Event extraction method and device, electronic equipment and computer readable medium

Abstract

The present disclosure relates to an event extraction method, device, electronic equipment and computer readable medium, and belongs to the technical field of natural language processing. The method comprises: performing core event extraction on original event text to obtain first model input data; inputting the first model input data into a first language representation model to identify an event trigger word, embedding the trigger word into a vector, performing multi-task learning on the trigger word embedded vector through a multi-task classifier to obtain the event trigger word and a first auxiliary task result; obtaining second model input data according to the trigger word embedded vector, the event trigger word and multimedia information data related to the original event text; inputting the second model input data into a second language representation model to identify an event element, embedding the element into a vector, performing multi-task learning on the element embedded vector through the multi-task classifier to obtain the event element and a second auxiliary task result. The present disclosure can realize comprehensive extraction of event information by using multi-task learning.

Description

Technical Field

[0001] This disclosure relates to the field of natural language processing technology, and more specifically, to an event extraction method, an event extraction device, an electronic device, and a computer-readable medium. Background Technology

[0002] Event extraction, which automatically extracts event elements of interest to users from unstructured natural language text, is a crucial component of information extraction. Currently, the level of event extraction is limited to processing event texts in a single language and domain; research on cross-language and multi-domain applications is largely in the exploratory stage. For example, research on Chinese event extraction mainly focuses on biomedicine, microblogs, news, and emergency situations. For domains containing large amounts of comprehensive information, such as short videos, cross-language event extraction research remains severely lacking.

[0003] Current event extraction methods suffer from drawbacks including high template construction costs, poor adaptability of technical solutions, and mechanically complex event description generation. Specifically, current methods require extensive manual construction for specific domains, necessitating the pre-construction of initial trigger word sets and the design of event category templates containing trigger words and event argument roles for each category's trigger word set. These tasks require significant time investment and prior knowledge from domain experts. Predefined event types and schemas are often suitable for specific datasets but are difficult to adapt effectively in real-world applications, especially in the open-domain short video ecosystem, resulting in high migration costs. Furthermore, generating structured event descriptions based on templates is rather mechanical and lacks diversity.

[0004] Therefore, there is an urgent need in this field for an event extraction method that can effectively extract event trigger words and elements, avoid large template construction costs, and improve the adaptability of technical solutions.

[0005] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this disclosure is to provide an event extraction method, event extraction device, electronic device, and computer-readable medium, which can at least to some extent effectively extract event trigger words and elements.

[0007] According to a first aspect of this disclosure, an event extraction method is provided, comprising:

[0008] The original event text is subjected to core event extraction to obtain the corresponding core event text, and the first model input data is obtained based on the core event text; wherein, the core event is the main event obtained after removing noise from the original event.

[0009] The first model input data is input into the first language representation model to identify event trigger words. The first language representation model is used to obtain the trigger word embedding vector. The trigger word embedding vector is then passed through the multi-task classifier corresponding to the first language representation model for multi-task learning to obtain the extracted event trigger words and the first auxiliary task result.

[0010] The second model input data is obtained based on the embedded vector of the trigger word, the event trigger word, and multimedia information data related to the original event text;

[0011] The second model input data is input into the second language representation model for event element identification. The second language representation model obtains the element embedded vector, and the element embedded vector is passed through the multi-task classifier corresponding to the second language representation model for multi-task learning to obtain the extracted event elements and the second auxiliary task results.

[0012] In one exemplary embodiment of this disclosure, the step of extracting core events from the original event text to obtain the corresponding core event text includes:

[0013] Obtain the model hint information corresponding to the core event extraction task, and input the model hint information corresponding to the core event extraction task and the original event text into the large language model to extract the core event;

[0014] Based on the core event extraction results output by the large language model, the core event text corresponding to the original event text is obtained.

[0015] In one exemplary embodiment of this disclosure, obtaining the first model input data based on the core event text includes:

[0016] The core event text and the original event text are combined to obtain the first model input data.

