Method for generating crisis help-seeking key information abstract based on intention detection

By comparing learning and graph attention networks, a highly relevant summary of key help-seeking information is generated, which solves the user intention identification problem in crisis events and improves the efficiency and accuracy of crisis response.

CN118939794APending Publication Date: 2024-11-12FUDAN UNIVERSITY
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202410996893.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-24
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

In crisis events, it is difficult for the existing technology to accurately and in-depth explore users' intentions for help, resulting in low efficiency in crisis management, overloading of rescue personnel information, and insufficient response capabilities.

Method used

Using the intent representation and classification method based on comparison learning, the help intention instance representation and category representation are generated through the BERT model, and a heterogeneous graph is constructed for graph attention network encoding, and a highly relevant help key information summary is generated using a two-stage decoder.

Benefits of technology

It improves the classification accuracy and summary generation quality of help-seeking intentions, enables rescuers and decision makers to respond more quickly and accurately, and improves the efficiency and accuracy of crisis response.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118939794A_ABST
    Figure CN118939794A_ABST
Patent Text Reader

Abstract

The invention discloses a crisis help-seeking key information abstract generation method based on intention detection. The method comprises the following steps: an intention characterization and classification stage: capturing similarity and difference between help seeking intention texts through comparative learning, and obtaining help seeking intention category representation through clustering so as to improve intention classification capability; in the abstract generation stage, an encoder-decoder is used as a basic framework, a heterogeneous graph composed of words, help information nodes and intention category nodes is constructed, initialization is carried out by utilizing node representations generated by an intention representation and classification model, and information updating is carried out by adopting graph attention. The decoder uses the intent category nodes and word nodes of the graph update as decoding control guidance to generate an abstract of intent control, thereby generating an abstract only containing key help information. According to the method, rescue workers and decision makers can respond more quickly and accurately, and the response efficiency of the society to crisis events is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of text processing, and in particular, relates to a method for generating a summary of key crisis help information based on intention detection. Background Art

[0002] (1) Crisis information processing technology

[0003] During crisis and disaster events, social media has become an important channel for information dissemination, carrying all kinds of information that is critical to rescue work. Effectively parsing and managing this information is crucial for the successful implementation of rescue operations. To meet this challenge, a series of studies have been devoted to identifying, classifying and processing crisis-related information from social media content.

[0004] Crisis information processing refers to the process of using technology to refine crisis information according to the needs of the scenario when a crisis event occurs. In the study of crisis information processing, social media plays an important role in information dissemination. Crisis management departments such as hospitals, fire departments, police stations, etc., specifically analyze social media to obtain the overall situation during the crisis, strengthen communication after the crisis, and provide information to relevant decision makers. For the public, social media has gradually become a channel for them to release information, communicate, and seek help during various types of crisis events. However, the massive amount of crisis information in social media has brought great challenges to emergency response. Due to the diverse characteristics of crisis information content, the complex information source channels, and the uneven information quality, if the crisis information is not processed, emergency personnel may not be able to respond to the information in a timely manner, and thus cannot effectively respond to the crisis.

[0005] Existing research has explored and studied information processing in crisis scenarios. Some studies have used big data to study the authenticity of Twitter tweets during the Hurricane Harvey crisis by studying the characteristics of user information, release time, and content. Other studies have used integer linear programming to summarize Weibo tweets within a period of time after the crisis. In addition, some studies have used the LDA topic analysis model to gain a deep understanding of people's emotional state and response measures for relevant departments to make decisions.

[0006] Although many existing studies have explored the authenticity, summary, and analysis of crisis information to a certain extent, the generation of summary for intention control in crisis scenarios still needs further exploration. By accurately and deeply mining and understanding the user's intention to seek help in a crisis environment, we can extract key information driven by intention and generate summaries to improve the efficiency of crisis management, reduce the information overload pressure of rescuers and decision makers, and improve the rapid response capability of rescue in crisis scenarios.

[0007] (2) Abstract Generation Technology

[0008] Summarization is the task of extracting the main information from the original text to form concise and refined summary content. For generative text summarization, early text summary generation mainly used neural networks for text summary generation. The encoder-decoder model architecture is the main one. These models use structures such as LSTM or GRU for sequence modeling. In order to better handle long texts and better focus on the key information in the input sequence, the attention mechanism was introduced into text summary generation. This mechanism enables the model to dynamically focus on different parts of the input sequence, thereby improving the generation quality.

