Event detection method and device fusing multi-source information

Abstract

The invention provides an event detection method and device fusing multi-source information, and the method is characterized in that the method comprises the steps: S1, carrying out the preprocessing of a designated text and each existing text; s2, inputting the pre-processed specified text and an existing trigger word extraction template into a candidate trigger word extraction model together to obtain candidate trigger words corresponding to the pre-processed specified text; step S3-S6, performing text matching on the pre-processed specified text and each pre-processed existing text to judge whether a matched text exists or not, and if yes, executing the step S3-S6; if yes, taking a preprocessed existing text matched with the preprocessed specified text as a matched text, and combining with a candidate trigger word extraction model, a trigger word extraction model and a candidate trigger word prompt template to obtain a trigger word, and if not, taking the candidate trigger word as the trigger word and entering the step S7; and S7, inputting the trigger word into the trigger word classification model to obtain an event type. In a word, the method can generate an accurate event detection result.

Description

Technical Field

[0001] The present invention relates to an event detection method, and in particular to an event detection method and device integrating multi-source information. Background Art

[0002] With the development of the Internet and economy, people can obtain a large amount of relevant news information in various fields on the Internet, and the relevant event information in different fields in these news is very important information for researchers in this field. Therefore, information extraction tasks such as event detection play a very important role in the field of news events. As a subtask of event extraction, event detection aims to obtain event trigger words and corresponding event types contained in unstructured language texts, where event trigger words are words that mark the occurrence of an event.

[0003] However, existing related research pays little attention to the multi-source nature of texts. Most of them conduct event detection in the context of single-source information and focus on the information contained in the single-source text itself. Therefore, there is a problem of incomplete event detection leading to inaccurate detection results. Summary of the invention

[0004] The present invention is made to solve the above-mentioned problem, and aims to provide an event detection method and device integrating multi-source information.

[0005] The present invention provides an event detection method integrating multi-source information, which is used to obtain the event type of the specified event according to a specified text and multiple existing texts containing the specified event, and has the following characteristics, including the following steps: step S1, preprocessing the specified text and each existing text respectively to obtain the corresponding preprocessed specified text and preprocessed existing text; step S2, inputting the preprocessed specified text and the existing trigger word extraction template into a candidate trigger word extraction model to obtain the candidate trigger word corresponding to the preprocessed specified text; step S3, performing text matching on the preprocessed specified text and each preprocessed existing text to determine whether there is a matching text, and if so, matching the preprocessed specified text with the existing text. The text matching preprocesses the existing text as the matching text and enters step S4. If not, the candidate trigger word is used as the trigger word and enters step S7; Step S4, for each matching text, the matching text and the trigger word extraction template are input into the candidate trigger word extraction model together to obtain the candidate trigger word corresponding to the matching text; Step S5, a candidate trigger word prompt template is constructed according to the candidate trigger words corresponding to each matching text; Step S6, each matching text, the candidate trigger word prompt template and the trigger word extraction template containing the candidate trigger words corresponding to the preprocessed specified text are input into the trigger word extraction model together to obtain the trigger word; Step S7, the trigger word is input into the trigger word classification model to obtain the event type.

[0006] The event detection method for fusing multi-source information provided by the present invention may also have the following features: wherein, in step S2, the preprocessed specified text and the trigger word extraction template are spliced ​​and the splicing result is used as the input of the candidate trigger word extraction model; in step S4, the matching text and the trigger word extraction template are spliced ​​and the splicing result is used as the input of the candidate trigger word extraction model; and in step S6, all matching texts, candidate trigger word prompt templates and trigger word extraction templates containing candidate trigger words corresponding to the preprocessed specified text are spliced, and the splicing result is used as the input of the trigger word extraction model.

[0007] The event detection method for fusing multi-source information provided by the present invention may also have the following feature: wherein both the candidate trigger word extraction model and the trigger word extraction model are constructed based on the existing generative pre-trained language model.

[0008] The event detection method for fusing multi-source information provided by the present invention may also have the following feature: in step S5, the candidate trigger word prompt template includes all candidate trigger words corresponding to each matching text.

