Event detection method and system capable of efficiently utilizing limited annotation data

Abstract

The invention discloses an event detection method and system capable of efficiently utilizing limited annotation data, and the method comprises the steps: extracting event knowledge from the annotation data of a training corpus, and training a pre-training language model according to the event knowledge, so as to learn the event knowledge; acquiring an event data set, and inputting the event data set into the trained pre-training language model to generate event type representation; the event type representation and the event knowledge are combined to enhance the meaning of the trigger word to obtain a target representation of the trigger word; obtaining conditional probability distribution according to the target representation, and detecting an event according to the conditional probability distribution to identify an event type; according to the method, the events are identified and classified by utilizing information related to the events in the tasks, so that the performance of an event detection model is improved, and the annotation knowledge of limited annotation data is fully utilized; the system also has the same beneficial effects.

Description

Technical Field

[0001] The present application relates to the technical field of information extraction in natural language processing, and in particular to an event detection method and system that efficiently utilizes limited annotated data. Background Art

[0002] The rapid development of information technology and artificial intelligence technology has led to an exponential growth in the global data resource pool and online data volume, which has brought great challenges to people in obtaining target information from massive data in real time and accurately. Events, as a basic carrier of information, are used to describe the interaction between specific people, things, and objects in a certain form at a specific time and place. They are usually manifested as the occurrence of actions, changes in things, or changes in state. With their dynamic representation and large amount of information, they play a key role in information transmission and expression. As a basic task in the fields of natural language processing, information extraction, knowledge graphs, etc., event detection tasks require event detection models to automatically identify words or phrases that trigger any event in the predefined event set from unstructured text sequences, and further classify them into specified event types, and finally present specific information related to the event in a structured form.

[0003] The technical paradigms adopted by the current existing event detection technologies can be divided into two categories: event detection paradigm based on deep neural networks and event detection paradigm based on pre-trained language models. The event detection paradigm based on deep neural networks has the challenges of high cost of obtaining labeled data and the inability of the amount of labeled data to meet the training requirements of supervised models. Although the event detection paradigm based on pre-trained language models alleviates the demand for large-scale labeled data for downstream tasks and utilizes coarse-grained knowledge directly related to the task, it does not fully utilize the large amount of fine-grained and precise task-related knowledge provided in the labeled data.

[0004] In view of this, it is a technical problem that needs to be urgently solved by those skilled in the art to provide an event detection method and system for efficiently utilizing limited annotated data while improving the performance of the event detection model and making full use of the annotation knowledge provided by the limited annotated data. Summary of the invention

[0005] In order to solve the above technical problems, the purpose of the present invention is to provide an event detection method and system that efficiently utilizes limited annotated data, and uses event-related information in the task to identify and classify events, thereby improving the performance of the event detection model while making full use of the annotation knowledge of limited annotated data.

[0006] The first object of the present invention is to provide an event detection method that efficiently utilizes limited annotated data; The technical solution provided by the present invention is as follows: An event detection method that efficiently utilizes limited labeled data includes the following steps: Extracting event knowledge from the labeled data of the training corpus, and training the pre-trained language model according to the event knowledge to learn the event knowledge; Acquire an event data set, and input the event data set into a trained language model to generate an event type representation; The event type representation and the event knowledge are combined to enhance the trigger word semantics to obtain the target event representation of the trigger word; A conditional probability distribution is obtained according to the target event representation, and an event is detected according to the conditional probability distribution to identify the event type.

[0007] Preferably, extracting event knowledge from the annotated data of the training corpus and training the pre-trained language model according to the event knowledge to learn the event knowledge specifically includes: Extract event type-related data based on the annotation information in the training corpus; Encoding the event type-related data through the trained language model to obtain a sentence global event vector and a single event vector; A sentence event representation is obtained according to the sentence global event vector and the single event vector, and the event type is identified using a classifier.

