Chapter-level event extraction method, device and equipment and storage medium

Abstract

The application discloses a chapter-level event extraction method and device, equipment and a storage medium. For a target chapter carrying title information, event parameter extraction is directly performed without trigger word identification. An event extraction result is composed of parameter values of various types of event parameters, event names and event types, and is more in line with chapter characteristics. Since trigger word identification is omitted, the entire processing procedure is simpler. In the event parameter extraction process, the target chapter is divided into sentences, and parameter values of various types of event parameters in each sentence are extracted in sequence according to an event parameter type template. Since event parameter extraction is performed at the sentence level, the marked event parameters can be prevented from being chaotic. For each type of event parameter, the parameter values of the type of event parameter extracted in each sentence are integrated to obtain integrated parameter values. The integrity of the parameter values of various types of event parameters extracted at the chapter level is ensured, and the problem of chaotic marked event parameters is avoided.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, and more specifically, to a method, apparatus, device, and storage medium for extracting document-level events. Background Technology

[0002] An event refers to a specific occurrence at a particular time and place involving one or more participants, often described as a change in state. In the field of natural language processing, event extraction is a common task, primarily involving the transformation of unstructured text describing event information into structured text. It has wide applications in areas such as online public opinion monitoring, emergency alerts, and intelligence gathering.

[0003] Existing event extraction methods are generally designed for sentence-level event extraction, with few solutions for document-level event extraction. The few document-level event extraction solutions that exist also follow the sentence-level approach. For example, they first extract trigger words from the entire document, then use this information to input the entire document into a sequence labeling model to label event parameters (also known as argument parameters). Finally, the event extraction result for the document is composed of the trigger words and event parameters.

[0004] Existing document-level event extraction schemes do not consider the characteristics of documents, and using the same scheme as sentence-level event extraction can easily lead to cumbersome processing procedures. Furthermore, the method of sequence labeling the entire document to obtain event parameters is prone to problems of disordered labeled event parameters. Summary of the Invention

[0005] In view of the above problems, this application is proposed to provide a method, apparatus, device, and storage medium for extracting chapter-level events, so as to solve the problems of cumbersome processing flow and easy confusion of labeled event parameters in existing chapter-level event extraction schemes. The specific solution is as follows:

[0006] Firstly, a chapter-level event extraction method is provided, including:

[0007] The target text is segmented into sentences to obtain a sentence sequence, and the target text carries title information;

[0008] According to the set event parameter type template, the parameter values ​​of each type of event parameter in each sentence of the sentence sequence are extracted sequentially. The event parameter type template contains several set event parameters.

[0009] For each type of event parameter, the parameter values ​​of the type of event parameter extracted from each sentence in the sentence sequence are integrated to obtain the integrated parameter value of the type of event parameter;

[0010] Based on the title information of the target chapter, determine the event name of the target chapter;

[0011] Obtain the domain information of the target text, and determine the event type of the target text based on the domain information;

[0012] The event extraction result of the target chapter is composed of the event type, event name, and integrated parameter values ​​of each type of event.

[0013] Secondly, a chapter-level event extraction device is provided, including:

[0014] Sentence segmentation units are used to segment the target text into sentences to obtain a sentence sequence;

[0015] An event parameter extraction unit is used to extract the parameter values ​​of each type of event parameter in each sentence of the sentence sequence according to a set event parameter type template. The event parameter type template contains several set event parameters of different types.

[0016] An event parameter integration unit is used to integrate the parameter values ​​of the event type extracted from each sentence in the sentence sequence for each type of event parameter, so as to obtain the integrated parameter value of the event type.

[0017] The event name determination unit is used to determine the event name of the target chapter based on the title of the target chapter;

[0018] An event type determination unit is used to obtain the domain information of the target chapter and determine the event type of the target chapter based on the domain information;

[0019] The event extraction result determination unit is used to compose the event extraction result of the target chapter from the event type, event name and integrated parameter values ​​of each type of event parameter.

[0020] Thirdly, a chapter-level event extraction device is provided, including: a memory and a processor;

[0021] The memory is used to store programs;

[0022] The processor is used to execute the program to implement the various steps of the chapter-level event extraction method described above.