[0009] The event detection method for integrating multi-source information provided by the present invention may also have the following features: wherein, the trigger word classification model includes a BERT pre-trained language model and a multi-layer perceptron, the BERT pre-trained language model generates a hidden layer representation vector of a sentence according to the trigger word, and the multi-layer perceptron maps the hidden layer representation vector of the sentence to a label dimension to obtain an event type.

[0010] The event detection method for fusing multi-source information provided by the present invention may also have the following feature: wherein, in step S1, the preprocessing includes removing redundant non-text content in the text.

[0011] The event detection method for integrating multi-source information provided by the present invention may also have the following features: wherein, according to a plurality of existing training texts and their corresponding target candidate trigger words, target trigger words and event type labels, the process of training a candidate trigger word extraction model, a trigger word extraction model and a trigger word classification model comprises the following steps: step T1, preprocessing all existing training texts, and constructing a training data set in combination with the corresponding real trigger words and event type labels; step T2, selecting existing training texts from the training data set and inputting the candidate trigger word extraction model to generate predicted candidate trigger words, and fine-tuning the candidate trigger word extraction model based on the loss calculated based on the predicted candidate trigger words and the corresponding target candidate trigger words; step T3, repeating step T2 until the first preset termination condition is reached, then obtaining a trained candidate trigger word extraction model, and entering step T4; step T4 Step T4, select existing training text from the training data set, combine the trained candidate trigger word extraction model to build the corresponding input data input trigger word extraction model to generate predicted trigger words, and calculate the loss based on the predicted trigger words and the corresponding target trigger words to fine-tune the trigger word extraction model; Step T5, repeat step T4 until the second preset termination condition is reached, then the trained trigger word extraction model is obtained, and enter step T6; Step T6, select existing training text from the training data set, combine the trained candidate trigger word extraction model and the trained trigger word extraction model to build the corresponding input data input trigger word classification model to generate predicted labels, and calculate the loss based on the predicted labels and the corresponding event type labels to fine-tune the trigger word classification model; Step T7, repeat step T6 until the third preset termination condition is reached, then the trained trigger word classification model is obtained.

[0012] The present invention also provides an event detection device that integrates multi-source information, which is used to obtain the event type of the specified event based on a specified text containing the specified event and multiple existing texts, and has the following characteristics: a preprocessing module, which is used to preprocess the specified text and each existing text respectively, to obtain the corresponding preprocessed specified text and preprocessed existing text; a candidate trigger word generation module, which includes a candidate trigger word extraction model and an existing trigger word extraction template, and is used to input the text and the trigger word extraction template into the candidate trigger word extraction model together to obtain the candidate trigger word corresponding to the text, and the text includes the preprocessed specified text and the matching text; a matching judgment module, which is used to compare the preprocessed specified text with each existing text respectively. A preprocessed existing text is used for text matching to determine whether there is a matching text. If so, the preprocessed existing text that matches the preprocessed specified text is used as the matching text. If not, the candidate trigger word is used as the trigger word; a template construction module is used to construct a candidate trigger word prompt template according to the candidate trigger words corresponding to each matching text; a trigger word generation module includes a trigger word extraction model, which is used to input each matching text, the candidate trigger word prompt template and the trigger word extraction template containing the candidate trigger word corresponding to the preprocessed specified text into the trigger word extraction model to obtain the trigger word; an event type generation module includes a trigger word classification model, which is used to input the trigger word into the trigger word classification model to obtain the event type.

[0013] Functions and Effects of the Invention

[0014] According to the event detection method and device for fusing multi-source information involved in the present invention, because, on the one hand, a matching text having the same designated event as the designated text is selected from the existing text through text matching; on the other hand, a trigger word extraction template, a candidate trigger word prompt template, a candidate trigger word extraction model and a trigger word extraction model are used to combine the matching text and the designated text to generate a trigger word, and the trigger word is input into the trigger word classification model to obtain the corresponding event type. Therefore, the event detection method and device for fusing multi-source information of the present invention can generate accurate event detection results. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a block diagram of an event detection device in an embodiment of the present invention;

[0016] Figure 2 It is a flowchart of training a candidate trigger word extraction model, a trigger word extraction model and a trigger word classification model in an embodiment of the present invention;

[0017] Figure 3 It is a flowchart of an event detection method for fusing multi-source information in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the technical means, creative features, objectives and effects of the present invention easy to understand, the following embodiments and the accompanying drawings specifically illustrate the event detection method and device for fusing multi-source information of the present invention.