[0008] Preferably, before acquiring the sentence event representation according to the sentence global event vector and the single event vector to identify the event type, the method further includes: An event knowledge learning loss function is constructed, the pre-trained language model is trained sentence by sentence using the annotated data in the training corpus, and the event knowledge learning loss function is optimized.

[0009] Preferably, the acquiring of an event data set and inputting the event data set into a trained pre-trained language model to generate an event type representation specifically includes: A set of annotated data is extracted from the training corpus as an event data set according to preset rules; The event data set is input into the trained pre-trained language model to generate event type representation.

[0010] Preferably, the step of jointly enhancing the trigger word semantics with the event type representation and the event knowledge to obtain a target representation of the trigger word specifically includes: Encode each word in the text sequence according to the trained language model to obtain the semantics of the trigger word itself and the event semantics of the text sequence; Obtaining a similarity score between the word-unit and each event type representation according to the word-unit and the event type representation; Acquire a trigger word representation with semantic enhancement of the event type according to the similarity score; The trigger word's own semantics, the event semantics of the text sequence, and the trigger word representation enhanced by the event type semantics are fused to obtain a target representation of the trigger word.

[0011] Preferably, acquiring a conditional probability distribution according to the target representation, and detecting an event according to the conditional probability distribution to identify an event type specifically includes: Calculate the conditional probability distribution of the trigger word in the task event type set according to the target representation of the trigger word; The event type triggered by the trigger word is identified based on the conditional probability distribution and the maximum probability principle.

[0012] Preferably, before identifying the event type triggered by the trigger word according to the conditional probability distribution combined with the maximum probability principle, the method further includes: An event trigger word feature learning loss function is constructed, the model is trained word by word using the annotated event trigger word data in the training corpus, and the event trigger word feature learning loss function is optimized.

[0013] A second object of the present invention is to provide an event detection system that efficiently utilizes limited annotated data; The technical solution provided by the present invention is as follows: An event detection system that efficiently utilizes limited labeled data, comprising: a learning module, an acquisition module, an enhancement module and a detection module; The learning module is used to extract event knowledge from the annotated data of the training corpus, and train the pre-trained language model according to the event knowledge to learn the event knowledge; The acquisition module is used to acquire an event data set and input the event data set into a trained language model to generate an event type representation; The enhancement module is used to enhance the trigger word semantics by combining the event type representation and the event knowledge to obtain a target representation of the trigger word; The detection module is used to obtain a conditional probability distribution according to the target representation, and detect an event according to the conditional probability distribution to identify an event type.

[0014] The third object of the present invention is to provide an electronic device; The technical solution provided by the present invention is as follows: An electronic device, comprising: at least one processor; and A memory communicatively connected to the at least one processor, the memory storing a computer program executable by the at least one processor, the computer program being executed by the at least one processor so that the at least one processor can perform any one of the method steps described in a method for efficiently detecting an event using limited labeled data.

[0015] A fourth object of the present invention is to provide a computer readable storage medium; The technical solution provided by the present invention is as follows: A computer-readable storage medium is used to store a computer program, wherein the computer program is used to enable a computer to execute any one of the method steps described in a method for efficiently detecting an event using limited labeled data.

[0016] The present invention provides an event detection method for efficiently utilizing limited annotated data, comprising: extracting event knowledge from the annotated data of a training corpus, and training a pre-trained language model based on the event knowledge to learn the event knowledge; obtaining an event data set, and inputting the event data set into the trained pre-trained language model to generate an event type representation; jointly enhancing the trigger word semantics with the event type representation and the event knowledge to obtain a target representation of the trigger word; obtaining a conditional probability distribution based on the target representation, and detecting events based on the conditional probability distribution to identify the event type; and using the method to utilize event-related information in a task to identify and classify events, thereby fully utilizing the annotated knowledge of limited annotated data while improving the performance of the event detection model.