[0009] In recent years, research has focused on pre-training language models. Some studies have proposed the BART model. Since there are differences between pre-training tasks and downstream text generation tasks when using generative models such as BERT, BART uses a new task of adding noise to destroy and reconstruct the original document to adapt to text generation. Other studies have proposed the T5 (Text-to-Text Transfer Transformer) model, which integrates various tasks into the same architecture through different prefix inputs and uses the method of masking phrases as pre-training targets. In addition, some studies have proposed the PEGASUS (Pre-training with Extracted Gap-sentences for Abstractive Summarization) model, which uses the method of masking sentences as pre-training targets and is specifically used for summary generation tasks. At the same time, some studies have explored controllable summary generation by introducing additional information. For example, some studies have proposed integrating topic encoding into the decoding process and generating corresponding summaries based on the topic. Other studies have proposed summaries based on focus points, using momentum encoders for contrastive learning, and constructing simple negative samples and hard negative samples to generate summaries for focus points. By integrating additional information, the relevance and quality of summaries can be improved. Summary of the invention

[0010] In view of the above-mentioned deficiencies of the prior art, the purpose of the present invention is to provide a method for generating a summary of key crisis help-seeking information based on intention detection. The method extracts the help-seeking intention category from the help-seeking information in a crisis scenario, and generates a corresponding help-seeking summary based on the category. The present invention uses a two-stage summary generation method to enable rescuers and decision makers to respond more quickly and accurately, thereby improving the efficiency of society's response to crisis events.

[0011] The technical solution of the present invention is specifically described as follows.

[0012] A method for generating summary of key information of crisis help requests based on intention detection, comprising an intention representation and classification stage based on contrastive learning and a summary generation stage of help request intention control;

[0013] 1) Intent representation and classification stage

[0014] The help text is encoded through the BERT model to obtain the help intention instance representation and word representation of the help text. Contrastive learning is used to compare different help instance intentions in the help information text, and between the intent category representation obtained through clustering and the help intention instance representation, to generate more accurate help intention instance representation and intent category representation;

[0015] 2) Summary Generation Stage

[0016] Based on word representation, help intention instance representation and intent category representation, a heterogeneous graph consisting of word nodes, help information nodes and intent category nodes is constructed. The graph is encoded using a graph attention network, and a decoder consisting of two stages: intent level decoding and word level decoding is used to fuse and summarize the intent. Finally, a summary of key help information that is highly relevant to the help intention is generated.

[0017] In the present invention, in the intent characterization and classification stage, the specific steps include:

[0018] (1) Convert the help-seeking information text input into a help-seeking intention instance representation, and classify the help-seeking intention instance representation;

[0019] (2) Conducting comparative learning of the intent of help-seeking information text instances, comparing different help-seeking intent representations within the same batch to improve the representation ability of intent instances, and obtaining the help-seeking instance intent representation;

[0020] (3) By using intention comparison, all generated help-seeking intention instance representations are clustered, and the help-seeking intention category representation is obtained by obtaining the cluster center. The help-seeking intention instances are compared with the help-seeking intention categories to obtain a better intention category representation.

[0021] In the present invention, in step (1), a BERT classifier is used to convert the help-seeking information text input into a help-seeking intention instance representation, and the intention classification is performed based on the help-seeking intention instance representation.

[0022] In the present invention, in step (2), when performing intent comparative learning of help information text instances, the intent instance representations obtained by inputting the same help information into the BERT classifier twice are taken as positive example pairs, and the intent instance representations of other help information with different intent classification labels in the same batch are taken as negative example pairs of the help information. By measuring the gap between the intent instance representations in the vector space, the distance between the distribution of the same type of intent instance representations in the space is shortened, and the distance between the distribution of the different types of intent instance representations in the space is increased, so as to form a more accurate help intention instance representation, thereby obtaining a better help intention instance representation.

[0023] In the present invention, in step (3), the intention comparison includes the following steps:

[0024] First, all generated help-seeking intent instance representations are clustered using K-means clustering, and the cluster centers after clustering represent the help-seeking intent category representations;

[0025] Then, the intent representation of the help-seeking information and the intent category representation corresponding to the help-seeking information category are taken as a positive pair, and the intent representation of the help-seeking information and the intent category representation of a different category are taken as a negative pair. By measuring the gap between the intent representation and the intent category representation in the vector space, the distribution of similar intent representations in space is made more compact, the distance between intent representations of different categories is made farther, and the cluster center, i.e., the intent category representation, is more accurate, thereby obtaining a better intent category representation.