[0017] In one exemplary embodiment of this disclosure, the step of performing multi-task learning on the embedded vector of the trigger word through a multi-task classifier corresponding to the first language representation model to obtain the extracted event trigger word and the first auxiliary task result includes:

[0018] The embedded vector of the trigger word is input into the multi-task classifier corresponding to the first language representation model. Through the event trigger word recognition task and the event classification task in the multi-task classifier, the extracted event trigger words and event classification results are obtained.

[0019] In one exemplary embodiment of this disclosure, obtaining the second model input data based on the trigger word embedded vector, the event trigger word, and multimedia information data related to the original event text includes:

[0020] The hidden state vector and position encoding vector corresponding to the embedded vector of the trigger word are combined to obtain the first input data;

[0021] Trigger word prompt information is obtained based on the event trigger word, and second input data is obtained based on the trigger word prompt information, the core event text, and multimedia information data related to the original event text;

[0022] The first input data and the second input data are combined to obtain the second model input data.

[0023] In one exemplary embodiment of this disclosure, obtaining trigger word prompt information based on the event trigger word includes:

[0024] Based on the event trigger words, the corresponding subject prompt information and object prompt information are obtained respectively.

[0025] In one exemplary embodiment of this disclosure, the step of performing multi-task learning on the embedded vector of the element through a multi-task classifier corresponding to the second language representation model to obtain the extracted event elements and the second auxiliary task result includes:

[0026] The embedded vector of the element is input into the multi-task classifier corresponding to the second language representation model. Through the event element identification task and event summary generation task in the multi-task classifier, the extracted event elements and the event summary generation results are obtained.

[0027] According to a second aspect of this disclosure, an event extraction apparatus is provided, comprising:

[0028] The core event extraction module is configured to extract core events from the original event text, obtain the corresponding core event text, and obtain the first model input data based on the core event text; wherein, the core event is the main event obtained after removing noise from the original event.

[0029] The event trigger word extraction module is configured to input the first model input data into the first language representation model to identify event trigger words, obtain the trigger word embedded vector through the first language representation model, and perform multi-task learning on the trigger word embedded vector through the multi-task classifier corresponding to the first language representation model to obtain the extracted event trigger words and the first auxiliary task result.

[0030] The input data construction module is configured to generate second model input data based on the embedded vector of the trigger word, the event trigger word, and multimedia information data related to the original event text.

[0031] The event element extraction module is configured to input the second model input data into the second language representation model to identify event elements, obtain element embedded vectors through the second language representation model, and perform multi-task learning on the element embedded vectors through the multi-task classifier corresponding to the second language representation model to obtain the extracted event elements and the second auxiliary task results.

[0032] In one exemplary embodiment of this disclosure, the core event extraction module includes:

[0033] The model prompt information acquisition unit is configured to acquire the model prompt information corresponding to the core event extraction task, and input the model prompt information corresponding to the core event extraction task and the original event text into the large language model for core event extraction;

[0034] The core event text extraction unit is configured to extract the core event text corresponding to the original event text by performing the core event extraction result output by the large language model.

[0035] In one exemplary embodiment of this disclosure, the core event extraction module further includes:

[0036] The first model input data determination unit is configured to combine the core event text and the original event text to obtain the first model input data.

[0037] In one exemplary embodiment of this disclosure, the event trigger word extraction module includes:

[0038] The trigger word recognition task execution unit is configured to input the embedded vector of the trigger word into the multi-task classifier corresponding to the first language representation model, and obtain the extracted event trigger word and event classification result through the event trigger word recognition task and event classification task in the multi-task classifier.

[0039] In one exemplary embodiment of this disclosure, the input data construction module includes:

[0040] The first input data determining unit is configured to combine the hidden state vector and the position encoding vector corresponding to the embedded vector of the trigger word to obtain the first input data;

[0041] The second input data determination unit is configured to perform the following: obtain trigger word prompt information based on the event trigger word, and obtain second input data based on the trigger word prompt information, the core event text, and multimedia information data related to the original event text.

[0042] The second model input data determination unit is configured to combine the first input data and the second input data to obtain the second model input data.

[0043] In one exemplary embodiment of this disclosure, the second model input data determining unit includes:

[0044] The trigger word prompt information determination unit is configured to obtain the corresponding subject prompt information and object prompt information based on the event trigger word.