[0023] Fourthly, a storage medium is provided on which a computer program is stored, which, when executed by a processor, implements the various steps of the chapter-level event extraction method as described above.

[0024] By employing the above technical solution, the document-level event extraction method of this application, for target documents carrying title information, takes into account the characteristics of documents that contain explicit title information, which can reflect the event name and event type. Based on this, it eliminates the need for trigger word identification in the first stage of sentence-level event extraction and can directly enter the event parameter extraction process. The event extraction result is composed of the parameter values ​​of each type of event parameter, the event name, and the event type. In other words, it provides a new descriptive form of document-level event extraction results, which is more in line with the characteristics of documents. Furthermore, since the identification of trigger words is eliminated, the entire processing flow is simpler.

[0025] Furthermore, in the event parameter extraction process, to avoid the problem of chaotic labeled event parameters caused by existing technologies that directly extract event parameters from the entire text using sequence labeling, this case first segments the target text into sentences to obtain a sentence sequence. Then, according to the set event parameter type template, the parameter values ​​of each type of event parameter in each sentence of the sentence sequence are extracted sequentially. Since the event parameter extraction is performed at the sentence level, the problem of chaotic labeled event parameters can be avoided. On this basis, for each type of event parameter, the parameter values ​​of that type of event parameter extracted from each sentence in the sentence sequence are integrated to obtain the integrated parameter value of that type of event parameter. This ensures the integrity of the parameter values ​​of each type of event parameter extracted at the text level and avoids the problem of chaotic labeled event parameters. Attached Figure Description

[0026] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0027] Figure 1 A flowchart illustrating the chapter-level event extraction method provided in this application embodiment;

[0028] Figure 2 This example illustrates the parameter value extraction results for each type of event parameter in a sentence sequence.

[0029] Figure 3 An example of a descriptive template for the event extraction results of a target chapter is provided.

[0030] Figure 4 This example illustrates the structure of a BERT pre-trained language model.

[0031] Figure 5 A schematic diagram of a chapter-level event extraction device provided in this application embodiment;

[0032] Figure 6 A schematic diagram of the structure of the chapter-level event extraction device provided in the embodiments of this application. Detailed Implementation

[0033] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] This application provides a chapter-level event extraction scheme, which is applicable to extracting events from chapters and obtaining event extraction results.

[0035] The proposed solution can be implemented based on a terminal with data processing capabilities, such as a mobile phone, computer, server, or cloud platform.

[0036] Next, combined Figure 1 As shown, the chapter-level event extraction method of this application may include the following steps:

[0037] Step S100: Segment the target text to obtain a sentence sequence.

[0038] The target chapter is the chapter from which event extraction needs to be performed. The target chapter carries title information. For example, the target chapter could be a news report, which contains explicit title information that reflects the event name and type.

[0039] In this embodiment, in order to ensure that the target text is processed completely and that there is no loss of information due to the limitation of text length, as well as problems such as confusion of labeled event parameters, the target text can first be processed into sentences to obtain the sentence sequence corresponding to the target text.

[0040] When performing sentence segmentation, various natural language processing tools can be used to segment the target text into several single sentences, ultimately obtaining the sentence sequence {sentence1, sentence2, ..., sentenceN}.

[0041] In this process, after segmenting the target text into sentences, the position of each sentence can be recorded, that is, the sentences are sorted in the sentence sequence according to their position in the target text.

[0042] Step S110: According to the set event parameter type template, extract the parameter values ​​of each type of event parameter in each sentence of the sentence sequence in sequence.

[0043] This application can pre-construct an event parameter type template based on the domain information of the chapter to be extracted. The event parameter type template can contain several set event parameter types, such as PER (person), LOC (location), TIME (time), MONEY (property), ACTION (action), and other event parameters.

[0044] Optionally, this application can construct a corresponding event parameter type template for each event type. The event parameter type template contains event parameters of various types that match the event type. In other words, event parameter type templates that match different event types can be constructed.

[0045] Based on this, the process of extracting the parameter values ​​of each type of event parameter in each sentence of the sentence sequence according to the set event parameter type template is as follows:

[0046] According to the event parameter type template corresponding to the event type of the target chapter, the parameter values ​​of each type of event parameter in each sentence of the sentence sequence are extracted sequentially.