[0026] In the present invention, in the intent representation and classification stage, the BERT classifier, intent instance comparison and intent comparison are trained simultaneously through a multi-task training strategy, and the final loss function is:

[0027]

[0028] In the formula, is the loss function of the BERT classifier, Use cross entropy loss function; is the loss function for intent instance comparison, which uses InfoNCE loss; is the loss function for intent comparison, which adopts InfoNCE loss. α, β, and γ are the weight parameters of BERT classifier, intent instance comparison, and intent comparison, respectively.

[0029] In the present invention, in the summary generation stage, an encoder-decoder architecture is used to generate the help information summary; the encoder is a document graph encoder, and the decoder is an intent fusion decoder; wherein:

[0030] For the encoder, we first construct a heterogeneous graph consisting of word nodes, help information nodes, and intent nodes, and then use the graph attention network to update the information of the node representation in the graph to improve the representation ability of the help information text;

[0031] For the decoder, the intent category representation generated by the encoder update is first decoded at the intent level to provide intent control over the summary, and then the word representation generated by the encoder is decoded to generate a fluent intent-controlled summary of key help information, that is, a summary that only contains key help information.

[0032] In the present invention, in the encoder, three steps are included:

[0033] First, a heterogeneous graph is constructed: the point set in the graph consists of three types of nodes, namely help information nodes, intent category nodes, and word nodes; the edge set consists of two types, the edges between word nodes and intent category nodes, and the edges between help information nodes and intent category nodes of their corresponding categories;

[0034] Secondly, initialize the representation of various node representations in the graph: for help information nodes, use the special mark in the BERT classifier <cls>The output of the corresponding position is used as the help information representation for initialization; for word nodes, the output corresponding to each word in the BERT classifier is used as the initialization representation of the word node; for intent category nodes, the intent category representation of the cluster center after clustering is used for initialization;

[0035] Then, the node representation is transformed to the same space, and a multi-layer graph attention network with a multi-head attention mechanism is used to update the graph. For each node, the node is updated by calculating the attention score between it and all neighboring nodes as the weight of the edge. The multi-head attention mechanism is used to enrich the association relationship between nodes. In order to prevent the gradient disappearance caused by the superposition of multiple layers of graph attention layers, layer normalization and residual network are used to connect different graph attention layers.

[0036] In the present invention, the decoder generates the summary of the intention control in an autoregressive manner; in the decoder, there are two steps, namely, intention level decoding and word level decoding;

[0037] 1) Intent-level decoding

[0038] The word sequence generated at all times so far is used as the generated partial summary. First, the generated summary is passed through the masked multi-head attention layer to build its context information. Then it is input into the intent fusion layer. The intent category node representation generated by the graph encoder is used as the input of this layer. Through the multi-head attention mechanism, the intent-aware fusion information is generated.

[0039] 2) Word-level decoding

[0040] The intent-aware fusion information generated by intent-level decoding is used as the query embedding input of the multi-head attention layer, and the word representation updated by the graph encoder is used as the key and value embedding input. After passing through the multi-head attention layer, it is input into the feedforward network layer to generate the candidate word probability at the next moment, and finally the word at that moment is generated through Softmax.

[0041] In the present invention, the cross entropy loss function is used to train the encoder-decoder architecture in the summary generation stage.

[0042] In the present invention, the cross entropy loss function is used to train the encoder-decoder architecture in the summary generation stage.

[0043] Compared with the prior art, the beneficial effect of the present invention is that: by using contrastive learning, the present invention compares the intent instance representation and the intent category representation, thereby improving the model's intent instance representation capability, thereby improving the classification accuracy of the intent to help, so that different categories of intent to help can be more accurately identified and processed. In addition, by using a graph attention network and a two-stage decoder, the quality of summary generation is improved by using the high-quality intent to help instance representation, intent to help category representation, and word representation generated in the first stage, and a summary that better reflects the user's intention can be generated, which performs particularly well in crisis scenarios, so that rescuers and decision makers can obtain key information more quickly and accurately, improving the efficiency and accuracy of crisis response. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a diagrammatic representation of the structure of the classification model intended for the present invention.

[0045] Figure 2 Generate a model structure diagram for the invention summary.

[0046] Figure 3 It is a diagrammatic representation of the structure of the document image encoder of the present invention.

[0047] Figure 4 It is a diagrammatic representation of the decoder structure of the present invention.

[0048] Figure 5 This is the purpose of the present invention to characterize and compare the classification model results.