[0045] In one exemplary embodiment of this disclosure, the event element extraction module includes:

[0046] The feature recognition task execution unit is configured to input the embedded vector of the feature into the multi-task classifier corresponding to the second language representation model, and obtain the extracted event features and the event summary generation results through the event feature recognition task and the event summary generation task in the multi-task classifier.

[0047] According to a third aspect of this disclosure, an electronic device is provided, comprising: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to execute the instructions to implement the event extraction method described in any of the preceding claims.

[0048] According to a fourth aspect of this disclosure, a computer-readable storage medium is provided that, when instructions in the computer-readable storage medium are executed by a processor of an electronic device, enables the electronic device to perform the event extraction method described in any of the preceding claims.

[0049] According to a fifth aspect of this disclosure, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the event extraction method described in any of the preceding claims.

[0050] The exemplary embodiments disclosed herein can have the following beneficial effects:

[0051] In the event extraction method of this exemplary embodiment, core events are extracted from the original event text, and first model input data is obtained based on the core event text. The first model input data is then input into a first language representation model for event trigger word recognition. The extracted event trigger words and the first auxiliary task result are obtained through a multi-task classifier corresponding to the first language representation model. Next, second model input data is obtained based on the trigger word embedded vector, the event trigger words, and multimedia information data related to the original event text. The second model input data is then input into a second language representation model for event element recognition. The extracted event elements and the second auxiliary task result are obtained through a multi-task classifier corresponding to the second language representation model. The event extraction method in this exemplary embodiment, on the one hand, strengthens the input text and reduces the impact of noise on event information extraction by extracting core events during the model input stage; on the other hand, the multi-task classifier optimizes the event trigger word extraction task and the event element extraction task, while supporting the expansion of other tasks; furthermore, by combining multimedia information data related to the original event text as auxiliary event elements, the original event element extraction task can be strengthened.

[0052] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0053] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0054] Figure 1 The schematic diagram illustrates the structure of an event extraction system according to a related embodiment of the present disclosure.

[0055] Figure 2 A flowchart illustrating an example embodiment of the event extraction method of this disclosure is shown.

[0056] Figure 3 A schematic diagram illustrating the flow of input data for constructing a model in an exemplary embodiment of this disclosure is shown.

[0057] Figure 4 The schematic diagram illustrates the structure of an event extraction system according to a specific embodiment of the present disclosure.

[0058] Figure 5 A block diagram of an event extraction apparatus according to an exemplary embodiment of the present disclosure is shown.

[0059] Figure 6 A schematic diagram of the structure of a computer system suitable for implementing the embodiments of the present disclosure is shown. Detailed Implementation

[0060] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0061] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein.

[0062] The following exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this disclosure more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a full understanding of embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure can be practiced with one or more of the specific details omitted, or other methods, components, apparatus, steps, etc., can be employed. In other instances, well-known technical solutions are not shown or described in detail to avoid obscuring various aspects of this disclosure.

[0063] Furthermore, the accompanying drawings are merely illustrative of this disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0064] In some relevant embodiments, Multi2OIE (Multilingual Open Information Extraction Based on Multi-Head Attention with BERT) can be used for event extraction. Multi2OIE is an open-domain event extraction method that does not require pre-defining event types and event schemas.

[0065] Figure 1 The diagram illustrates the structure of an event extraction system according to a related embodiment of this disclosure. This scheme treats information extraction as a sequence labeling task. Event extraction is divided into event trigger word extraction and event argument extraction, using a BERT (Bidirectional Encoder Representations from Transformers) model based on multi-head attention. The input is a text segment, and the outputs are sequence labels for event trigger words and event arguments based on BIO annotations. BIO annotation refers to labeling each element as "BX", "IX", or "O". "BX" indicates that the segment containing this element belongs to type X and the element is at the beginning of the segment; "IX" indicates that the segment containing this element belongs to type X and the element is in the middle of the segment; and "O" indicates that it does not belong to any type. This method can achieve multilingual event extraction based on BERT models for different languages.

[0066] like Figure 1 As shown, the loss function for training this model can include two parts: the predicate verb part of the extracted triples and the subject-object part of the triples.