[0017] The present invention also provides an event detection system that efficiently utilizes limited annotated data. Since the system and the event detection method that efficiently utilizes limited annotated data solve the same technical problem, belong to the same technical concept, and should have the same beneficial effects, they will not be described in detail here. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0019] Figure 1 Schematic diagram of a flow chart of an event detection method for efficiently utilizing limited annotated data in an embodiment of the present invention; Figure 2 Another schematic diagram of a flow chart of an event detection method for efficiently utilizing limited annotated data in an embodiment of the present invention; Figure 3 It is a structural schematic diagram of an event detection system that efficiently utilizes limited annotated data in an embodiment of the present invention; Figure 4 A schematic structural diagram of an electronic device in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] In order to enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below. Obviously, the described embodiments are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0021] It should be noted that when an element is referred to as being "fixed on" or "set on" another element, it can be directly on the other element or indirectly set on the other element; when an element is referred to as being "connected to" another element, it can be directly connected to the other element or indirectly connected to the other element.

[0022] It should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0023] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of this application, "multiple" and "several" mean two or more, unless otherwise clearly and specifically defined.

[0024] It should be noted that the structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification for people familiar with this technology to understand and read, and are not used to limit the conditions under which this application can be implemented. Therefore, they have no substantive technical significance. Any structural modification, change in proportional relationship or adjustment of size should still fall within the scope of the technical content disclosed in this application without affecting the effects and purposes that can be achieved by this application.

[0025] like Figure 1 to Figure 2As shown, an embodiment of the present invention provides an event detection method that efficiently utilizes limited labeled data, including the following steps: S1. extracting event knowledge from the annotated data of the training corpus, and training the pre-trained language model according to the event knowledge to learn the event knowledge; S2. Obtain an event data set, and input the event data set into a trained pre-trained language model to generate an event type representation; S3. Jointly enhancing the trigger word semantics by representing the event type and the event knowledge to obtain a target representation of the trigger word; S4. Acquire a conditional probability distribution according to the target representation, and detect events according to the conditional probability distribution to identify event types.

[0026] In step S1 to step S4, event knowledge is first extracted from the annotated data of a given training corpus, and a pre-trained language model is trained to learn event knowledge. Then, an event data set is obtained from the training corpus, and the event data set is input into the trained pre-trained language model to generate an event type representation. The learned event knowledge and event type representation are used to jointly enhance the trigger word semantics, thereby obtaining a target representation of the trigger word, aiming to make full use of the event type semantics. Finally, the conditional probability distribution is calculated according to the target representation, and the event is detected according to the conditional probability distribution to identify the event type. The identified event type is then used to train the model so that it learns the event trigger word features to fully model the event trigger word feature space. Through the above steps, the event-related information in the task is used to identify and classify the event, thereby improving the performance of the event detection model while making full use of the annotation knowledge of limited annotated data.

[0027] Preferably, extracting event knowledge from the annotated data of the training corpus and training the pre-trained language model according to the event knowledge to learn the event knowledge specifically includes: A1. Extract event type related data based on the annotation information in the preset corpus; In step A1, assume that the training corpus D = {s1, s2, …, s n} and the event type set E={l1,l2,…,l M}, where s i represents the i-th sentence in the training corpus, n is the total number of sentences in the training corpus, l i is the name of the i-th event type, M is the total number of event types in the event type set; the annotation information in a given corpus is mainly divided into two categories, one is all event-related annotation data DE containing noise words n ={sni}, if DE n Empty, indicating sentence s ni Without trigger words, if DE n Not empty, indicating sentence s ni There is at least one trigger word in the DE; the other is single event directly related annotation data DE that does not contain noise words p = , where t is the sentence s ni The number of trigger words in is the labeled data directly related to the event, k is the type of data directly related to the event, DE=DE n ∪DE p According to the annotation information in the input sentence, the event type related data DE is extracted through the instance extraction module in the event type network, and the automatic extraction of these annotation data is realized by designing two types of maskers, namely the global masker and the event type masker; the global masker takes all the contexts in the sentence as event-related annotation data, and the event type masker is composed of a group of event maskers, each of which is responsible for extracting the annotation data without noise words directly related to the event, such as the trigger word masker, the event masker, etc.; the event type network in this embodiment is composed of an instance extraction module, an event encoder, a semantic fusion module and a classifier.