[0019] This embodiment provides an event detection device that integrates multi-source information, hereinafter referred to as an event detection device, for obtaining the event type of a specified event based on a specified text and multiple existing texts containing the specified event. The specified text and existing text corresponding to the event detection device are texts of corresponding fields selected according to actual needs.

[0020] Among the texts corresponding to many different fields, financial texts have more unique features than other types of texts. That is, in today's Internet era, financial texts often derive multiple different texts containing the same event during the dissemination process, that is, multi-source information. For example, multiple news media will produce multiple different reports on the same financial event, but the news events contained in them are consistent.

[0021] Therefore, in this embodiment, the event detection device of this embodiment is described by taking a financial event as a designated event, a news report containing the designated financial event as a designated text, and other multiple news reports as existing texts.

[0022] Figure 1 is a block diagram of an event detection device in an embodiment of the present invention.

[0023] like Figure 1 As shown, the event detection device 100 includes a preprocessing module 11, a candidate trigger word generation module 12, a matching judgment module 13, a template construction module 14, a trigger word generation module 15, an event type generation module 16 and a general control module 17 for controlling the above modules.

[0024] The preprocessing module 11 is used to preprocess the designated text and each existing text respectively to obtain the corresponding preprocessed designated text and preprocessed existing text.

[0025] The preprocessing includes removing the redundant contents of the text, ie, the designated text and each existing text, which are not the main text. In this embodiment, the preprocessing also includes converting the text into a format that conforms to the input form of the subsequent model.

[0026] The candidate trigger word generation module 12 includes a candidate trigger word extraction model and an existing trigger word extraction template, which is used to input the text and the trigger word extraction template into the candidate trigger word extraction model to obtain the candidate trigger words corresponding to the text, and the text includes the pre-processed specified text and the matching text.

[0027] When the text is a preprocessed specified text, the preprocessed specified text and the trigger word extraction template are spliced ​​and the splicing result is used as the input of the candidate trigger word extraction model. When the text is a matching text, the matching text and the trigger word extraction template are spliced ​​and the splicing result is used as the input of the candidate trigger word extraction model.

[0028] The format of the trigger word extraction template in this embodiment is "the event trigger word is <trigger>.",in" <trigger>" is a special mark that acts as a placeholder. In the splicing result of this embodiment, the text and the trigger word extraction template are separated by the "[SEP]" interval identifier. The output of the trigger word extraction template of this embodiment is "The event trigger word is <trigger>.",in" <trigger>" is a candidate trigger word, and when multiple candidate trigger words are generated, the multiple candidate trigger words are connected by "and".

[0029] The candidate trigger word extraction model is constructed based on the existing generative pre-trained language model. The generative pre-trained language model consists of two parts: an encoder and a decoder. The encoder is used to encode and model the input sequence, and the decoder is used to generate the target sequence according to the output of the encoder.

[0030] The matching judgment module 13 is used to perform text matching between the preprocessed designated text and each preprocessed existing text to determine whether there is a matching text. If so, the preprocessed existing text that matches the preprocessed designated text is used as the matching text. If not, the candidate trigger word is used as the trigger word.

[0031] In this embodiment, the matching judgment module 13 comprehensively evaluates whether the two texts contain the same event through text matching methods such as semantic matching or other feature matching. If so, the two texts are mutually matching texts, that is, there are matching texts. If not, there are no matching texts.

[0032] The template construction module 14 is used to construct a candidate trigger word prompt template according to the candidate trigger words corresponding to each matching text. In this embodiment, the candidate trigger word prompt template is "the event trigger word may be<TriggerFrom Same Event> .",in"<TriggerFrom Same Event> " is the corresponding candidate trigger word, and when there are multiple candidate trigger words corresponding to the matching text, the multiple candidate trigger words are connected by "and".