[0047] Alternatively, this application can also construct a unified event parameter type template for each event type. That is, different event types correspond to the same event parameter type template, which can contain various types of event parameters that may appear in various event types. Table 1 below illustrates a case where different event types correspond to the same event parameter type template:

[0048] Table 1

[0049]

[0050]

[0051] It should be noted that Table 1 above is merely an example, and event parameter types can be added or removed according to actual needs. Furthermore, Table 1 only illustrates a few optional event types; other event types can be added based on the specific scenario.

[0052] In this step, for each type of event parameter contained in the template, the parameter value of that type of event parameter is extracted sequentially from each sentence in the sentence sequence, thereby obtaining the parameter value of each type of event parameter extracted from each sentence.

[0053] Take the sentence "Xiaoming bought a mobile phone this morning" as an example:

[0054] The parameter values ​​for each type of event extracted from this sentence are as follows:

[0055] PER: Xiaoming; TIME: This morning; MONEY: A mobile phone; ACTION: Purchase.

[0056] Since a sentence sequence typically contains multiple sentences, parameter values ​​for each type of event are extracted from each sentence separately. The final extraction result can be as follows: Figure 2 As shown, from the first sentence Sentence1 to the last sentence SentenceN in the sentence sequence, the parameter values ​​of each type of event parameter were extracted according to the event parameter type template. Figure 2 The column for parameter values ​​is not filled. You can fill it with the extracted parameter values ​​based on the extraction results. Figure 2 (A column of parameter values).

[0057] Step S120: For each type of event parameter, integrate the parameter values ​​of the type of event parameter extracted from each sentence in the sentence sequence.

[0058] Specifically, by integrating the parameter values ​​of each type of event parameter, the integrated parameter value of that type of event parameter can be obtained.

[0059] In one alternative implementation, for each type of event parameter in the event parameter type template:

[0060] The parameter values ​​of the type event parameters extracted from each sentence can be combined sequentially according to the order of the sentences in the sentence sequence to obtain the integrated parameter value of the type event parameters. If the parameter value of the type event parameter extracted from a certain sentence is empty, a preset character (such as padding) can be used to replace the parameter value of the type event parameter.

[0061] Reference Figure 3 This example provides a template for describing the results of event extraction from a target text. The dashed box section uses the PER (Person) event parameter as an example to illustrate the result after parameter value integration. Figure 3 The example demonstrates the use of square brackets "[]" to represent the collection of integrated parameter values ​​at the document level, and parentheses "()" to represent event parameter values ​​for sentence-level recognition. If a sentence contains multiple event parameters of the same type, they are separated by "\", such as... Figure 3 The example of per1\per2 in the first parentheses indicates that two different characters are extracted from the first sentence, represented by per1 and per2 respectively. If the parameter value of a certain type of event extracted in a sentence is null, it can be represented using the special character "PADDING".

[0062] for Figure 3 The event parameters such as "LOC (Location)," "TIME (Time)," "MONEY (Finance)," and "ACTION (Action)" have results similar in format to those of "PER (Person)." Figure 3 No example provided.

[0063] Step S130: Determine the event name of the target chapter based on the title information of the target chapter.

[0064] Specifically, given the characteristics of a text, namely that its title information can clearly express the key information of the event, the event name can be determined based on the title information of the target text in this embodiment.

[0065] In one optional scenario, the title information of the target chapter can be directly used as the event name of the target chapter. Alternatively, the theme of the title information of the target chapter can be extracted and used as the event name of the target chapter. When extracting the theme from the title information, various existing theme extraction tools can be used.

[0066] like Figure 3 As shown, the event names of the identified target chapters can be filled into the second row.

[0067] Step S140: Obtain the domain information of the target chapter, and determine the event type of the target chapter based on the domain information.

[0068] Specifically, the target text may directly contain information about its domain, or its domain can be determined based on how it was obtained. Alternatively, natural language processing methods can be used to determine the domain information of the target text.

[0069] After determining the domain information to which the target text belongs, the event type of the target text can be determined based on its domain information.

[0070] In one alternative approach, a target event type that is closest to the domain information of the target article can be selected from a set event type template as the event type of the target article. The event type template can contain several set event types, as illustrated in the first column of Table 1 above.