[0067] Extract the predicate verb of the triplet in Figure 1 In the left half of the graph, the input text is first encoded using BERT, and the encoded hidden vector is input into an intermediate layer, which is used to predict the label of the predicate verb.

[0068] The main guest of the three-element combination is drawn. Figure 1 The right half of the input consists of three parts: the BERT encoding result of the first part, the average vector of the BERT encoding of the predicate verb, and the positional encoding vector indicating whether to perform prediction (if it is a predicate verb, the second part does not need to be predicted). The overall structure of the middle processing is similar to that of a transformer.

[0069] However, the above event extraction scheme has the following drawbacks:

[0070] (1) This scheme is mainly designed for processing open domain event extraction datasets and is not suitable for scenarios with high text noise, such as short video scenarios in actual business.

[0071] (2) Only one event trigger word can be extracted at a time. If there are multiple trigger words, the process needs to be repeated multiple times.

[0072] (3) It is only applicable to the task of extracting event trigger words / event elements. It is difficult to support other tasks by transferring the model.

[0073] To address the aforementioned issues, this example implementation first provides an event extraction method. (Reference) Figure 2 As shown, the above event extraction method may include the following steps:

[0074] Step S210. Extract the core events from the original event text to obtain the corresponding core event text, and obtain the first model input data based on the core event text.

[0075] Among them, the core event is the main event obtained after removing noise from the original event.

[0076] Step S220. Input the first model input data into the first language representation model to identify event trigger words, obtain the trigger word embedding vector through the first language representation model, and perform multi-task learning on the trigger word embedding vector through the multi-task classifier corresponding to the first language representation model to obtain the extracted event trigger words and the first auxiliary task results.

[0077] Step S230. Obtain the second model input data based on the embedded vector of the trigger word, the event trigger word, and the multimedia information data related to the original event text.

[0078] Step S240. Input the second model input data into the second language representation model to identify event elements, obtain the element embedded vector through the second language representation model, and perform multi-task learning on the element embedded vector through the multi-task classifier corresponding to the second language representation model to obtain the extracted event elements and the second auxiliary task results.

[0079] In the event extraction method of this exemplary embodiment, core events are extracted from the original event text, and first model input data is obtained based on the core event text. The first model input data is then input into a first language representation model for event trigger word recognition. The extracted event trigger words and the first auxiliary task result are obtained through a multi-task classifier corresponding to the first language representation model. Next, second model input data is obtained based on the trigger word embedded vector, the event trigger words, and multimedia information data related to the original event text. The second model input data is then input into a second language representation model for event element recognition. The extracted event elements and the second auxiliary task result are obtained through a multi-task classifier corresponding to the second language representation model. The event extraction method in this exemplary embodiment, on the one hand, strengthens the input text and reduces the impact of noise on event information extraction by extracting core events during the model input stage; on the other hand, the multi-task classifier optimizes the event trigger word extraction task and the event element extraction task, while supporting the expansion of other tasks; furthermore, by combining multimedia information data related to the original event text as auxiliary event elements, the original event element extraction task can be strengthened.

[0080] Below, in conjunction with Figures 3 to 4 The steps described above in this example implementation will be explained in more detail.

[0081] In step S210, the core event is extracted from the original event text to obtain the corresponding core event text, and the first model input data is obtained based on the core event text.

[0082] In this example implementation, the core event refers to the main event obtained after removing relevant noise from the original event. In scenarios with high text noise, such as short video scenarios, considering the excessive noise in short video text, there is a lot of descriptive information besides the core event description, which will introduce significant input noise to model training. Therefore, a large model can be used to purify the core events during the model's input stage.

[0083] In this example implementation, model hints corresponding to the core event extraction task can be obtained, and these hints, along with the original event text, are input into a large language model for core event extraction. Based on the core event extraction results output by the large language model, the core event text corresponding to the original event text is obtained. Extracting core events can reduce noise in the input text.

[0084] The model prompts for the core event extraction task are based on the prompts set by the large language model for the task. These prompts, along with the original event text, are input into the large language model to extract the core event text corresponding to the original event text. The large language model can be, for example, ChatGPT-3.5 (Chat Generative Pre-trained Transformer) or a similar large model.