[0028] A2. Encode the event type related data to obtain a sentence global event vector and a single event vector; In step A2, the event encoder encodes the extracted event type related data DE to obtain its high-dimensional space word embedding. In order to make the present invention more user-friendly, the pre-trained language model BERT with a small number of parameters is selected as the basic encoder, and the model is initialized using its parameters. Then, the event type related data DE is encoded to obtain two sets of word vectors representing the sentence global event vector DE n '={wni} and the single event vector set DE p '= .

[0029] A3. Obtain a sentence event representation based on the sentence global event vector and the single event vector to identify the event type.

[0030] In step A3, the semantic fusion module combines all single event vectors DE p 'Fuse into sentence global event vector DE n 'Thus, we get the event type; first, we fuse all the representations in a single event vector to get the representation DE of the event in the sentence p '= , then, calculate the similarity score between the representation of each event mentioned in the sentence and the global event representation of the sentence , and then use weighted summation to get the weighted representation of the sentence event ; Finally, the sentence event weighted representation is combined with the obtained sentence global event vector DE n 'Fuse and get the final event representation of the sentence ; Then the classifier is based on the obtained sentence event representation , identify the event type involved in the specific; first, use a forward neural network to extract the target event type features in the sentence event representation, and obtain the sentence event set representation ; Next, the conditional probability distribution of the sentence on the set of task event types is calculated based on the obtained sentence event concentrated representation; finally, the event type triggered by the sentence is identified based on the decoding rules.

[0031] Preferably, before acquiring the sentence event representation according to the sentence global event vector and the single event vector to identify the event type, the method further includes: An event knowledge learning loss function is constructed, the pre-trained language model is trained sentence by sentence using the annotated data in the training corpus, and the event knowledge learning loss function is optimized.

[0032] In the actual application process, the purpose of training the model to learn event knowledge is to optimize the loss function of the event type predicted by the sentence event representation and the event type actually mentioned in the sentence. The optimized loss function is as follows:

[0033] Where Y is the sentence s i The set of real event types, Y' is the sentence s i The set of predicted event types, y i Indicates event l i For the sentence i The true probability of the true event type, p i Indicates event l i For the sentence i The predicted probability of the real event type.

[0034] Preferably, the acquiring of an event data set and inputting the event data set into a trained pre-trained language model to generate an event type representation specifically includes: A set of annotated data is extracted from the training corpus as an event data set according to preset rules; The event data set is input into the trained pre-trained language model to generate event type representation.

[0035] In actual application, based on a set of predefined rules, an event data set is extracted from the training corpus for each event in the task event set: When extracting data for each event type, the extracted sentences must meet the following four rules: (1) There must be a trigger word in the sentence; (2) There can only be one trigger word in the sentence; (3) The number of data in the event type data set is determined by the type of trigger word for that event type; (4) For multiple sentences containing the same trigger word and triggering the same event type, the sentence containing the most argument types is selected based on the argument types. Under the guidance of these four criteria, an event data set S is extracted from the training corpus. E =[S l1 , S l2 ,…, S lM ], where S li =[s i1 ,s i2 ,…,s ip ], s ip For event l i Sentences containing the pth trigger word.

[0036] Then, we use the event types learned from the event type knowledge and the extracted event dataset S E Generate event type representation, represented by a matrix EM=[EM1, EM2,…, EM M ]; First, the event dataset S E The data in the sequence of event types is sorted by each event type S li All sentences are input into the encoder of the event type network for encoding, and the event representation w of each sentence is obtained. ip Then, all sentence event representations of each event type are fused to obtain the final representation EM of the event type. i =[ w i1 ,w i2 ,…, w ip ]; Finally, the representations of all event types are stored in the event type representation storage matrix in the form of key-value pairs (event type name: event type representation) to facilitate subsequent calls.