[0071] like Figure 3 As shown, the event types of the target chapter can be filled into the first row.

[0072] It should be noted that there is no necessary order between steps S130-S140 and steps S100-S120. Figure 1 Only one optional case is shown.

[0073] Step S150: The event extraction result of the target chapter is composed of the event type, event name and integrated parameter values ​​of each type of event parameter.

[0074] Specifically, in this embodiment, the event extraction result of the target chapter can be composed of the event type, event name, and integrated parameter values ​​of each type of event parameter of the target chapter. For example... Figure 3 This example provides a template for the event extraction results of a target chapter.

[0075] The document-level event extraction method provided in this application, for target documents carrying title information, takes into account the characteristics of documents, namely, that they contain explicit title information, which can reflect the event name and event type. Based on this, it eliminates the need for trigger word identification in the first stage of sentence-level event extraction and can directly enter the event parameter extraction process. The event extraction result is composed of the parameter values ​​of each type of event parameter, the event name, and the event type. In other words, it provides a new descriptive form of document-level event extraction results, which is more in line with the characteristics of documents. Furthermore, since the identification of trigger words is eliminated, the entire processing flow is simpler.

[0076] Furthermore, in the event parameter extraction process, to avoid the problem of chaotic labeled event parameters caused by existing technologies that directly extract event parameters from the entire text using sequence labeling, this case first segments the target text into sentences to obtain a sentence sequence. Then, according to the set event parameter type template, the parameter values ​​of each type of event parameter in each sentence of the sentence sequence are extracted sequentially. Since the event parameter extraction is performed at the sentence level, the problem of chaotic labeled event parameters can be avoided. On this basis, for each type of event parameter, the parameter values ​​of that type of event parameter extracted from each sentence in the sentence sequence are integrated to obtain the integrated parameter value of that type of event parameter. This ensures the integrity of the parameter values ​​of each type of event parameter extracted at the text level and avoids the problem of chaotic labeled event parameters.

[0077] In some embodiments of this application, an optional implementation of step S110 above, which extracts the parameter values ​​of each type of event parameter in each sentence of the sentence sequence, is described.

[0078] Specifically, in this embodiment, a pre-trained sequence labeling model can be used to label the parameters of each type of event in each sentence of the sentence sequence, thereby obtaining the parameter values ​​of each type of event.

[0079] The sequence labeling model can be trained using training sentences labeled with the event parameter type labels of each word in the sentence as training data.

[0080] The sequence labeling module can use pre-trained language models such as BERT; the structure can be referenced. Figure 4 As shown:

[0081] Among them, E1, E2...E N Let T1, T2...T be the input vectors of the model. N This is the output vector of the model.

[0082] The BERT pre-trained language model uses the Transformer feature extractor, which can include multiple stacked units, each consisting of a self-attention mechanism and a feedforward neural network.

[0083] In the self-attention mechanism, each word corresponds to three different vectors: a query vector (Q), a key vector (K), and a value vector (V), all of equal length. The embedding vectors are multiplied by three different weight matrices W. q W k W v The importance score for each word is obtained by multiplying the query vector and the key vector: score = q·k

[0084] The attention value is obtained by smoothing the score using SoftMax, and the smoothed result is multiplied by the value vector.

[0085]

[0086] d k Let Q and K be one dimension. The penalty factor ensures that the inner product of Q and K does not become excessively large. Through self-attention operations, the original input matrix... It became a new matrix

[0087] In practice, the attention-encoded vector Z is typically calculated first using an attention mechanism, and then Z is fed into a feedforward neural network layer to train downstream tasks.

[0088] FFN(Z) = max(0, ZW1+b1) + b2

[0089] The BERT pre-trained language model uses a multi-head attention mechanism, which consists of multiple self-attention mechanisms, to acquire sentence-level semantic information.

[0090] Multihead=concat(head1, head2,..., headn ) ·W

[0091]

[0092] To further address the challenges of training deep neural networks, the Transformer module can also employ "short-circuit connections" and "layer normalization." "Short-circuit connections" refer to passing information from the previous layer to the next without error, thus resolving the vanishing gradient problem in deep neural networks. "Layer normalization" involves normalizing the activation values ​​of each layer, which accelerates the model training process and allows the model to converge more quickly.