[0049] Figure 6 Generate model comparison results for this invention summary.

[0050] Figure 7 Evaluation results for users of the system of the present invention. DETAILED DESCRIPTION

[0051] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0052] The present invention provides a method for generating summary of key information of crisis help-seeking based on intention detection. For help-seeking information in crisis scenarios, the help-seeking intention category is extracted, and a corresponding help-seeking summary is generated according to the category. The method includes two stages: an intention representation and classification stage based on contrastive learning and a summary generation stage of help-seeking intention control. In the intention representation and classification stage based on contrastive learning, the similarities and differences between help-seeking intention texts are captured by contrastive learning, and the help-seeking intention category representation is obtained by clustering, thereby improving the intention classification capability. In the summary generation stage of help-seeking intention control, a heterogeneous graph consisting of words, documents, and intention nodes is constructed based on an encoder-decoder as a basic architecture, and each node representation generated by the intention representation and classification model is initialized, and the graph attention is used to update the information. The decoder constructs a two-stage decoder, and uses the updated intention category node and word node of the graph as a decoding control guide to generate an intention-controlled summary. Thus, a summary containing only key help-seeking information is generated. The present invention enables rescuers and decision makers to respond more quickly and accurately through a two-stage summary generation method, thereby improving the efficiency of social response to crisis events. The details are as follows.

[0053] A method for generating summary of key information of crisis help based on intention detection, the specific steps are as follows:

[0054] (1) An intent representation and classification model based on contrastive learning was designed. Contrastive learning was used to compare different help-seeking intent instances, the intent category representations obtained through clustering and the help-seeking intent instances, and to generate intent categories and their intent representations.

[0055] (2) A key information summary generation model for help-seeking intention control is designed. A heterogeneous graph of help-seeking information-intent-word is constructed, and the graph is encoded using a graph attention network. A decoder consisting of two stages, intent-level decoding and word-level decoding, is used to fuse the intent and generate a summary.

[0056] Finally, a summary of help-seeking information that is highly relevant to the help-seeking intention can be generated.

[0057] 1. Intent representation and classification model based on contrastive learning

[0058] The overall structure of the model is as follows Figure 1 As shown in the figure. The model as a whole can be divided into three parts: BERT classifier, intent instance comparison and intent comparison. The BERT classifier converts the help information input into intent instance representation, and classifies the intent according to the help intent instance representation. Then, through intent instance comparison, the different help intent instance representations in the same batch are compared to improve the representation ability of intent instances. At the same time, intent comparison is used to cluster all the generated intent instance representations, obtain the representation of the help intent category by obtaining the cluster center, and compare the help intent instance representation with the intent category representation to obtain a better intent category representation.

[0059] a) BERT Classifier

[0060] In order to model the input help information, the BERT model is used to encode the help information. The core structure of the BERT model consists of multiple Transformer encoder layers, each of which contains a multi-head self-attention mechanism. The model learns the relationship in the context during the pre-training process to achieve efficient modeling of the context. <cls>The corresponding output is used as the intent instance representation of the input help information. In order to apply the BERT model to the help intention classification task, a classification layer is added on top of the BERT model. Based on the pre-training, the model is fine-tuned to adapt it to the intention detection task in crisis scenarios. For the BERT classifier, the cross entropy loss function is used for training. The cross entropy loss function formula is as follows:

[0061]

[0062] Where K is the number of categories, p i is the probability of the ith category.

[0063] b) Intent instance comparison

[0064] In this paper, the intention contrast learning between help information text instances is introduced. The intention instance representations obtained by inputting the same help information into BERT twice are taken as positive pairs. Since the dropout is different twice, the output representations are also different. However, since the content of the help information and the intention classification results are consistent, the two different representations should be similar. At the same time, other help information in the same small batch is taken as the negative pair of the help information. Due to the different contents, the help intentions corresponding to the text may also be different, and the corresponding intention instance representations should also have differences. Since the same small batch may contain samples with the same intention classification label as the help information, and the intention representations of samples with the same label should be similar, in order to reduce the impact of such samples with the same label, by masking the samples corresponding to the same label, only samples with different intention labels are compared as negative pairs, reducing the impact of false negative samples on contrastive learning. By measuring the gap between the intention instance representations in the vector space, the distance between the distribution of the same type of intention instance representations in the space is reduced, and the distance between the distribution of the different types of intention instance representations in the space is increased, forming a more accurate help intention instance representation, thereby improving the classification effect of the model. For intent instance comparison, contrastive learning InfoNCE loss is used for training. The loss function of intent instance comparison is as follows:

[0065]

[0066] Among them, sim(·) is the cosine similarity, h is the intention expression of a help-seeking message, and h + is the intention representation of the same help information after different dropouts, h neg It represents other help-seeking intents that are different from the help-seeking intent category label.

