News event discovery method based on event review

Through event review and noise sentence occlusion technology, the news clustering results are updated, which solves the problem of insufficient clustering accuracy of news flow data and improves the event discovery performance of the news retrieval system.

CN118916486AInactive Publication Date: 2024-11-08HANGZHOU DIANZI UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410879559.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-02
Publication Date
2024-11-08
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The prior art is difficult to apply for long-term use when processing news flow data, and the accuracy of clustering results is insufficient, so it is impossible to effectively explore the problem of changes in clustering results over time, resulting in low event discovery performance in news retrieval systems.

Method used

Using the news event discovery method based on event review, better embedded representation is obtained through the article coding model, and combined with the noise sentence occlusion technology, falsely analysed and allocated news articles, updated clustering results, and improved clustering accuracy .

Benefits of technology

It improves the performance of event discovery in the news retrieval system, improves the similarity and clustering accuracy of news articles for the same event, reduces interference from noisy sentences, and realizes self-correction and data flow updates.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118916486A_ABST
    Figure CN118916486A_ABST
Patent Text Reader

Abstract

The invention discloses a news event discovery method based on event review, and the method comprises the steps: obtaining an embedded representation which can better represent the event content of a news article through a proposed article coding model, and carrying out the re-analysis of the news article which may have misjudgment through the combination of an event review and unrelated sentence shielding method, and according to an analysis result, redistributing to a corresponding cluster or generating a new event cluster, so that part of misjudged news is clustered to a correct cluster, and an event clustering space is updated. According to the method, the problem that the clustered events cannot be updated along with time in the process of discovering and clustering the news events by the news retrieval system is solved, historical news data are clustered into more accurate event clusters by continuously reviewing and correcting errors, and the performance of discovering the news events in the news retrieval system is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of text processing, and relates to a news event discovery method based on event review, and more specifically to a method for solving the problem that news retrieval systems have numerous news relations, complex sentence contexts, and are difficult to discover news belonging to the same type of events. Background Art

[0002] Nowadays, news articles covering real-time events are published through a large number of online platforms. The huge flow of news articles makes it difficult for people to quickly find news articles belonging to the same event, resulting in a long time spent on sorting out news articles. Real-time discovery of news articles with unique topics can not only bring great benefits to individuals tracking news reports, but also help organizations make strategic decisions. Therefore, the task of event discovery in online news reports has become increasingly important.

[0003] The current online news event discovery methods mainly obtain the embedded representation of articles and gradually cluster them. Most methods use static and explicit features for embedding representation and clustering, which is undoubtedly difficult to apply in the long term for the ever-changing news stream information, and the cost of obtaining the displayed features is also high. Susik et al. proposed the SCStory framework in an article published at the 2023 WWW conference. SCStory can already discover and cluster news events from the article stream in an unsupervised manner and achieve the best performance. However, how to further improve the accuracy of clustering results on streaming data is a key issue that needs to be continuously addressed. Summary of the invention

[0004] One purpose of the present invention is to address the problem that the prior art only performs conventional clustering on news stream data and is unable to discover that the clustering results of clustered news articles change over time. A news event discovery method based on event review is provided, which uses a proposed article encoding model to obtain an embedded representation that can better represent the event content of news articles, and then combines event review and noise sentence masking methods to re-analyze news articles that may have misjudgments, and reallocate them to corresponding clusters or generate new event clusters based on the analysis results, thereby clustering some of the misjudged news into correct clusters, ultimately improving the performance of event discovery in news retrieval systems.