[0093] Conditional random fields (CRFs) are a common algorithm in sequence labeling tasks, and therefore can be used to label event parameter types. Linear-chain CRFs are frequently used in sequence labeling models; they are discriminative models that predict output sequences based on input sequences. For a given sequence X(x1, x2, ... x... n ), whose corresponding label is Y(y1, y2, ... y n If the following conditions are met:

[0094] P(Y i |X, Y1, ..., Y i-1 Y i+1 , ...Y n )=P(Y i |X, Y i-1 Y i+1 )

[0095] Let P(N, K) be the weight matrix output by the decoding layer, and then the evaluation score S(x, y) can be obtained, i.e.

[0096]

[0097] Where A is the transition matrix, k is the number of labels, and n is the sequence length. The maximum probability of sequence label y can be calculated using the SoftMax function, i.e.

[0098]

[0099] During training, the maximum likelihood method can be used to solve for the maximum posterior probability of P(y|x):

[0100]

[0101] In some embodiments of this application, based on the aforementioned chapter-level event extraction scheme, a further chapter-level event correlation analysis scheme is provided.

[0102] When there are multiple target chapters for which event extraction is required, events are extracted for each target chapter in accordance with the aforementioned embodiment. After obtaining the event extraction results for each target chapter, correlation analysis can be performed on multiple target chapters based on the event extraction results of multiple target chapters.

[0103] This embodiment provides two methods for correlation analysis:

[0104] First, conduct event correlation analysis from a semantic perspective.

[0105] In this embodiment, the semantic dimension can be further subdivided into the event type perspective and the event parameter perspective.

[0106] From the perspective of event type:

[0107] Event type can be used as a classification criterion. Based on the event extraction results of multiple target chapters, multiple target chapters can be divided into at least one set of events of the same type.

[0108] As can be seen from the aforementioned event extraction process of the target chapter, the final event extraction result includes the event types of the target chapter. Therefore, when conducting event correlation analysis, the event type can be used as a classification condition to divide multiple target chapters into sets of events of the same type.

[0109] From the perspective of event parameters:

[0110] If several events share similar or nearly identical event parameter information, then these events are considered similar. For example, suppose we know events A, B, and C: if events A and B occur at similar or identical times, then events A and B are similar in the dimension of "time"; if events A and C occur at similar or identical locations, then events A and C are similar in the dimension of "location"; if events A, B, and C involve the same person, then events A, B, and C are similar in the dimension of "person". To determine the similarity between the parameters of two events, we can use semantic similarity measurement methods to directly calculate the similarity of the event parameter values ​​between each pair of events.

[0111] The specific implementation process may include:

[0112] S1. For each type of event parameter, calculate the semantic similarity of the integrated parameter values ​​of the type of event parameters for each pair of target chapters.

[0113] When calculating the semantic similarity of integrated parameter values, a semantic similarity measurement method can be used.

[0114] S2. Determine whether there is a semantic relationship between two target chapters whose semantic similarity exceeds a set similarity threshold in terms of the type of event parameter.

[0115] The similarity threshold can be set according to the actual situation, such as 100% or other values ​​less than 1.

[0116] Second, conduct event correlation analysis from a statistical perspective.

[0117] The semantic-dimensional correlation analysis described above analyzes the relationships between multiple events from a single dimension, while the statistical-dimensional correlation analysis considers semantic information from multiple dimensions simultaneously to determine the relationships between multiple events. This embodiment illustrates two statistical-dimensional correlation methods, as shown in Table 2 below:

[0118] Table 2

[0119]

[0120] The specific implementation process may include:

[0121] S1. For any two target chapters, compare the integrated parameter values ​​of each type of event parameter to see if they are the same.

[0122] S2. Based on the comparison results, it is determined that there is a co-occurrence relationship between the two target chapters that meet the first condition, and that there is a sequential relationship between the two target chapters that meet the second condition.

[0123] The first and second conditions mentioned above are as follows:

[0124] The first condition is that the integrated parameter values ​​of the time parameter and the location parameter in the event extraction results of the two target chapters are the same.