[0037] Preferably, the step of enhancing the trigger word semantics by using the event type representation and the event knowledge to obtain a target representation of the trigger word specifically includes: B1. Encode each word in the text sequence according to the trained language model to obtain the semantics of the trigger word itself and the event semantics of the text sequence; In step B1, the event type network that learns event type knowledge is used to train the text sequence s iEncode each word in the trigger word to obtain its own semantics w i and text sequence s i The event semantics w si , where w i is the i-th word x i High-dimensional space word vector.

[0038] B2. Obtaining a similarity score between a word unit and each event type representation according to the word unit and the event type representation; B3. Acquire the trigger word representation of the event type semantic enhancement according to the similarity score; In step B2 to step B3, the event type representation stored in the event type representation storage matrix is ​​used to enhance the trigger word semantics. The event type representation storage matrix stores M key-value pairs (l i , E.M. i ), where key l i Indicates event type l i Query, EM i For event type l i Representation; by calculating the word x i The similarity scores of all event types in the event type set are normalized, and then the weighted summation formula is used to obtain the semantically enhanced representation of its event type w i e .

[0039] B4. Fusing the trigger word's own semantics, the event semantics of the text sequence, and the trigger word representation enhanced by the event type semantics to obtain a target representation of the trigger word.

[0040] In step B4, finally, according to the trigger word's own semantics w i , event semantics si and the trigger word representation w enhanced by event type semantics i e Fusion is performed to obtain the final representation of the target trigger word, that is, the representation of the trigger word enhanced by event semantics. i ’ ; The event type triggered by the trigger word is the input text sequence s i The fact that a subset of event types is involved can be used to i The representation of the trigger word enhances the semantics of the trigger word. i , the event semantics involved in the input text sequence w si And the event type represents the enhanced semantics w i e Fusion is performed to obtain the target representation w of the trigger word i ’=[w i , w si , w i e ], where the fusion operation can be either the concatenation of vectors or the summation of vectors.

[0041] Preferably, acquiring a conditional probability distribution according to the target representation, and detecting an event according to the conditional probability distribution to identify an event type specifically includes: Calculate the conditional probability distribution of the target representation of the trigger word in the task event type set according to the target representation of the trigger word; The event type triggered by the trigger word is identified according to the conditional probability distribution combined with the heuristic rule.

[0042] In actual application, the target representation w is obtained based on the semantic enhancement of the word type. i ’ , the event classifier calculates its conditional probability distribution on the set of task event types, and uses the softmax function to simultaneously complete the recognition and classification of trigger words; given an input text sequence si, the event classifier processes the output of the multi-source semantic enhancement module token by token, and generates the score p of the i-th token on different event types i =EClassifier(FFN(w i ))=[p i1 ,p i2 ,…p iM ] and identify the event type triggered by the word based on the principle of maximum probability.

[0043] Preferably, before identifying the event type triggered by the trigger word according to the conditional probability distribution combined with the maximum probability principle, the method further includes: Construct an event trigger word feature learning loss function, train the model word by word using the annotated event trigger word data in the training corpus, and optimize the loss function.

[0044] In actual application, the trigger word feature learning loss function is optimized. By minimizing its loss value, the model can fully learn the trigger word features to fully model the event trigger word feature space. The optimized trigger word feature learning loss function is as follows:

[0045] Among them, y ij ∈{0, 1} indicates whether the true event type of the i-th word is event l j , n represents the total number of tokens in sentence si.