[0005] The technical solution adopted by the present invention is as follows:

[0006] An object of the present invention is to provide a news event discovery method based on event review, the method comprising the following steps:

[0007] Step (1), obtain the initial sentence embedding representation of the news text:

[0008] First, the news article text is segmented into a group of sentences, then the sentences are encoded using a pre-trained language model, and finally truncated or padded to make the number of sentences in the article a fixed value, thus obtaining the initial sentence embedding representation of the news text;

[0009] Step (2), using the article encoding model to learn the initial sentence embedding representation, obtain sentence embedding with event indicativeness, and construct an event indicative embedding representation model for the article;

[0010] Step (3), clustering the event-indicative embedding representation of the article, assigning it to a certain event cluster or initializing it as a new event cluster, and constructing a new event cluster space;

[0011] Step (4) updates the event clustering space, recalculates the embedding representation of each event cluster center and the confidence score of the article-event pair, and then obtains the final event discovery result; specifically:

[0012] 4-1 Use the event review and re-clustering mechanism to review the event clustering space, correct the events that are clustered incorrectly, and update the event clustering space;

[0013] 4-2 Update the event indicative embedding model in step (2), mask the sentences with lower sentence weights in the article, and further update the event clustering space updated in step 4-1;

[0014] 4-3 Recalculate the embedding representation of each event cluster center and the confidence score of the article-event pair according to the event clustering space updated in step 4-3, and then obtain the final event discovery result.

[0015] Preferably, step (2) is specifically:

[0016] First, the LSTM neural network is used to learn the initial sentence embedding representation to obtain the sentence representation LS(E i ), and then use a multi-head self-attention layer to embed the initial sentence representation E i and LS(E i ) are learned separately, and two different sentence embeddings with event indicative MHS (E i )、MHS(LS(E i ));

[0017] Then, for MHS(E i )、MHS(LS(E i )) After layer normalization, two event-indicative sentence embedding representations MC are obtained through residual connections and linear layers. iand MLC i ;

[0018] Embedding two sentences to represent MC i and MLC i Sum and average to get the final sentence embedding representation C with event indication i ;

[0019] In the attention block, learn the attention weights So as to carry out C i The weighted pooling of is used to obtain the event-indicative embedding representation of the article.

[0020] Preferably, step 4-1 is specifically:

[0021] Given a threshold η, for the confidence score Articles with a probability lower than the threshold η are extracted from the existing target event clusters, and then the center embedding representations of the retained event clusters and the confidence scores of the article-event pairs are updated. Finally, the extracted articles are redistributed according to step (3);

[0022] Preferably, step 4-2 is specifically:

[0023] 4-2-1 Update event indicative embedding model; specifically:

[0024] Take any two identical events S a Two articles by i and d u ; According to the sentence weight mw i About Article i Sort in descending order and take the first 50% of the sentences; sort the articles according to the sentence weights. u Sort in descending order and take the last 50% of the sentences; then concatenate the above extracted sentences to construct a new article sample d p ; Sentence normalization weight mw i Obtained through a multi-head self-attention layer of the event-indicative embedding model;

[0025] Use the constructed new article sample d p and existing articles in the event clustering space, updating the article encoding model in step (2);

[0026] 4-2-2 Mask the sentences with lower sentence weights in the article, and then update the central embedding representation of each event cluster;

[0027] The specific operation of masking is: combining the sentence weight mw i About News iThe lower ranked sentences in the matrix are masked according to the ratio θ:

[0028]

[0029] in Indicates news article d i After normalization weight mw i Ranked after |d i | sentence, Expressing couplets

[0030] son Perform masking operation; θ∈(0,1).

[0031] Another object of the present invention is to provide a news event discovery device based on event review, comprising a memory, a processor, and a news event discovery program based on event review stored on the memory and executable on the processor, wherein the news event discovery program based on event review implements the steps of the above method when executed by the processor.

[0032] Another object of the present invention is to provide a storage medium storing a news event discovery program based on event review, wherein the news event discovery program based on event review implements the steps of the above method when executed by a processor.

[0033] The technical solution provided by the present invention has the following beneficial effects:

[0034] In view of the problem that the prior art only discovers and clusters related news events in sequence but cannot review the discovered events and check whether there are clustering errors, the present invention proposes a news event discovery method based on event review, which can perform self-correction and update clustering following the data stream, thereby improving the performance of retrieving and discovering news articles about the same event in the news retrieval system.