[0085] After obtaining the core event text corresponding to the original event text, the core event text and the original event text can be combined to obtain the input data for the first model.

[0086] After refining the event text using large model prompting learning, the core event text can be used together with the original event text as input to strengthen the input text.

[0087] In step S220, the first model input data is input into the first language representation model to identify event trigger words. The first language representation model is used to obtain the embedded vector of the trigger word, and the embedded vector of the trigger word is passed through the multi-task classifier corresponding to the first language representation model for multi-task learning to obtain the extracted event trigger words and the results of the first auxiliary task.

[0088] In this example implementation, the first language representation model is used for the identification and extraction of event trigger words. The first language representation model can be the XLM-Roberta (Unsupervised Cross-lingual Representation Learning at Scale) Encoder, or other multilingual deep learning models, such as XLM (Cross-lingual Language Model Pretraining)-BERT, etc., and this example implementation does not impose any specific limitations.

[0089] In this example implementation, the embedded vector of the trigger word can be input into the multi-task classifier corresponding to the first language representation model. Through the event trigger word recognition task and the event classification task in the multi-task classifier, the extracted event trigger words and event classification results can be obtained.

[0090] Input the purified core event text together with the original event text into a first language representation model, for example, XLM-Roberta Encoder, to convert them into corresponding trigger word embedding vectors, then perform multi-task learning through a MultiTask Head. While completing the trigger word recognition task, other auxiliary tasks are extended, such as event classification task or other possible auxiliary tasks, and then the extracted event trigger words and the first auxiliary task result, for example, the event classification result, are obtained.

[0091] The multi-task classifier can optimize the event trigger word extraction task, support the expansion of other tasks such as event topic multi-classification task, and can also optimize the event trigger word recognition task in the training stage, so as to improve the recognition performance of the model. Event topics in the event topic multi-classification task may include: politics, sports, finance, technology and other similar Hetu categories. The event trigger word extraction task can adopt the extraction of event trigger words based on BIO sequence labeling.

[0092] In step S230, second model input data is obtained according to the trigger word embedding vectors, the event trigger words, and the multimedia information data related to the original event text.

[0093] In this exemplary embodiment, the text understanding capability of a large model can be utilized to generate auxiliary event elements through prompt learning, so as to enhance the original event element extraction.

[0094] In this exemplary embodiment, the multimedia information data refers to audio and video information data related to the original event text. Taking short videos as an example, compared with formal news headlines and text content, the headlines and description texts of short videos have strong randomness and poor standardization from creators, resulting in incomplete event text description of short videos. Therefore, multimedia information data can be obtained according to the titles of short videos, OCR (Optical Character Recognition) recognition results, and ASR (Automatic Speech Recognition) recognition results of the video, and the data is input as Context Input.

[0095] In this exemplary embodiment, as Figure 3 shown, obtaining the second model input data according to the trigger word embedding vectors, the event trigger words, and the multimedia information data related to the original event text may specifically include the following steps:

[0096] Step S310: combining the hidden state vectors corresponding to the trigger word embedding vectors and the position encoding vectors to obtain first input data.

[0097] In this example implementation, the second model input data can mainly include two parts. The first part is input embeddings, which are the hidden state vectors corresponding to the trigger word embedding vectors. These mainly include the hidden states of the intermediate layer vectors in the process of extracting the event trigger word, as well as the position encoding vectors of the first language representation model output.

[0098] Step S320. Obtain trigger word prompt information based on the event trigger word, and obtain second input data based on the trigger word prompt information, core event text, and multimedia information data related to the original event text.

[0099] The other part of the input data for the second model is text input, which mainly consists of the purified core event text, multimedia information data related to the original event text, and trigger word prompts set based on the trigger words extracted in the aforementioned steps.

[0100] In this example implementation, corresponding subject and object prompts can be obtained based on the event trigger words. The subject prompts are used to help the model extract the text subject, while the object prompts are used to help the model extract the text object. By setting appropriate event trigger words, precise extraction of the desired text content can be achieved.

[0101] Step S330. Combine the first input data and the second input data to obtain the second model input data.

[0102] Finally, by combining the input data from the two parts mentioned above, we can obtain the input data for the second model.