[0067] c) Intentional comparison

[0068] This part is divided into two steps.

[0069] First, all the help-seeking information sets are encoded through BERT to obtain the intent instance representations of all the help-seeking information and cluster them using K-means clustering. The cluster centers after clustering represent the help-seeking intent category representations.

[0070] Then, establish intent comparison. Since the intent instance representation of the help information under each intent category should be similar to its corresponding intent category representation, the intent instance representation of the help information and the intent category representation corresponding to the help information category are taken as positive example pairs. For a certain help information, its intent instance representation should be quite different from the intent category representation of a different intent category. The intent instance representation of the help information and the intent category representation of a different category are taken as negative example pairs. Similarly, by measuring the gap between the intent instance representation and the intent category representation in the vector space, the representations of the same type of intent instances are distributed more compactly in space, the representations of intent instances of different categories are farther apart, and the cluster center, i.e., the intent category representation, is more accurate, thereby obtaining a better intent category representation. For intent comparison, the contrastive learning InfoNCE loss is also used for training. The loss function of intent comparison is as follows:

[0071]

[0072] Among them, h is the intention representation of a help-seeking message, c is the intention category representation corresponding to the help-seeking message, K is the number of intention categories, and F is the intention category corresponding to the help-seeking message.

[0073] Finally, the above three parts of BERT classifier, intent instance comparison and intent comparison are trained simultaneously through a multi-task training strategy, and the final loss function is:

[0074]

[0075] 2. Key Information Summary Generation Model for Help-Seeking Intention Control

[0076] The model uses an encoder-decoder architecture to generate a summary of the help information. For the encoder, the aforementioned contrastive learning-based intent representation and classification model is used to encode the help information, obtain the corresponding word representation, intent instance representation, and intent category representation, and construct a graph composed of these three types of representations. The node representation in the graph is updated through the graph attention network to improve the representation ability of the short text of the help information. For the decoder, the intent category representation generated by the encoder is first decoded at the intent level to provide intent control for the summary. Then the words are decoded to generate a fluent intent-controlled help information summary. The overall structure of the model is as follows: Figure 2 shown.

[0077] a) Document Graph Encoder

[0078] The document graph encoder structure is as follows Figure 3 shown.

[0079] First, the graph required by the document graph encoder is constructed. The point set consists of three types of nodes, namely help information nodes, intent category nodes, and word nodes. The edge set consists of two types, the edges between word nodes and intent category nodes, and the edges between help information nodes and intent category nodes of their corresponding categories.

[0080] Secondly, the node representation in the graph is initialized. For each type of node in the graph, the intent representation based on contrastive learning and the classification model output are used for initialization. The specific method is as follows.

[0081] For help information nodes, special tags in the BERT classifier are used <cls>The output at the corresponding position is initialized as the help information representation.

[0082] For word nodes, the output corresponding to each word in the BERT classifier is used as the initialization representation of the word node.

[0083] For the intent category nodes, the intent category representation of the cluster center in the model is used for initialization.

[0084] Then, given the constructed graph and the node initialization representation, the graph nodes will be updated. A multi-layer graph attention network with a multi-head attention mechanism is used to update the graph. Since the graph attention network is only applicable to homogeneous graphs, and the graph used here is a heterogeneous graph, before updating the graph nodes, the node representation needs to be transformed to the same space to satisfy the graph attention network. The calculation method is as follows:

[0085]

[0086] in, are all learnable parameters. is the representation of the ith node in the lth layer. For each node, the attention score between it and all neighboring nodes is calculated as the weight of the edge to update the node. The multi-head attention mechanism is also used to enrich the association relationship between nodes. The calculation method is as follows:

[0087]

[0088] Among them, LeakyReLU(·) is the activation function, is a trainable mapping, W is a trainable parameter matrix, [·||·] is a concatenation operation, is the set of all neighboring nodes of node i. Similar to the graph attention network, in order to prevent the gradient from disappearing due to the superposition of multiple graph attention layers, layer normalization and residual network are used to connect different layers. The calculation method is as follows:

[0089]

[0090] b) Decoder for intent fusion

[0091] The decoder structure of the intention fusion is as follows Figure 4 As shown in Figure 2, the decoder uses an autoregressive approach to generate a summary of intent control. In this decoder, there are two steps, namely intent level decoding and word level decoding.