[0035] The present invention proposes a new article event-indicative embedding model, combines LSTM with a multi-head self-attention mechanism to extract the deep relationship between news text sentences, and finally obtains a better article embedding representation, thereby improving the performance of representing news articles in the news retrieval system.

[0036] This discovery combines the sentence normalization weights obtained by the multi-head self-attention module to mask sentences with lower weights, making the overall representation of news articles closer to the central idea. This improves the similarity of news articles belonging to the same event in the news retrieval system when updating clustering clusters, reduces the interference of noise sentences, and improves the accuracy of news clustering in the news retrieval system. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1The overall flow chart of the method of the present invention is as follows;

[0038] Figure 2 The overall architecture diagram of steps (1) to (5) of the present invention is as follows: DETAILED DESCRIPTION

[0039] The present invention will be further analyzed below in conjunction with the accompanying drawings.

[0040] A news event discovery method based on event review proposes an article encoding model. By combining LSTM and multi-head self-attention mechanism, it realizes the fusion of multi-level semantic information and dependency information between context sentences, thereby obtaining a better embedded representation of the article with event indicativeness, improving the accuracy of the content representation of news articles, and thus improving the performance of the news retrieval system; an event review re-clustering method is proposed, which reviews low-confidence events in each window of the news stream data and re-clustering them, so that a certain number of incorrectly clustered news data can be re-clustered into the correct event cluster, further improving the performance of the news retrieval system; a sentence-level noise information masking method is also proposed, which masks the less influential noise sentences in the article, so that the final representation of the article is more in line with the central theme, thereby improving the performance of the news retrieval system. Figure 1 The following processes are included:

[0041] The news text is pre-trained with a language model to obtain the required text sentence representation vector, and then the sentence representation vector is passed through the article encoding model to obtain the article representation vector. The article representation vector is allocated or initialized, and the current news text is allocated to the event clustering space. Finally, the updated clustering space is obtained based on event review, text sentence masking, and model training. The specific architecture is as follows: Figure 2 The following steps are involved:

[0042] Step (1) obtains the initial sentence embedding representation of the news text. First, use Spacy to parse the news article text and divide the article into a group of sentences. Then use the existing pre-trained language model (Sentence-BERT or Sentence-T5) to encode the sentences in the news article to obtain the initial embedding representation of each sentence. Finally, truncation or padding is performed to make the number of sentences in the article a fixed value to obtain the initial sentence embedding representation of the news text. Specifically:

[0043] Given a news article d i , this article was written by |d i | sentences, segment the article text to get the corresponding sentence groups Given the pre-trained language model to be used and the maximum number of sentences L, we can get article d i The initial sentence embedding representation Ei for:

[0044]

[0045] Among them, ISE represents the initial sentence embedding representation of the article, and PSE is the pre-trained language model used. is the dth article i i | sentences, TP(·,L) means truncating or padding the article sentences to match the specified length L, and the final article initial sentence embedding E i It can be expressed as:

[0046]

[0047] in is the embedding representation of the Lth sentence in article i, h e is the output dimension of the pre-trained language model;

[0048] Step (2) obtains the event-indicative embedding representation of the article, uses the article encoding model to learn the initial sentence embedding representation, obtains the sentence embedding with event indicativeness, and then obtains the event-indicative embedding representation of the article from the sentence embedding with event indicativeness, specifically:

[0049] First, the LSTM neural network is used to process the initial sentence embedding representation to obtain a sentence representation LS(E i ), and then use the multi-head self-attention layer to embed the existing sentence representation E i And the sentence dependency representation LS(E i ) are studied separately to obtain two different levels of sentence dependency representation MHS (E i ) and MHS(LS(E i )), which is calculated as follows:

[0050]

[0051] Among them, H p and The calculation of is as follows:

[0052]

[0053] Where n is the number of attention heads, h e is a hidden dimension, is a learnable parameter, the superscript T indicates transposition; p∈[1,n];