[0033] The candidate trigger word prompt template includes all candidate trigger words corresponding to each matching text.

[0034] The trigger word generation module 15 includes a trigger word extraction model, which is used to input each matching text, the candidate trigger word prompt template and the trigger word extraction template containing the candidate trigger word corresponding to the pre-processed specified text into the trigger word extraction model to obtain the trigger word.

[0035] Among them, all matching texts, candidate trigger word prompt templates and trigger word extraction templates containing candidate trigger words corresponding to the preprocessed specified text are spliced, and the splicing result is used as the input of the trigger word extraction model. In the splicing result of this embodiment, the matching text, candidate trigger word prompt template and trigger word extraction template are separated by the "[SEP]" interval identifier.

[0036] The trigger word extraction model is constructed based on the existing generative pre-trained language model.

[0037] The event type generation module 16 includes a trigger word classification model, which is used to input the trigger word into the trigger word classification model to obtain the event type.

[0038] The trigger word classification model includes the BERT pre-trained language model and a multi-layer perceptron. The BERT pre-trained language model generates a hidden layer representation vector of the sentence based on the trigger word. The multi-layer perceptron maps the hidden layer representation vector of the sentence to the label dimension to obtain the event type.

[0039] In this embodiment, the candidate trigger word extraction model, the trigger word extraction model and the trigger word classification model are trained in sequence according to the specified order.

[0040] Figure 2 It is a flowchart of training a candidate trigger word extraction model, a trigger word extraction model and a trigger word classification model in an embodiment of the present invention.

[0041] like Figure 2 As shown, the process of training a candidate trigger word extraction model, a trigger word extraction model and a trigger word classification model according to multiple existing training texts and their corresponding target candidate trigger words, target trigger words and event type labels includes the following steps:

[0042] Step T1: preprocess all existing training texts and construct a training dataset based on the corresponding real trigger words and event type labels.

[0043] Step T2, select existing training text from the training data set to input the candidate trigger word extraction model to generate predicted candidate trigger words, and calculate the loss based on the predicted candidate trigger words and the corresponding target candidate trigger words to fine-tune the candidate trigger word extraction model.

[0044] Step T3, repeat step T2 until the first preset termination condition is reached, then a trained candidate trigger word extraction model is obtained, and enter step T4.

[0045] Step T4, select existing training text from the training data set, combine the trained candidate trigger word extraction model to build the corresponding input data input trigger word extraction model to generate predicted trigger words, and calculate the loss based on the predicted trigger words and the corresponding target trigger words to fine-tune the trigger word extraction model.

[0046] Step T5, repeat step T4 until the second preset termination condition is reached, then a trained trigger word extraction model is obtained, and enter step T6.

[0047] Step T6, select existing training text from the training data set, combine the trained candidate trigger word extraction model and the trained trigger word extraction model to build the corresponding input data input trigger word classification model to generate prediction labels, and calculate the loss based on the prediction label and the corresponding event type label to fine-tune the trigger word classification model. The loss calculation used for each model in this embodiment is the cross entropy loss.

[0048] Step T7, repeat step T6 until the third preset termination condition is reached, and then a trained trigger word classification model is obtained. In this embodiment, the first preset termination condition, the second preset termination condition and the third preset termination condition are all whether the iteration round reaches the corresponding maximum iteration round.

[0049] The master control module 17 stores a control program for controlling the operation of each module.

[0050] The following describes the process of an event detection method for fusing multi-source information using the event detection device 100 in conjunction with the accompanying drawings.

[0051] Figure 3 It is a flowchart of an event detection method for fusing multi-source information in an embodiment of the present invention.

[0052] like Figure 3 As shown, the event detection method integrating multi-source information includes the following steps:

[0053] Step S1 : using the preprocessing module 11 to preprocess the designated text and each existing text respectively, to obtain the corresponding preprocessed designated text and preprocessed existing text.