[0125] The second condition is that the integrated parameter values ​​of the character parameters and location parameters in the event extraction results of the two target chapters are the same, while the integrated parameter values ​​of the time parameters are different.

[0126] Of course, the above only illustrates two types of event relationships. In addition, other event relationships can be derived by analyzing the relationships from the dimensions of other event parameters. This embodiment will not exhaustively list them.

[0127] In the above embodiments of this application, after completing the chapter-level event extraction, two methods for analyzing event relationships are further introduced from the semantic and statistical dimensions, respectively. These methods establish relationships between events belonging to multiple target chapters, providing a new approach to assist users in analyzing the relationships between target chapters and between events. Moreover, the results of this event relationship analysis have strong feasibility and reliability.

[0128] The document-level event extraction apparatus provided in the embodiments of this application is described below. The document-level event extraction apparatus described below and the document-level event extraction method described above can be referred to in correspondence.

[0129] See Figure 5 , Figure 5 This is a schematic diagram of a chapter-level event extraction device disclosed in an embodiment of this application.

[0130] like Figure 5 As shown, the device may include:

[0131] Sentence segmentation unit 11 is used to segment the target text into sentences to obtain a sentence sequence;

[0132] The event parameter extraction unit 12 is used to extract the parameter values ​​of each type of event parameter in each sentence of the sentence sequence according to the set event parameter type template. The event parameter type template contains several set event parameters of different types.

[0133] Event parameter integration unit 13 is used to integrate the parameter values ​​of the type of event parameter extracted from each sentence in the sentence sequence for each type of event parameter to obtain the integrated parameter value of the type of event parameter;

[0134] Event name determination unit 14 is used to determine the event name of the target chapter based on the title of the target chapter;

[0135] Event type determination unit 15 is used to obtain the domain information of the target chapter and determine the event type of the target chapter based on the domain information;

[0136] The event extraction result determination unit 16 is used to compose the event extraction result of the target chapter from the event type, event name and integrated parameter values ​​of each type of event parameter.

[0137] Optionally, the process by which the event parameter extraction unit sequentially extracts the parameter values ​​of each type of event parameter in each sentence of the sentence sequence may include:

[0138] A pre-trained sequence labeling model is used to label the event parameters of each type in each sentence of the sentence sequence, and the parameter values ​​of each type of event parameter are obtained.

[0139] The sequence labeling model is trained using training sentences labeled with event parameter type tags to which each word in the sentence belongs as training data.

[0140] Optionally, the process by which the event parameter integration unit integrates the parameter values ​​of the event type extracted from each sentence in the sentence sequence to obtain the integrated parameter value of the event type for each type of event parameter may include:

[0141] For each type of event parameter in the event parameter type template:

[0142] According to the sorting order of each sentence in the sentence sequence, the parameter values ​​of the type event parameters extracted from each sentence are combined sequentially to obtain the integrated parameter value of the type event parameters. If the parameter value of the type event parameter extracted from a certain sentence is empty, then a preset character is used to replace the parameter value of the type event parameter.

[0143] Optionally, the process by which the event name determination unit determines the event name of the target chapter based on the title information of the target chapter may include:

[0144] The title information of the target chapter is used as the event name of the target chapter;

[0145] or,

[0146] Extract the theme from the title information of the target chapter and use it as the event name of the target chapter.

[0147] Optionally, the process by which the event type determination unit determines the event type of the target text based on the domain information to which the target text belongs may include:

[0148] In the set event type template, select the target event type that is closest to the domain information of the target chapter, and use it as the event type of the target chapter;

[0149] The event type template contains several defined event types.

[0150] Optionally, when there are multiple target chapters for event extraction, the apparatus of this application may further include:

[0151] The event association analysis unit is used to perform event association analysis from semantic and / or statistical dimensions based on the event extraction results of multiple target chapters after obtaining the event extraction results of each target chapter.

[0152] Optionally, the process by which the event association analysis unit performs event association analysis from a semantic dimension based on the event extraction results of multiple target chapters may include:

[0153] Using event type as the classification condition, based on the event extraction results of multiple target chapters, the multiple target chapters are divided into at least one set of events of the same type;

[0154] And / or,

[0155] For each type of event parameter, calculate the semantic similarity of the integrated parameter values ​​of the event parameters of that type for each pair of target chapters;

[0156] It is determined that two target texts with a semantic similarity exceeding a set similarity threshold have a semantic association on the type of event parameter.