[0092] Intent-level decoding

[0093] For intent-level decoding, the word sequence generated at all times so far is used as the generated partial summary. First, the context information of the generated summary is constructed through the masked multi-head attention layer. Then it is input into the intent fusion layer, and the intent category node representation generated by the graph encoder is used as the input of this layer. Through the multi-head attention mechanism, intent-aware fusion information is generated. The calculation formula of this part is as follows:

[0094] h i =MaskMultiHeadAttn(e i , E [:i-1] , E [:i-1] )

[0095] v i =MultiHeadAttn(u i , H I , H I )

[0096] Among them, h i is the output of the multi-head attention layer with mask, v i H is the output of the help-seeking intention fusion layer. I is the help-seeking intent category representation after the graph encoder updates, E [:i-1] Generate summary word embeddings for the first i-1 words.

[0097] Word-level decoding

[0098] For word-level decoding, the representation of the fused intent generated by the intent-level decoding is used as the query embedding input of the multi-head attention layer, and the word representation updated by the graph encoder is used as the key and value embedding input. After passing through the multi-head attention layer, it is input into the feedforward network layer to generate the probability of candidate words at the next moment, and finally the word at that moment is generated through Softmax. The calculation formula for this part is as follows:

[0099] o i =MultiHeadAttn(v i , H W , H W )

[0100] x i =W2tanh(W1o i +b1)+b2

[0101] p i =Softmax(x i )

[0102] Among them, i is the output of the multi-head attention layer, x i is the output of the feedforward network layer, W and b are the learnable parameter matrices, and H W The word representation updated for the graph encoder.

[0103] For the entire model, the cross-entropy loss function is used to train the model. The formula of the cross-entropy loss function is as follows:

[0104]

[0105] The following are specific embodiments.

[0106] Embodiment 1

[0107] The data set is constructed by using the crisis help-seeking information data published on Weibo during the ×× period in the first half of 2022. The original data is collected by retrieving and crawling Weibo posts with tags such as "#××× help-seeking on ××", "#×× help-seeking", "#Weibo ×× help-seeking platform" and so on between March 2022 and July 2022. And it is judged whether each Weibo contains help-seeking information by manual review one by one, and non-help-seeking information such as providing help, information sharing, expressing emotions, etc. in the collected original data is excluded. Next, the content of the help-seeking information in the data set is processed. The forwarded and duplicate Weibo content is deleted, and at the same time, the Weibo content that only contains emoticons, topic tags, labels or URLs is deleted. The topic tags, username references, short links of the Weibo ×× help-seeking platform, etc. that are contained in a large number of help-seeking information in the Weibo are removed, and the blank characters contained in the help-seeking information are removed. After processing the original data, a total of 5018 effective help-seeking information are collected.

[0108] The data set is analyzed and annotated. The help-seeking information in the data set is divided into help-seeking intentions such as medical treatment, medicine, food and daily necessities, pet supplies, mother and baby supplies, travel by car, sending items, and others. And the data set is manually annotated to generate reference help-seeking summary information. The data set is divided into a training set, a validation set and a test set according to the ratio of 6:1:3. The training set contains 3013 help-seeking information in total, the validation set contains 506 help-seeking information in total, and the test set contains 1499 help-seeking information in total.

[0109] For the intention representation and classification model based on contrast learning, the constructed data set is used to compare with the existing classification models. The detailed results are as Figure 5 shown. Compared with other models, this model benefits from the association between different help-seeking information texts provided by contrast learning for the model. The spatial distribution learned by the model is more compact, the intention instance representation and the intention category representation are more reasonable, and there are different degrees of improvement in indicators such as accuracy, precision and recall.

[0110] For the key information summary generation model for help-seeking intention control, the constructed data set is used to compare with the existing summary models. The detailed results are as Figure 6 Compared with other models, the graph encoder used in this model can generate better representations for short texts. The global relationship brought by the graph encoder enables the model to better represent intent and words, and then generate summaries that are more relevant to intent in the decoder. At the same time, compared with other models pre-trained for summary tasks, the summaries generated by this model can better reflect the user's intentions and are more suitable for summarizing users' help texts in crisis scenarios.