[0054] Step S2, using the candidate trigger word generation module 12 to input the preprocessed designated text and the existing trigger word extraction template into the candidate trigger word extraction model to obtain the candidate trigger word corresponding to the preprocessed designated text.

[0055] In step S3, the matching judgment module 13 is used to perform text matching between the preprocessed specified text and each preprocessed existing text to determine whether there is a matching text. If so, the preprocessed existing text that matches the preprocessed specified text is used as the matching text and the process proceeds to step S4. If not, the candidate trigger word is used as the trigger word and the process proceeds to step S7.

[0056] Step S4, using the candidate trigger word generation module 12 for each matching text, inputting the matching text and the trigger word extraction template into the candidate trigger word extraction model together to obtain the candidate trigger word corresponding to the matching text.

[0057] Step S5: The template construction module 14 is used to construct a candidate trigger word prompt template according to the candidate trigger words corresponding to each matching text.

[0058] Step S6, using the trigger word generation module 15 to input each matching text, the candidate trigger word prompt template and the trigger word extraction template containing the candidate trigger word corresponding to the pre-processed designated text into the trigger word extraction model to obtain the trigger word.

[0059] Step S7: Use the event type generation module 16 to input the trigger word into the trigger word classification model to obtain the event type.

[0060] In this embodiment, existing data is used to test the performance of the event detection method of integrating multi-source information in this embodiment, namely Ours method, and the existing Text2Event method and DEGREE_ed method in the trigger word extraction task and event classification task, respectively. The results are shown in the following table:

[0061]

[0062]

[0063] In the above table, the first column is each task, the second column is each method, and the third to fifth columns are the calculation results of precision, recall and F1 of each method on each task. For example, the cell in the third column of the second row indicates that the precision of the Text2Event method on the trigger word extraction task is 72.97%. It can be seen that the event detection method that integrates multi-source information has better trigger word extraction and event classification effects than the existing methods.

[0064] Functions and Effects of the Embodiments

[0065] According to the event detection method and device for fusing multi-source information involved in this embodiment, on the one hand, a matching text having the same designated event as the designated text is selected from the existing text through text matching; on the other hand, a trigger word extraction template, a candidate trigger word prompt template, a candidate trigger word extraction model and a trigger word extraction model are used to combine the matching text and the designated text to generate a trigger word, and the trigger word is input into the trigger word classification model to obtain the corresponding event type. In short, this method can generate accurate event detection results.

[0066] Those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.< / trigger> < / trigger> < / trigger> < / trigger>

Claims

1. An event detection method integrating multi-source information, for obtaining the event type of a specified event according to a specified text containing the specified event and a plurality of existing texts, characterized in that: The following steps are involved: Step S1, preprocessing the designated text and each of the existing texts respectively to obtain corresponding preprocessed designated text and preprocessed existing text; Step S2, inputting the preprocessed designated text and the existing trigger word extraction template into a candidate trigger word extraction model to obtain candidate trigger words corresponding to the preprocessed designated text; Step S3, performing text matching between the preprocessed designated text and each of the preprocessed existing texts to determine whether there is a matching text. If so, the preprocessed existing text that matches the preprocessed designated text is used as the matching text and the process proceeds to step S4. If not, the candidate trigger word is used as the trigger word and the process proceeds to step S7. Step S4, for each matching text, inputting the matching text and the trigger word extraction template into the candidate trigger word extraction model to obtain the candidate trigger word corresponding to the matching text; Step S5, constructing a candidate trigger word prompt template according to the candidate trigger words corresponding to each of the matching texts; Step S6, inputting each of the matching texts, the candidate trigger word prompt template and the trigger word extraction template containing the candidate trigger word corresponding to the pre-processed designated text into a trigger word extraction model to obtain the trigger word; Step S7: input the trigger word into a trigger word classification model to obtain the event type.

2. The event detection method for fusing multi-source information according to claim 1 is characterized in that: in, In step S2, the preprocessed designated text and the trigger word extraction template are spliced ​​and the splicing result is used as the input of the candidate trigger word extraction model. In step S4, the matching text and the trigger word extraction template are concatenated and the concatenated result is used as the input of the candidate trigger word extraction model. In the step S6, all the matching texts, the candidate trigger word prompt templates and the trigger word extraction templates containing the candidate trigger words corresponding to the preprocessed specified texts are spliced, and the splicing result is used as the input of the trigger word extraction model.