[0054] Then, the two different levels of MHS output MHS (E i)、MHS(LS(E i )) After layer normalization, two event-indicative sentence embedding representations EC are obtained by combining residual connections and linear layers. i and MLC i , which is expressed as:

[0055]

[0056] Where W mc ,W mlc ,b mc ,b mlc is a learnable parameter, Indicates MC i The Lth element of Indicates MLC i The Lth element of ; LN represents layer normalization; tanh represents the activation function;

[0057] Embedding two sentences to represent MC i and MLC i Sum and average to get the final sentence embedding representation C with event indication i :

[0058]

[0059] In the attention block, learn the attention weights So as to carry out C i ’s weighted pooling to obtain the event-indicative embedding representation of the article:

[0060]

[0061] The attention weight The calculation formula is:

[0062]

[0063] Where W a ,w a ,V a is a learnable parameter;

[0064] Step (3): embedding the event-indicative representation of the article Perform clustering operations, assign it to a certain event cluster or initialize it as a new event cluster, and construct a new event cluster space; specifically:

[0065] Get the embedding representation of each existing news event cluster center

[0066]

[0067] Where S a represents the target event cluster;

[0068] Then, the confidence level of the new article assigned to each cluster is determined by comparing the embedding relevance of the new article. The embedding relevance uses the confidence score of the article-event pair. i Cluster S for a target event a If the confidence score of an article-event pair exceeds the set threshold, then the article d i Will be added to the current target event cluster S a Otherwise, the article d i Initialize the seed article for a new target event cluster;

[0069] Confidence score The calculation is as follows:

[0070]

[0071] Where cos(·,·) is the calculation of cosine similarity;

[0072] Step (4) uses the event review and re-clustering mechanism to review the event clustering space obtained after allocation, and reallocates the articles to update the event clustering space, specifically:

[0073] Given a threshold η, for articles with confidence scores lower than the threshold η, they will be extracted from the existing event clusters, the central embedding representation of each event cluster and the confidence score of the article-event pair will be updated, and then the extracted articles will be redistributed according to the method in step (3);

[0074] Step (5) trains and updates the event indicative embedding model in step (2), masks sentences with lower sentence weights in the article, and further updates the event clustering space updated in step (4), specifically:

[0075] During the training process, the sentence normalization weight mw is obtained through the multi-head self-attention layer of the model i , which is obtained by adding the weights of two different levels of multi-head self-attention layers;

[0076] Take any two identical events S j Two articles by i and d u ; Normalize the weight mw according to the sentence i About Article i Sort in descending order and take the first 50% of the sentences; sort the articles according to the sentence weights. uSort in descending order and take the last 50% of the sentences; then concatenate the above extracted sentences to construct a new article sample d p :

[0077]

[0078] in and They are article d i and d u After the sentence normalization weight mw i The mth and nth sentences after sorting; the normalized weight of the sentence mw i Obtained through the multi-head self-attention layer of the event-indicative embedding model; concat means concatenation;

[0079] Use the constructed new article samples and the existing event clustering samples in the window to train the event-indicative embedding model in step (2). Here, the contrast loss function is used as the loss function of the model for training:

[0080]

[0081] Where (d i ,S a ) is an article-event pair, τ is a set parameter, Indicates that in the time window The event clustering space in represents the data available for training, Indicates that except event S a Event Hubs embed representation outside of .

[0082] Then, combined with the attention weight of the sentence About News i The lower ranked sentences in the matrix are masked according to the ratio θ:

[0083]

[0084] in Indicates news article d i After the attention weight Ranked after |d i | sentence, Express the sentence Perform masking operation; θ∈(0,1).

[0085] Finally, recalculate the embedding representation of each event cluster center;

[0086] Step (6) obtains the final event discovery result based on the event clustering space updated in step (5) and the confidence score of the corresponding event cluster center.

[0087] The performance evaluation of this invention uses the News14 dataset, the WCEP18 dataset, and the WCEP19 dataset respectively. The experiment uses B 3 -F1, Adjusted Rand Index (ARI) and Adjusted Mutual Information (AMI) were used as evaluation metrics.