[0111] At the same time, in order to verify the capabilities in actual scenarios, a crisis help platform was built, which includes a shared document that can be edited online by multiple people at the same time. In the platform, a crisis help key information summary generation system based on intent detection was implemented. The system aims to classify the help information in the online community according to its help intention, and generate a concise summary of the help content based on its help intention. The system is built with Go and Python and deployed on a Linux server. After the user enters in the shared document, it is generated. First, the data is preprocessed to remove the tags, website hyperlinks, emoticons, etc. contained in it. Then, the cleaned help information is input into the intention representation and classification model based on contrastive learning to generate the intention category of the help information. Next, the intention representation and help information representation generated by the model are input into the key information summary generation model controlled by the help intention to generate a key information summary that is highly related to the help intention. And the classification and summary cells corresponding to the help information in the shared document are updated for display. Thus, the corresponding help intention category is automatically generated, and the corresponding core help purpose summary is generated according to the intention, so that rescue personnel can quickly understand the needs of the help seeker and improve the response speed of help.

[0112] In order to evaluate the user experience and effectiveness of the system in actual applications, a questionnaire survey on the use of the platform was conducted.

[0113] The questionnaire mainly includes the following key questions:

[0114] 1. How relevant are the generated help-seeking intent categories (i.e., the categories of help needed) to the help-seeking information?

[0115] 2. Is the generated help summary fluent?

[0116] 3. How relevant is the generated help-seeking content summary to the original help-seeking information content?

[0117] 4. How well does the generated help-seeking content summary match the help-seeking intent category (i.e., the category of help needed)?

[0118] Each question has five different levels of answers, from low to high: very poor, poor, average, good, and very good. Users need to evaluate each help-seeking information category and key information summary when they generate a help-seeking information. Ten volunteers were recruited to evaluate the system. Users need to evaluate the content generated by the system every time they generate a help-seeking information. Each volunteer is required to publish and generate at least five help-seeking information categories and summaries on the platform.

[0119] The 50 questionnaires generated by these 10 volunteers were analyzed, and the distribution of the results of each question was as follows: Figure 7 The results show that most users think the system performs well in identifying the relevance of intent categories and help-seeking information, indicating that the system can accurately identify user intent and associate it with the corresponding help-seeking information. Most users think that the system-generated summary performs well in terms of relevance to help-seeking information. Most users said that the system performs well in terms of practicality and can help quickly understand and respond to help-seeking information in crisis scenarios.< / cls> < / cls> < / cls>

Claims

1. A method for generating summary of key information of crisis assistance based on intention detection, characterized in that: It includes the intention representation and classification stage based on contrastive learning and the summary generation stage of help-seeking intention control; 1) Intent representation and classification stage The help text is encoded through the BERT model to obtain the help intention instance representation and word representation of the help text; contrastive learning is used to compare the different help intention instance representations in the help information text, and the intent category representation obtained through clustering with the help intention instance representation, to generate more accurate help intention instance representation and intent category representation; 2) Summary Generation Stage Based on word representation, help intention instance representation and intent category representation, a heterogeneous graph consisting of word nodes, help information nodes and intent category nodes is constructed. The graph is encoded using a graph attention network, and a decoder consisting of two stages: intent level decoding and word level decoding is used to fuse and summarize the intent. Finally, a summary of key help information that is highly relevant to the help intention is generated.

2. The method for generating summary of key crisis assistance information according to claim 1, characterized in that: In the intent representation and classification phase, the specific steps include: (1) Input the help information text into the BERT model and convert it into help intent instance representation and word representation, and then classify the help intent based on the help intent instance representation; (2) Conducting comparative learning of the intent of help-seeking information text instances, comparing the representations of different help-seeking intent instances in the same batch to improve the representation ability of intent instances and obtain more accurate representations of the intent of help-seeking instances; (3) By using intention comparison, all generated help-seeking instance intention representations are clustered, and the help-seeking intent category representation is obtained by obtaining the cluster center. The help-seeking intent instance representation and the help-seeking intent category representation are compared and learned to obtain a better intent category representation.

3. The method for generating summary of key crisis assistance information according to claim 2, characterized in that: In step (1), the BERT classifier is used to convert the help information text input into a help intent instance representation, and the intent classification is performed based on the help intent instance representation.

4. The method for generating summary of key crisis assistance information according to claim 2, characterized in that: In step (2), when performing comparative learning of the intent of the help information text instance, the intent representation obtained by inputting the same help information into the BERT classifier twice is used as a positive pair, and other help information with different intent classification labels in the same batch is used as a negative pair of the help information. By measuring the gap between the intent representations in the vector space, the distance between the distribution of the same type of intent representations in the space is shortened, and the distance between the distribution of the representations of different categories of intent in the space is increased, so as to form a more accurate help intention instance representation, thereby obtaining a better help intention instance representation.