[0157] Optionally, the process by which the above-mentioned event correlation analysis unit performs event correlation analysis from a statistical dimension based on the event extraction results of multiple target chapters may include:

[0158] For any two target chapters, compare the integrated parameter values ​​of each type of event parameter to see if they are the same.

[0159] Based on the comparison results, it was determined that two target texts satisfying the first condition have a co-occurrence relationship in terms of events, and that two target texts satisfying the second condition have a sequential relationship in terms of events, wherein:

[0160] The first condition is that the integrated parameter values ​​of the time parameter and the location parameter in the event extraction results of the two target chapters are the same.

[0161] The second condition is that the integrated parameter values ​​of the character parameter and the location parameter in the event extraction results of the two target chapters are the same, while the integrated parameter values ​​of the time parameter are different.

[0162] The document-level event extraction device provided in this application embodiment can be applied to document-level event extraction devices, such as terminals: mobile phones, computers, etc. Optionally, Figure 6 The hardware structure block diagram of the chapter-level event extraction device is shown below. Figure 6 The hardware structure of a chapter-level event extraction device may include: at least one processor 1, at least one communication interface 2, at least one memory 3, and at least one communication bus 4.

[0163] In this embodiment of the application, the number of processor 1, communication interface 2, memory 3, and communication bus 4 is at least one, and processor 1, communication interface 2, and memory 3 communicate with each other through communication bus 4;

[0164] Processor 1 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention.

[0165] Memory 3 may include high-speed RAM, and may also include non-volatile memory, such as at least one disk storage device;

[0166] The memory stores a program, which the processor can call. The program is used for:

[0167] The target text is segmented into sentences to obtain a sentence sequence, and the target text carries title information;

[0168] According to the set event parameter type template, the parameter values ​​of each type of event parameter in each sentence of the sentence sequence are extracted sequentially. The event parameter type template contains several set event parameters.

[0169] For each type of event parameter, the parameter values ​​of the type of event parameter extracted from each sentence in the sentence sequence are integrated to obtain the integrated parameter value of the type of event parameter;

[0170] Based on the title information of the target chapter, determine the event name of the target chapter;

[0171] Obtain the domain information of the target text, and determine the event type of the target text based on the domain information;

[0172] The event extraction result of the target chapter is composed of the event type, event name, and integrated parameter values ​​of each type of event.

[0173] Optionally, the refined and extended functions of the program can be found in the description above.

[0174] This application embodiment also provides a storage medium that can store a program suitable for execution by a processor, the program being used for:

[0175] The target text is segmented into sentences to obtain a sentence sequence, and the target text carries title information;

[0176] According to the set event parameter type template, the parameter values ​​of each type of event parameter in each sentence of the sentence sequence are extracted sequentially. The event parameter type template contains several set event parameters.

[0177] For each type of event parameter, the parameter values ​​of the type of event parameter extracted from each sentence in the sentence sequence are integrated to obtain the integrated parameter value of the type of event parameter;

[0178] Based on the title information of the target chapter, determine the event name of the target chapter;

[0179] Obtain the domain information of the target text, and determine the event type of the target text based on the domain information;

[0180] The event extraction result of the target chapter is composed of the event type, event name, and integrated parameter values ​​of each type of event.

[0181] Optionally, the refined and extended functions of the program can be found in the description above.

[0182] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0183] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0184] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily 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 this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for extracting events at the chapter level, characterized in that, include: The target text is segmented into sentences to obtain a sentence sequence, and the target text carries title information; According to the set event parameter type template, the parameter values ​​of each type of event parameter in each sentence of the sentence sequence are extracted sequentially. The event parameter type template contains several set event parameters. For each type of event parameter in the event parameter type template: according to the sorting order of each sentence in the sentence sequence, the parameter values ​​of the type of event parameter extracted from each sentence are combined sequentially to obtain the integrated parameter value of the type of event parameter. If the parameter value of the type of event parameter extracted from a certain sentence is empty, then a preset character is used to replace the parameter value of the type of event parameter. Based on the title information of the target chapter, determine the event name of the target chapter; Obtain the domain information of the target text, and determine the event type of the target text based on the domain information; The event extraction result of the target chapter is composed of the event type, event name, and integrated parameter values ​​of each type of event.