[0088] B 3 -F1 is a variant of the F1-score, a commonly used evaluation indicator for binary classification models in machine learning. Its accuracy for entity i is Precision i and recall i The calculation formula is:

[0089]

[0090] The final accuracy Precision and recall Recall are calculated as:

[0091]

[0092]

[0093] The last B 3 -F1 calculation formula is:

[0094]

[0095] AMI is an adjusted version of normalized information, which can eliminate the bias caused by random clustering results. The calculation formula is:

[0096]

[0097] Where Ω represents the clustering partition, C represents the real category partition, I(Ω; C) represents the mutual information, H(Ω) and H(C) represent the entropy of the clustering cluster and the real category respectively, and E[I(Ω; C)] represents the expected value of the mutual information.

[0098] ARI is an adjusted version of the Rand Index, which can also eliminate the bias caused by random clustering results. The calculation formula is:

[0099]

[0100] Among them, n ij represents the number of identical samples in the i-th cluster and the j-th true category, a i represents the number of samples in the i-th cluster, b j represents the number of samples in the jth true category, It represents the number of combinations of any two samples from n.

[0101] Table 1 Effects of using the Sentence-BERT pre-trained language model to discover news events on the above three datasets

[0102]

[0103] Table 2 Effects of using Sentence-T5 to train language models for news event discovery on the above three datasets

[0104]

[0105] From the table using Sentence-BERT, we can conclude that our invention is better than the current best model in B of the three datasets. 3 -F1 are improved by 2.0%, 1.4%, and 1.9% respectively; on AMI, they are improved by 2.3%, 1.1%, and 1.5% respectively; on ARI, they are improved by 3.0%, 1.9%, and 2.4% respectively.

[0106] From the table using Sentence-T5, we can conclude that our invention is better than the best model in B of the three datasets. 3 -F1, 4.0%, 3.8%, 3.5% respectively; and on AMI, 6.3%, 4.9%, 6.1% respectively.

[0107] In addition, the experiment of the present invention made the following four settings: using SCStory's article encoding model to replace the article encoding model proposed in the present invention to verify the superiority of the article encoding model proposed in the present invention; removing the event review and re-aggregation method to verify the effectiveness of the event review and re-aggregation method; removing the sentence masking method to verify the effectiveness of the sentence masking method.

[0108] Table 3 Verification experimental results of three innovative points based on the SBERT pre-trained language model of the present invention

[0109]

Claims

1. A news event discovery method based on event review, Features The method comprises the following steps: Step (1), obtain the initial sentence embedding representation of the news text: First, the news article text is segmented into a group of sentences, then the sentences are encoded using a pre-trained language model, and finally truncated or padded to make the number of sentences in the article a fixed value, thus obtaining the initial sentence embedding representation of the news text; Step (2), using the article encoding model to learn the initial sentence embedding representation, obtain sentence embedding with event indicativeness, and construct an event indicative embedding representation model for the article; Step (3), clustering the event-indicative embedding representation of the article, assigning it to a certain event cluster or initializing it as a new event cluster, and constructing a new event cluster space; Step (4) updates the event clustering space, recalculates the embedding representation of each event cluster center and the confidence score of the article-event pair, and then obtains the final event discovery result; specifically: 4-1 Use the event review and re-clustering mechanism to review the event clustering space, correct the events that are clustered incorrectly, and update the event clustering space; 4-2 Update the event indicative embedding model in step (2), mask the sentences with lower sentence weights in the article, and further update the event clustering space updated in step 4-1; 4-3 Recalculate the embedding representation of each event cluster center and the confidence score of the article-event pair according to the event clustering space updated in step 4-3, and then obtain the final event discovery result.