5. The method for generating summary of key crisis assistance information according to claim 2, characterized in that: In step (3), the intention comparison includes the following steps: First, all generated help-seeking intent instance representations are clustered using K-means clustering, and the cluster centers after clustering represent the help-seeking intent category representations; Then, the intent instance representation of the help-seeking information and the intent category representation corresponding to the help-seeking information category are taken as the positive pair, and the intent instance representation of the help-seeking information and the intent category representation of different categories are taken as the negative pair. By measuring the gap between the intent instance representation and the intent category representation in the vector space, the distribution of the same type of intent instance representations in the space is made more compact, the distance between the intent instance representations of different categories is farther, and the cluster center, i.e., the intent category representation, is more accurate, thereby obtaining a better intent category representation.

6. The method for generating summary of key crisis assistance information according to claim 2, characterized in that: In the intent representation and classification stage, the BERT classifier, intent instance comparison, and intent comparison are trained simultaneously through a multi-task training strategy. The final loss function is: In the formula, is the loss function of the BERT classifier, Use cross entropy loss function; is the loss function for intent instance comparison, which uses InfoNCE loss; is the loss function for intent comparison, which adopts InfoNCE loss. α, β, and γ are the weight parameters of BERT classifier, intent instance comparison, and intent comparison, respectively.

7. The method for generating summary of key crisis assistance information according to claim 1, characterized in that: In the summary generation stage, an encoder-decoder architecture is used to generate the help information summary; the encoder is a document graph encoder, and the decoder is an intent fusion decoder; where: For the encoder, we first construct a heterogeneous graph consisting of word nodes, help information nodes, and intent nodes, and then use the graph attention network to update the information of the node representation in the graph to improve the representation ability of the help information text; For the decoder, the intent category representation generated by the encoder update is first decoded at the intent level to provide intent control over the summary, and then the word representation generated by the encoder is decoded to generate a fluent intent-controlled summary of key help information, that is, a summary that only contains key help information.

8. The method for generating summary of key crisis assistance information according to claim 7, characterized in that: In the encoder, there are three steps: First, a heterogeneous graph is constructed: the point set in the graph consists of three types of nodes, namely help information nodes, intent category nodes, and word nodes; the edge set consists of two types, the edges between word nodes and intent category nodes, and the edges between help information nodes and intent category nodes of their corresponding categories; Secondly, initialize the representation of various node representations in the graph: for help information nodes, use the special mark in the BERT classifier <cls> The output of the corresponding position initializes the representation of the help information;< / cls> For word nodes, the output corresponding to each word in the BERT classifier is used as the initialization representation of the word node; for intent category nodes, the intent category representation of the cluster center after clustering is used for initialization; Then, the node representation is transformed to the same space, and a multi-layer graph attention network with a multi-head attention mechanism is used to update the graph. For each node, the attention score between it and all neighboring nodes is calculated as the weight of the edge to update the node. A multi-head attention mechanism is used to enrich the correlation between nodes. In order to prevent the gradient disappearance caused by the superposition of multiple graph attention layers, layer normalization and residual network are used to connect different graph attention layers.

9. The method for generating summary of key crisis assistance information according to claim 7, characterized in that: The decoder uses an autoregressive approach to generate a summary of intent control. In this decoder, there are two steps: intent-level decoding and word-level decoding. 1) Intent-level decoding The word sequence generated at all times so far is taken as the generated partial summary. First, the generated summary is passed through the masked multi-head attention layer to build its context information. Then it is input into the intent fusion layer. The intent category node representation generated by the graph encoder is used as the input of this layer. Through the multi-head attention mechanism, the intent-aware fusion information is generated. 2) Word-level decoding The intent-aware fusion information generated by intent-level decoding is used as the query embedding input of the multi-head attention layer, and the word representation updated by the graph encoder is used as the key and value embedding input. After passing through the multi-head attention layer, it is input into the feedforward network layer to generate the candidate word probability at the next moment, and finally the word at that moment is generated through Softmax.

10. The method for generating summary of key crisis assistance information according to claim 7, characterized in that: The encoder-decoder architecture is trained using a cross-entropy loss function during the summary generation phase.

Citation Information

Cited By

  • New view synthesis method, system and device

    CN121190678A

  • A new view synthesis method, system and device

    CN121190678B