the event elements into corresponding semantic roles according to the event template to achieve the extraction of semantic roles; Step 53: Extract key elements of the event to form an event quintuple, which includes event type, event sentence, time, place and person; Step 54: Convert the sentence into a dependency syntax tree, and the dependency syntax tree is represented by a directed complete graph, which represents the structure of syntactic dependency by finding a maximum spanning tree.

2. The event detection and extraction method based on entity and trigger word fusion recognition according to claim 1 is characterized in that: The step 1 comprises: Search for words before the pronoun in the event text, analyze the contextual relationship between the pronoun and the antecedent, and combine the pronoun and the antecedent into a pronoun chain based on the contextual relationship. If multiple pronouns point to the same antecedent, they will be linked.

3. The event detection and extraction method based on entity and trigger word fusion recognition according to claim 1 is characterized in that: Before step 1, the method further includes: Based on the event category labels in the event extraction dataset, event-related data is collected from social media, and an event classification dataset is formed through data annotation. The event classification dataset is divided into training set, validation set and test set; the training set is used to train the proposed event detection and classification model; the test set is used to test the accuracy of the event detection and classification model; the validation set is used to determine the network structure and adjust the model's hyperparameters to obtain the current optimal model.

4. The event detection and extraction method based on entity and trigger word fusion recognition according to claim 1 is characterized in that: The step 2 comprises: A word embedding model is used to map different words into vector space, and the distance between word vectors is calculated to measure the similarity between words. When the similarity between verbs in the event text and trigger words of multiple categories exceeds the specified threshold, the category with the highest trigger word similarity is selected as the candidate event category.

5. The event detection and extraction method based on entity and trigger word fusion recognition according to claim 1 is characterized in that: The step 4 comprises: A semantic dictionary is introduced to annotate and classify the semantic roles in the event text, and entity disambiguation is combined to realize named entity recognition of the event text to identify the key element entities of the event; entity disambiguation is used to process the identified entity information.

6. The event detection and extraction method based on entity and trigger word fusion recognition according to claim 1 is characterized in that: The step 7 comprises: Step 71: In the process of role labeling model training, firstly, the corpus is vectorized and preprocessed to generate word vectors and use them as network layer inputs of the semantic role label learning model, and the feature representation of the input word vector sequence is learned through the network layer; Step 72: using the output value calculated by the LSTM network in the semantic role label learning model, that is, the corpus feature, as the input data of the conditional random field in the semantic role label learning model, and obtaining the initial semantic role label after the conditional random field and post-processing; using IOBES to label the semantic role; Step 73: vectorize the corpus to obtain word vectors, input the word vectors into the part-of-speech learning model, use the part-of-speech tag as the target vector, and obtain the vector representation of the part-of-speech tag through supervised learning; Step 74: Input the word vector obtained by one-hot encoding vectorization and the part-of-speech vector representation obtained by supervised learning into the semantic role label learning model for fusion to obtain a vector representation with part-of-speech features, and input the fused vector representation into the role labeling model for training. After training, the final semantic role label of the word is obtained through conditional random field processing.

7. An event detection and extraction system based on entity and trigger word fusion recognition, implementing the event detection and extraction method based on entity and trigger word fusion recognition according to any one of claims 1 to 6, characterized in that: include: An extraction module is used to pre-process the event text data and extract trigger words from the event text data; The measurement module is used to map the event text to the vector space based on the word embedding model, and construct an event text classifier to measure the similarity between all the words with verb parts of speech and trigger words in the event text, so as to realize the classification and detection of event sentences and event categories; A building module is used to build an event detection model to obtain event categories in event text data; The recognition module is used to identify key element entities of the event based on the acquired event category by using named entity recognition technology; The representation module is used to construct a dependency syntactic tree to represent the syntactic dependencies of a sentence; A conversion module, used to convert the dependency syntax tree into a dependency graph, and then convert the dependency graph into a dependency subgraph, the score of the dependency syntax tree is equal to the sum of the scores of each subordinate dependency subgraph; The semantic role labeling module is used to form a role labeling model using the network layer of the long short-term memory network, conditional random fields and post-processing; the corpus is input into the role labeling model for training to obtain the initial semantic role label, the learned part-of-speech feature information is added, and the vector representation of the part-of-speech tag is obtained through supervised learning with the part-of-speech tag as the target vector; the preprocessed word vector and the part-of-speech vector representation learned through supervised learning are input into the role labeling model for fusion to obtain a vector representation with part-of-speech features, and the obtained vector representation is input into the role labeling model for training. After training, the final semantic role label of the word is obtained through conditional random field processing.

Images