2. The method according to claim 1, It is characterized in that Step (2) is specifically: First, the LSTM neural network is used to learn the initial sentence embedding representation to obtain the sentence representation LS(E i ), and then use a multi-head self-attention layer to embed the initial sentence representation E i and LS(E i ) are learned separately, and two different sentence embeddings with event indicative MHS (E i )、MHS(LS(E i )); Then, for MHS(E i )、MHS(LS(E i )) After layer normalization, two event-indicative sentence embedding representations MC are obtained through residual connections and linear layers. i and MLC i ; Embedding two sentences to represent MC i and MLC i Sum and average to get the final sentence embedding representation C with event indication i ; In the attention block, learn the attention weights So as to carry out C i The weighted pooling of is used to obtain the event-indicative embedding representation of the article.

3. The method according to claim 2, It is characterized in that In step (2), two different levels of sentence embedding MHS with event indicativeness (E i )、MHS(LS(E i )) is calculated as follows: Among them, H p and The calculation of is as follows: Where n is the number of attention heads, h e is a hidden dimension, is a learnable parameter, and the superscript T indicates transposition; p∈[1,n].

4. The method according to claim 2, It is characterized in that The two event-indicative sentence embeddings in step (2) represent MC i and MLC i The calculation is as follows: Where W mc , W mlc ,b mc , b mlc is a learnable parameter, Indicates MC i The Lth element of Indicates MLC i The Lth element of ; LN represents layer normalization; tanh represents the activation function.

5. The method according to claim 1, It is characterized in that Step (3) is specifically: Get the embedding representation of the cluster centers of each existing news event Where S a represents the target event cluster; Then, the confidence level of the new article assigned to each cluster is determined by comparing the embedding relevance of the new article. The embedding relevance uses the confidence score of the article-event pair. i Cluster S for a target event a If the confidence score of an article-event pair exceeds the set threshold, then the article d i Will be added to the current target event cluster S a Otherwise, the article d i Initialize the seed article for a new target event cluster; The confidence score of an article-event pair for a target event cluster is calculated as follows: Where cos(·,·) represents the cosine similarity function.

6. The method according to claim 1, It is characterized in that Step 4-1 is as follows: Given a threshold η, for the confidence score Articles with a probability lower than the threshold η are extracted from the existing target event clusters, and then the center embedding representations of the retained event clusters and the confidence scores of the article-event pairs are updated. Finally, the extracted articles are redistributed according to step (3).

7. The method according to claim 1, It is characterized in that Step 4-2 is specifically: 4-2-1 Update event indicative embedding model; specifically: Take any two identical events S a Two articles by i and d u ; Normalize the weight mw according to the sentence i About Article i Sort in descending order and take the first 50% of the sentences; sort the articles according to the sentence weights. u Sort in descending order and take the last 50% of the sentences; then concatenate the above extracted sentences to construct a new article sample d p ; Sentence normalization weight mw i Obtained through a multi-head self-attention layer of the event-indicative embedding model; Use the constructed new article sample d p and existing articles in the event clustering space, updating the article encoding model in step (2); 4-2-2 Mask the sentences with lower sentence weights in the article, and then update the central embedding representation of each event cluster; The specific operation of the masking is: combining the sentence attention weight About News i The lower ranked sentences in the matrix are masked according to the ratio θ: in Indicates news article d i After the attention weight Ranked after |d i | sentence, Express the sentence Perform masking operations; θ∈(0,1).

8. The method according to claim 1, Features The loss function of the event-indicative embedding model is specifically: Where (d i , S a ) is an article-event pair, τ is a set parameter, Indicates that in the time window The event clustering space in represents the result space of article-event pairs, Indicates that except event S a Event Hubs embed representation outside of .

9. An event discovery device based on event review, comprising a memory, a processor, and an event discovery program based on event review stored in the memory and executable on the processor, wherein the event discovery program based on event review, when executed by the processor, implements the steps of the method described in any one of claims 1 to 8.

10. A storage medium storing an event discovery program based on event review, wherein the event discovery program based on event review, when executed by a processor, implements the steps of the method according to any one of claims 1 to 8.