[0103] In this example implementation, by combining multimedia information data related to the original event text as auxiliary event elements, the original event element extraction task can be enhanced. Simultaneously, by utilizing large model prompting learning to enhance event element extraction, the understanding capabilities of the large model can be leveraged to obtain its output through questioning, which can then be used as an input and merged with the existing process, thereby further enhancing the original event element extraction task.

[0104] Continue to refer to Figure 2 As shown, in step S240, the second model input data is input into the second language representation model to identify event elements. The second language representation model is used to obtain the element embedded vector, and the element embedded vector is used to perform multi-task learning through the multi-task classifier corresponding to the second language representation model to obtain the extracted event elements and the second auxiliary task results.

[0105] In this example implementation, the second language representation model is used for extracting event elements. The second language representation model can be generated using mBART (Multilingual Bert) or other multilingual generative models, such as m-T5 (multilingual T5), etc. This example implementation does not impose any specific limitations.

[0106] In this example implementation, the embedded vector of the element can be input into the multi-task classifier corresponding to the second language representation model. Through the event element identification task and the event summary generation task in the multi-task classifier, the extracted event elements and the event summary generation results can be obtained.

[0107] This study utilizes a second language representation model, such as the mBART model, for multi-task learning of event element generation. The input data is fed into the second language representation model and converted into corresponding element embedding vectors. A multi-task classifier is then used for multi-task learning. While completing the event element recognition task, other auxiliary tasks are extended, such as event summarization or other possible tasks. The result is the extracted event elements and the generated event summary. The event element extraction task predicts the position of event elements in the original text using spans (position labels), while the event summarization task generates an event summary based on extracted event trigger words and video text. Both tasks are suitable for generative models. A multi-task classifier can optimize the event element extraction task and support the extension of other tasks, such as event summarization. Furthermore, the event element recognition task can be optimized during the training phase to improve the model's recognition performance.

[0108] Figure 4 The schematic diagram illustrates the structure of an event extraction system according to a specific embodiment of the present disclosure. This event extraction system is used in the field of short videos and mainly includes the following four main modules: a short video event text purification module based on a large model, a trigger word recognition module based on multi-task learning, a short video event element extraction and enhancement module based on a large model, and an event element generation module based on multi-task learning.

[0109] The short video event text purification module based on a large model is used to purify the core events during the model's input stage, and then use them together with the original text as input, thereby strengthening the input text.

[0110] The short video event element extraction enhancement module based on a large model leverages the text understanding capabilities of the large model to generate auxiliary event elements through cue learning, thereby enhancing the original event element extraction module. This module can use the short video title, OCR recognition results, and ASR video results as context input, combined with the intermediate layer vectors of the event trigger word extraction model, the refined core events, and cue words set based on the trigger words as common input.

[0111] The trigger word recognition module and the event element generation module for multi-task learning mainly complete the learning tasks of event type, trigger word, event element, and event summary, and can also be extended to other types of learning tasks.

[0112] The technology proposed in this solution can be used in the following scenarios:

[0113] (1) Event clustering and context mining: After extracting basic events, news context can be generated through clustering and event elements;

[0114] (2) Event Hot List Generalization: By aggregating events at different levels, we can provide more related videos for a particular video within an event, thereby supporting different businesses. For example, after some S-level events occur, the horizontal bar generalization capability based on hot events can quickly expand the penetration rate of the hot list.

[0115] It should be noted that although the steps of the method in this disclosure are described in a specific order in the accompanying drawings, this does not require or imply that the steps must be performed in that specific order, or that all the steps shown must be performed to achieve the desired result. Additional or alternative steps may be omitted, multiple steps may be combined into one step, and / or a step may be broken down into multiple steps.

[0116] Furthermore, this disclosure also provides an event extraction apparatus. (See reference...) Figure 5 As shown, the event extraction device may include a core event extraction module 510, an event trigger word extraction module 520, an input data construction module 530, and an event element extraction module 540. Wherein:

[0117] The core event extraction module 510 is configured to extract core events from the original event text, obtain the corresponding core event text, and obtain the first model input data based on the core event text; wherein, the core event is the main event obtained after removing noise from the original event.