[0046] like Figure 3As shown, an embodiment of the present invention provides an event detection system that efficiently utilizes limited labeled data, including: a learning module, an acquisition module, an enhancement module and a detection module; The learning module is used to extract event knowledge from the annotated data of the training corpus, and train the model using the extracted event knowledge to learn the event knowledge; The acquisition module is used to represent the event type and jointly enhance the trigger word semantics to obtain the event type representation; The enhancement module is used to enhance the trigger word semantics by combining the event type representation and the event knowledge to obtain a target representation of the trigger word; The detection module is used to obtain a conditional probability distribution according to the target representation, and detect the event according to the conditional probability distribution to identify the event type; In the actual application process, a learning module, an acquisition module, an enhancement module and a detection module are set up, among which the acquisition module is connected to the learning module and the enhancement module respectively, and the enhancement module is also connected to the detection module; the learning module will extract event knowledge from the labeled data of the training corpus, and use the event knowledge extracted by the learning module to train the model so that it can learn event knowledge and transfer the learned event knowledge to the enhancement module; the enhancement module will jointly enhance the trigger word semantics with the event type representation and event knowledge, and after obtaining the target representation of the trigger word, the target representation of the trigger word is transmitted to the detection module; the detection module will obtain the conditional probability distribution according to the target representation of the trigger word, and detect the event according to the conditional probability distribution to identify the event type, and then use the identified event type to train the model so that it can learn the event trigger word features to fully model the event trigger word feature space; through this system, the event-related information in the task is used to identify and classify events, so as to improve the performance of the event detection model while making full use of the annotation knowledge of limited annotated data.

[0047] Furthermore, the present application also discloses an electronic device. Figure 4 This is a structural diagram of an electronic device according to an exemplary embodiment. The content in the diagram cannot be regarded as any limitation on the scope of use of the present application.

[0048] Figure 4A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 is used to store a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the event detection method for efficiently utilizing limited annotated data disclosed in any of the aforementioned embodiments. In addition, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0049] In this embodiment, the power supply 23 is used to provide working voltage for each hardware device on the electronic device 20; the communication interface 24 can create an event detection channel between the electronic device 20 and external devices for efficient use of limited annotation data. The communication protocol it follows is any communication protocol that can be applied to the technical solution of the present application and is not specifically limited here; the input and output interface 25 is used to obtain external input data or output data to the outside world. The specific interface type can be selected according to specific application needs and is not specifically limited here.

[0050] In addition, the memory 22, as a carrier for storing resources, can be a read-only memory, a random access memory, a disk or an optical disk, etc. The resources stored thereon may include an operating system 221, a computer program 222 and data 223, etc. The storage method can be temporary storage or permanent storage.

[0051] Among them, the operating system 221 is used to manage and control the hardware devices and computer programs 222 on the electronic device 20 to realize the operation and processing of the data 223 in the memory 22 by the processor 21, which can be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program that can be used to complete the event detection method for efficiently utilizing limited annotation data performed by the electronic device 20 disclosed in any of the aforementioned embodiments, the computer program 222 can further include a computer program that can be used to complete other specific tasks. In addition to including data transmitted from an external device received by the event detection device for efficiently utilizing limited annotation data, the data 223 can also include data collected by its own input and output interface 25, etc.

[0052] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0053] Furthermore, the present application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, the aforementioned disclosed event detection method for efficiently utilizing limited annotated data is implemented. The specific steps of the method can be referred to the corresponding contents disclosed in the aforementioned embodiments, and will not be repeated here.

[0054] It should be understood that the use of "method", "device", "unit" and / or "module" in this application is only a method for distinguishing different components, elements, parts, portions or assemblies at different levels. However, if other words can achieve the same purpose, the word can be replaced by other expressions.

[0055] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "a kind" and / or "the" do not refer to the singular, but also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of clearly identified steps and elements, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements. The elements defined by the sentence "includes a..." do not exclude the existence of other identical elements in the process, method, commodity or device that includes the elements.

[0056] In the following, the terms "first" and "second" are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features.

[0057] If a flow chart is used in the present application, the flow chart is used to illustrate the operations performed by the system according to the embodiment of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, each step can be processed in reverse order or simultaneously. At the same time, other operations can also be added to these processes, or a certain step or several steps of operations can be removed from these processes.