3. The event detection method for fusing multi-source information according to claim 1 is characterized in that: in, The candidate trigger word extraction model and the trigger word extraction model are both constructed based on an existing generative pre-trained language model.

4. The event detection method for fusing multi-source information according to claim 1 is characterized in that: in, In the step S5, the candidate trigger word prompt template includes all the candidate trigger words corresponding to each of the matching texts.

5. The event detection method for fusing multi-source information according to claim 1 is characterized in that: in, The trigger word classification model includes a BERT pre-trained language model and a multi-layer perceptron. The BERT pre-trained language model generates a hidden layer representation vector of a sentence according to the trigger word, The multi-layer perceptron maps the hidden layer representation vector of the sentence to a label dimension to obtain the event type.

6. The event detection method for fusing multi-source information according to claim 1 is characterized in that: in, In the step S1, the preprocessing includes removing redundant non-text content in the text.

7. The event detection method for fusing multi-source information according to claim 1, Features: in, The process of training the candidate trigger word extraction model, the trigger word extraction model and the trigger word classification model according to a plurality of existing training texts and their corresponding target candidate trigger words, target trigger words and event type labels comprises the following steps: Step T1, preprocessing all the existing training texts, and constructing a training data set in combination with the corresponding real trigger words and the event type labels; Step T2, selecting the existing training text from the training data set and inputting it into the candidate trigger word extraction model to generate predicted candidate trigger words, and fine-tuning the candidate trigger word extraction model based on the loss calculated based on the predicted candidate trigger words and the corresponding target candidate trigger words; Step T3, repeating step T2 until the first preset termination condition is reached, then obtaining the trained candidate trigger word extraction model, and proceeding to step T4; Step T4, selecting the existing training text from the training data set, combining the trained candidate trigger word extraction model to construct the corresponding input data input into the trigger word extraction model to generate the predicted trigger word, and calculating the loss based on the predicted trigger word and the corresponding target trigger word to fine-tune the trigger word extraction model; Step T5, repeating step T4 until the second preset termination condition is reached, then obtaining the trained trigger word extraction model, and proceeding to step T6; Step T6, selecting the existing training text from the training data set, combining the trained candidate trigger word extraction model and the trained trigger word extraction model to construct the corresponding input data input into the trigger word classification model to generate a predicted label, and calculating the loss based on the predicted label and the corresponding event type label to fine-tune the trigger word classification model; Step T7, repeat step T6 until the third preset termination condition is reached, and the trained trigger word classification model is obtained.

8. An event detection device integrating multi-source information, used for obtaining the event type of a specified event based on a specified text containing the specified event and a plurality of existing texts, characterized in that: include: A preprocessing module, used for preprocessing the designated text and each of the existing texts respectively to obtain corresponding preprocessed designated text and preprocessed existing text; A candidate trigger word generation module, comprising a candidate trigger word extraction model and an existing trigger word extraction template, for inputting a text and the trigger word extraction template together into the candidate trigger word extraction model to obtain a candidate trigger word corresponding to the text, wherein the text comprises the preprocessed specified text and the matching text; A matching judgment module, used for performing text matching between the preprocessed designated text and each of the preprocessed existing texts to determine whether there is a matching text, if so, taking the preprocessed existing text that matches the preprocessed designated text as the matching text, if not, taking the candidate trigger word as the trigger word; A template construction module, used to construct a candidate trigger word prompt template according to the candidate trigger words corresponding to each of the matching texts; A trigger word generation module, comprising a trigger word extraction model, for inputting each of the matching texts, the candidate trigger word prompt templates and the trigger word extraction template containing the candidate trigger word corresponding to the pre-processed designated text into the trigger word extraction model to obtain the trigger word; The event type generation module includes a trigger word classification model, which is used to input the trigger word into the trigger word classification model to obtain the event type.

Images