2. The method according to claim 1, characterized in that, The event parameter type template set is a preset event parameter type template corresponding to the event type of the target chapter; Alternatively, the event parameter type template set is a preset unified event parameter type template corresponding to each event type.

3. The method according to claim 1, characterized in that, The step of sequentially extracting the parameter values ​​of each type of event parameter in each sentence of the sentence sequence includes: A pre-trained sequence labeling model is used to label the event parameters of each type in each sentence of the sentence sequence, and the parameter values ​​of each type of event parameter are obtained. The sequence labeling model is trained using training sentences labeled with event parameter type tags to which each word in the sentence belongs as training data.

4. The method according to claim 1, characterized in that, The process of determining the event name of the target chapter based on its title information includes: The title information of the target chapter is used as the event name of the target chapter; or, Extract the theme from the title information of the target chapter and use it as the event name of the target chapter.

5. The method according to claim 1, characterized in that, The process of determining the event type of the target text based on the domain information of the target text includes: In the set event type template, select the target event type that is closest to the domain information of the target chapter, and use it as the event type of the target chapter; The event type template contains several defined event types.

6. The method according to any one of claims 1-5, characterized in that, The number of target chapters is multiple, and after obtaining the event extraction results for each target chapter, the following is also included: Based on the event extraction results of multiple target chapters, event correlation analysis is performed from semantic and / or statistical dimensions.

7. The method according to claim 6, characterized in that, Based on the event extraction results from multiple target chapters, event association analysis is performed from a semantic dimension, including: Using event type as the classification condition, based on the event extraction results of multiple target chapters, the multiple target chapters are divided into at least one set of events of the same type; And / or, For each type of event parameter, calculate the semantic similarity of the integrated parameter values ​​of the event parameters of that type for each pair of target chapters; It is determined that two target texts with a semantic similarity exceeding a set similarity threshold have a semantic association on the type of event parameter.

8. The method according to claim 6, characterized in that, Based on the event extraction results from multiple target chapters, event correlation analysis is performed from a statistical perspective, including: For any two target chapters, compare the integrated parameter values ​​of each type of event parameter to see if they are the same. Based on the comparison results, it was determined that two target texts satisfying the first condition have a co-occurrence relationship in terms of events, and that two target texts satisfying the second condition have a sequential relationship in terms of events, wherein: The first condition is that the integrated parameter values ​​of the time parameter and the location parameter in the event extraction results of the two target chapters are the same. The second condition is that the integrated parameter values ​​of the character parameter and the location parameter in the event extraction results of the two target chapters are the same, while the integrated parameter values ​​of the time parameter are different.

9. A document-level event extraction device, characterized in that, include: Sentence segmentation units are used to segment the target text into sentences to obtain a sentence sequence; An event parameter extraction unit is used to extract the parameter values ​​of each type of event parameter in each sentence of the sentence sequence according to a set event parameter type template. The event parameter type template contains several set event parameters of different types. The event parameter integration unit is used to, for each type of event parameter in the event parameter type template, sequentially combine the parameter values ​​of the type of event parameter extracted from each sentence according to the sorting order of each sentence in the sentence sequence to obtain the integrated parameter value of the type of event parameter. If the parameter value of the type of event parameter extracted from a certain sentence is empty, then a preset character is used to replace the parameter value of the type of event parameter. The event name determination unit is used to determine the event name of the target chapter based on the title of the target chapter; An event type determination unit is used to obtain the domain information of the target chapter and determine the event type of the target chapter based on the domain information; The event extraction result determination unit is used to compose the event extraction result of the target chapter from the event type, event name and integrated parameter values ​​of each type of event parameter.

10. A document-level event extraction device, characterized in that, include: Memory and processor; The memory is used to store programs; The processor is used to execute the program to implement each step of the chapter-level event extraction method as described in any one of claims 1 to 7.

11. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements each step of the chapter-level event extraction method as described in any one of claims 1 to 7.

Images