[0118] The event trigger word extraction module 520 is configured to input the first model input data into the first language representation model to identify event trigger words, obtain the trigger word embedding vector through the first language representation model, and perform multi-task learning by passing the trigger word embedding vector through the multi-task classifier corresponding to the first language representation model to obtain the extracted event trigger words and the first auxiliary task results;

[0119] The input data construction module 530 is configured to obtain the second model input data based on the embedded vector of the trigger word, the event trigger word, and multimedia information data related to the original event text;

[0120] The event element extraction module 540 is configured to input the second model input data into the second language representation model to identify event elements, obtain element embedded vectors through the second language representation model, and perform multi-task learning on the element embedded vectors through the multi-task classifier corresponding to the second language representation model to obtain the extracted event elements and the second auxiliary task results.

[0121] In some exemplary embodiments of this disclosure, the core event extraction module 510 may include a model prompt information acquisition unit and a core event text extraction unit. Wherein:

[0122] The model prompt information acquisition unit is configured to acquire the model prompt information corresponding to the core event extraction task, and input the model prompt information corresponding to the core event extraction task and the original event text into the large language model for core event extraction;

[0123] The core event text extraction unit is configured to obtain the core event text corresponding to the original event text based on the core event extraction results output by the large language model.

[0124] In some exemplary embodiments of this disclosure, the core event extraction module 510 may further include a first model input data determination unit, configured to combine the core event text and the original event text to obtain the first model input data.

[0125] In some exemplary embodiments of this disclosure, the event trigger word extraction module 520 may include a trigger word recognition task execution unit, which is configured to input the embedded vector of the trigger word into a multi-task classifier corresponding to the first language representation model, and obtain the extracted event trigger words and event classification results through the event trigger word recognition task and event classification task in the multi-task classifier.

[0126] In some exemplary embodiments of this disclosure, the input data construction module 530 may include a first input data determination unit, a second input data determination unit, and a second model input data determination unit. Wherein:

[0127] The first input data determining unit is configured to combine the hidden state vector corresponding to the trigger word embedded vector and the position encoding vector to obtain the first input data;

[0128] The second input data determination unit is configured to obtain trigger word prompt information based on the event trigger word, and to obtain second input data based on the trigger word prompt information, the core event text, and multimedia information data related to the original event text;

[0129] The second model input data determination unit is configured to combine the first input data and the second input data to obtain the second model input data.

[0130] In some exemplary embodiments of this disclosure, the second input data determining unit may include a trigger word prompt information determining unit, configured to obtain corresponding subject prompt information and object prompt information according to the event trigger word.

[0131] In some exemplary embodiments of this disclosure, the event element extraction module 540 may include an element recognition task execution unit, configured to input the embedded vector of the element into a multi-task classifier corresponding to the second language representation model, and obtain the extracted event elements and the event summary generation result through the event element recognition task and the event summary generation task in the multi-task classifier.

[0132] The specific details of each module / unit in the above-mentioned event extraction device have been described in detail in the corresponding method embodiment section, and will not be repeated here.

[0133] Figure 6 A schematic diagram of the structure of a computer system suitable for implementing the embodiments of the present disclosure is shown.

[0134] It should be noted that, Figure 6 The computer system 600 of the electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0135] like Figure 6 As shown, the computer system 600 includes a central processing unit (CPU) 601, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 602 or programs loaded from storage section 608 into random access memory (RAM) 603. The RAM 603 also stores various programs and data required for system operation. The CPU 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0136] The following components are connected to I / O interface 605: an input section 606 including a keyboard, mouse, etc.; an output section 607 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN card, modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to I / O interface 605 as needed. A removable medium 611, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 610 as needed so that computer programs read from it can be installed into storage section 608 as needed.

[0137] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 609, and / or installed from removable medium 611. When the computer program is executed by central processing unit (CPU) 601, it performs various functions defined in the system of this disclosure.

[0138] It should be noted that the computer-readable medium disclosed herein may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0139] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0140] In another aspect, this disclosure also provides a computer-readable medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The computer-readable medium carries one or more programs that, when executed by the electronic device, cause the electronic device to perform the methods described in the above embodiments.