[0058] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An event detection method that efficiently utilizes limited labeled data, characterized in that: The steps include: Extracting event knowledge from the labeled data of the training corpus, and training the pre-trained language model according to the event knowledge to learn the event knowledge; Acquire an event data set, and input the event data set into a trained pre-trained language model to generate an event type representation; The event type representation and the event knowledge are combined to enhance the trigger word semantics to obtain a target representation of the trigger word; A conditional probability distribution is obtained according to the target representation, and an event is detected according to the conditional probability distribution to identify the event type.

2. The event detection method for efficiently utilizing limited annotated data according to claim 1, characterized in that: The extracting event knowledge from the labeled data of the training corpus and training the pre-trained language model according to the event knowledge to learn the event knowledge specifically includes: Extract event type-related data based on the annotation information in the training corpus; Encoding the event type-related data through the trained language model to obtain a sentence global event vector and a single event vector; A sentence event representation is obtained according to the sentence global event vector and the single event vector, and the event type is identified using a classifier.

3. The event detection method for efficiently utilizing limited annotated data according to claim 2, characterized in that: Before acquiring the sentence event representation according to the sentence global event vector and the single event vector to identify the event type, the method further includes: An event knowledge learning loss function is constructed, the pre-trained language model is trained sentence by sentence using the annotated data in the training corpus, and the event knowledge learning loss function is optimized.

4. The event detection method for efficiently utilizing limited annotated data according to claim 1, characterized in that: The acquiring of an event data set and inputting the event data set into a trained pre-trained language model to generate an event type representation specifically includes: A set of annotated data is extracted from the training corpus as an event data set according to preset rules; The event data set is input into the trained pre-trained language model to generate event type representation.

5. The event detection method for efficiently utilizing limited annotated data according to claim 1, characterized in that: The step of jointly enhancing the trigger word semantics with the event type representation and the event knowledge to obtain the target representation of the trigger word specifically includes: Encode each word in the text sequence according to the trained language model to obtain the semantics of the trigger word itself and the event semantics of the text sequence; Obtaining a similarity score between the word-gram and each event type representation according to the word-gram and the event type representation; Acquire a trigger word representation with semantic enhancement of the event type according to the similarity score; The trigger word's own semantics, the event semantics of the text sequence, and the trigger word representation enhanced by the event type semantics are fused to obtain a target representation of the trigger word.

6. The event detection method for efficiently utilizing limited annotated data according to claim 1, characterized in that: The acquiring of a conditional probability distribution according to the target representation and detecting an event according to the conditional probability distribution to identify an event type specifically includes: Calculating a conditional probability distribution in a set of task event types based on the target representation of the trigger word; The event type triggered by the trigger word is identified based on the conditional probability distribution and the maximum probability principle.

7. The event detection method for efficiently utilizing limited annotated data according to claim 6, characterized in that: Before identifying the event type triggered by the trigger word according to the conditional probability distribution combined with the maximum probability principle, the method further includes: An event trigger word feature learning loss function is constructed, the model is trained word by word using the annotated event trigger word data in the training corpus, and the event trigger word feature learning loss function is optimized.

8. An event detection system that efficiently utilizes limited annotated data, characterized in that: include: Learning module, acquisition module, enhancement module and detection module; The learning module is used to extract event knowledge from the annotated data of the training corpus, and train the pre-trained language model according to the event knowledge to learn the event knowledge; The acquisition module is used to acquire an event data set and input the event data set into a trained language model to generate an event type representation; The enhancement module is used to enhance the trigger word semantics by combining the event type representation and the event knowledge to obtain a target representation of the trigger word; The detection module is used to obtain a conditional probability distribution according to the target representation, and detect events according to the conditional probability distribution to identify event types.

9. An electronic device, characterized in that: include: at least one processor; as well as A memory communicatively connected to the at least one processor, the memory storing a computer program executable by the at least one processor, the computer program being executed by the at least one processor so that the at least one processor can execute the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The storage medium is used to store a computer program, and the computer program is used to enable a computer to execute the method according to any one of claims 1 to 7.

Images