[0141] It should be noted that although several modules for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more modules described above can be embodied in one module. Conversely, the features and functions of one module described above can be further divided and embodied by multiple modules.

[0142] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein.

[0143] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. An event extraction method, characterized in that, include: The original event text is subjected to core event extraction to obtain the corresponding core event text, and the first model input data is obtained based on the core event text; wherein, the core event is the main event obtained after removing noise from the original event. The first model input data is input into the first language representation model to identify event trigger words. The first language representation model is used to obtain the trigger word embedding vector. The trigger word embedding vector is then passed through the multi-task classifier corresponding to the first language representation model for multi-task learning to obtain the extracted event trigger words and the first auxiliary task result. The second model input data is obtained based on the embedded vector of the trigger word, the event trigger word, and multimedia information data related to the original event text; The second model input data is input into the second language representation model for event element identification. The second language representation model obtains the element embedded vector, and the element embedded vector is passed through the multi-task classifier corresponding to the second language representation model for multi-task learning to obtain the extracted event elements and the second auxiliary task results.

2. The event extraction method according to claim 1, characterized in that, The process of extracting core events from the original event text to obtain the corresponding core event text includes: Obtain the model hint information corresponding to the core event extraction task, and input the model hint information corresponding to the core event extraction task and the original event text into the large language model to extract the core event; Based on the core event extraction results output by the large language model, the core event text corresponding to the original event text is obtained.

3. The event extraction method according to claim 1, characterized in that, The process of obtaining the first model input data based on the core event text includes: The core event text and the original event text are combined to obtain the first model input data.

4. The event extraction method according to claim 1, characterized in that, The step of passing the embedded vector of the trigger word through a multi-task classifier corresponding to the first language representation model for multi-task learning to obtain the extracted event trigger word and the result of the first auxiliary task includes: The embedded vector of the trigger word is input into the multi-task classifier corresponding to the first language representation model. Through the event trigger word recognition task and the event classification task in the multi-task classifier, the extracted event trigger words and event classification results are obtained.

5. The event extraction method according to claim 1, characterized in that, The process of obtaining the second model input data based on the embedded vector of the trigger word, the event trigger word, and multimedia information data related to the original event text includes: The hidden state vector and the position encoding vector corresponding to the embedded vector of the trigger word are combined to obtain the first input data; Trigger word prompt information is obtained based on the event trigger word, and second input data is obtained based on the trigger word prompt information, the core event text, and multimedia information data related to the original event text; The first input data and the second input data are combined to obtain the second model input data.

6. The event extraction method according to claim 5, characterized in that, The step of obtaining trigger word prompt information based on the event trigger word includes: Based on the event trigger words, the corresponding subject prompt information and object prompt information are obtained respectively.

7. The event extraction method according to claim 1, characterized in that, The step of performing multi-task learning on the embedded vector of the element through the multi-task classifier corresponding to the second language representation model to obtain the extracted event elements and the second auxiliary task results includes: The embedded vector of the element is input into the multi-task classifier corresponding to the second language representation model. Through the event element identification task and event summary generation task in the multi-task classifier, the extracted event elements and the event summary generation results are obtained.

8. An event extraction device, characterized in that, include: The core event extraction module is configured to extract core events from the original event text, obtain the corresponding core event text, and obtain the first model input data based on the core event text; wherein, the core event is the main event obtained after removing noise from the original event. The event trigger word extraction module is configured to input the first model input data into the first language representation model to identify event trigger words, obtain the trigger word embedded vector through the first language representation model, and perform multi-task learning on the trigger word embedded vector through the multi-task classifier corresponding to the first language representation model to obtain the extracted event trigger words and the first auxiliary task result. The input data construction module is configured to generate second model input data based on the embedded vector of the trigger word, the event trigger word, and multimedia information data related to the original event text. The event element extraction module is configured to input the second model input data into the second language representation model to identify event elements, obtain element embedded vectors through the second language representation model, and perform multi-task learning on the element embedded vectors through the multi-task classifier corresponding to the second language representation model to obtain the extracted event elements and the second auxiliary task results.

9. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the event extraction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are executed by the processor of the electronic device, the electronic device is enabled to perform the event extraction method as described in any one of claims 1 to 7.

Images