A Progressive Opinion Extraction Method and System for Hot Topics

Abstract

The present invention discloses a progressive opinion extraction method and system for hot topics. The method is as follows: providing prior knowledge; constructing a seed event structure diagram based on the prior knowledge, where the diagram contains opinion information nodes and edges representing the relationships between opinion elements; combining the event structure diagram and data in the current stage to train and predict an opinion extraction model, and giving the prediction results of the data in the current stage after the training ends; removing the opinions that already exist in the event structure diagram, and submitting the new opinions to experts for confirmation; screening the confirmation results returned by the experts, removing the unqualified opinions, and adding the qualified opinions to the event structure diagram; returning to the opinion extraction step again, and repeating this cycle until the opinion extraction model converges. The present invention uses historical information for opinion extraction from new texts under the same topic, can effectively mitigate the impact of unbalanced hot topic datasets on neural models, and can obtain high-quality opinion information with a small amount of labeled data.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence, and particularly relates to a progressive opinion extraction method and system for hot topics. Background Art

[0002] With the popularization of intelligent terminals, social media has become a common platform for people to express opinions and emotions. As the main information manifestation form in the web 2.0 era, social texts include Weibo tweets, Tieba posts, Zhihu comments, etc., containing different opinions and emotions of netizens on specific hot topics. In the fields of public opinion analysis, e-commerce sales, etc., extracting users' opinions, perspectives, and emotional information on specific topics has become an important but extremely challenging task. However, existing opinion extraction algorithms mainly rely on statistical models or neural networks, and it is difficult to efficiently and accurately extract opinion information in the case of complex and changeable social media data.

[0003] Most current opinion extraction methods are based on statistical model-related methods (Reference: Zhuang L, Jing F, Zhu X Y. "Movie review mining and summarization." [C] / / Proceedings of the 15th ACM international conference on Information and knowledge management. 2006:43-50), that is, regarding the entire opinion extraction problem as a sequence labeling problem, extracting relevant features through complex manual annotation or neural network modules. These features include syntactic dependency relations, part-of-speech tagging, domain word vectors, etc., and then using the hidden Markov model or conditional random field to obtain the target output sequence. The method based on statistical models relies on complex manual features and has poor domain transfer performance, and it is difficult to be widely and practically applied in the scenario of social text opinion extraction tasks.

[0004] With the rise of neural networks, their ability to fit complex features has been favored by people, and relevant researchers have begun to use neural network models for opinion extraction. For example, related work (Reference: Ma D, Li S, Wu F, et al. "Exploring sequence-to-sequence learning in aspect term extraction." [C] / / Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019: 3538-3547.) proposed to model the opinion triple extraction task as an end-to-end sequence output, using long short-term memory neural units and attention mechanisms to capture the complex relationships between opinion elements. However, the end-to-end model structure is too complex to be iterated easily, and the distribution of relevant discussion opinions on specific topics is unbalanced - most people hold similar opinions, and complex neural models relying on labeled data perform poorly on unbalanced data sets. Peng et al. (Reference: Peng H, Xu L, Bing L, et al. "Knowing what, how and why: A near complete solution for aspect-based sentiment analysis." [C] / / Proceedings of the AAAI Conference on Artificial Intelligence: volume 34. 2020: 8600-8607.) proposed to split the opinion triple extraction task into two stages. In the first stage, words and sentiment information are extracted, and in the second stage, relevant elements are combined. Although the stage-based method can effectively simplify the model structure, it is prone to cascade errors and severs the interconnection between subtasks. Summary of the Invention

[0005] In view of the above problems, the present invention proposes a progressive opinion extraction method and system for hot topics. The method and system utilize historical information stored in an event structure graph to assist opinion extraction, and construct and update the event structure graph through a progressive framework. The historical information is incorporated into the neural network model through a dual attention mechanism to complete high-quality opinion extraction for new texts under the same hot topic.

[0006] The present invention solves the above technical problems through the following technical means:

[0007] A progressive opinion extraction method for hot topics, comprising the following steps:

[0008] S01. Provide prior knowledge: Divide the social texts under a hot topic into multiple stages, give the set of perspective categories and some viewpoints corresponding to the topic. For each perspective category, add a small number of seed words that can reflect the category information.

[0009] S02. Construct the seed event structure diagram: Based on the above prior knowledge, construct the seed event structure diagram, which includes view information nodes and edges representing the relationships between views.

[0010] S03. Train and predict the view extraction model: Combine the updated event structure diagram or seed event structure diagram in the previous stage to train and predict the current stage text stream. After training, give the prediction result of the current stage text stream.

[0011] S04. Confirm new viewpoints based on human-machine collaboration: Use the word clustering filtering algorithm based on word similarity to filter the prediction result, filter out the viewpoints that already exist in the event structure diagram, and submit the new viewpoints to experts for confirmation.

[0012] S05. Update the event structure diagram: Screen the confirmation results returned by experts, remove the viewpoints marked as unqualified, add the qualified viewpoints to the event structure diagram, and then return to S03 again. Repeat this process until the view extraction model converges. Take the prediction result of the converged view extraction model as the final view extraction result.

[0013] The present invention uses the event structure diagram and the progressive framework to help extract new text viewpoints under the same hot topic by using historical information, and realizes high-quality view extraction under the hot topic.

[0014] In the above S01, if the hot topic is general and universal, experts can directly give relevant prior information. If the hot topic is highly professional, it is necessary to learn relevant professional knowledge, invite domain experts, and then give more accurate prior knowledge. Specifically, in the above S01, human experts give the set of perspective categories S = {ac1, ac2,..., ac n} according to their own knowledge for a specific topic. For each perspective category ac i ∈S, a small set of seed words c k that can reflect the category information is provided.

[0015] In the above S02, first connect the internal elements of the view triple (perspective word, view word, and sentiment word), and then judge the perspective category of the view based on the similarity calculation. If the similarity calculation result between the view and the perspective category is greater than the given threshold then assign the perspective category to the view.

[0016] Specifically, in S02, the edges between the lexical nodes in the event structure graph are constructed through the following steps:

[0017] 2.1) Construct the internal edges of the view. Take the elements in the view triple - the perspective word, the view word, and the sentiment word at j , ot j , se j as nodes respectively, and take the at j , ot j edge and the ot j , se j edge. Here, 1 ≤ j ≤ M, and M is the total number of views in the event structure graph.

[0018] 2.2) Calculate the similarity between the perspective word and the perspective category. For each perspective word at j , calculate its cosine similarity wordSim z with the seed word ac j,z in category k, and take the average value of the similarities of all seed words as the similarity categorySim j,k between the perspective word and the perspective category, and transform all the perspective category similarities using the sigmoid function. The similarity calculation formula is as follows:

[0019]

[0020]

[0021] where categoryScore j,k represents the similarity score between the perspective word and the perspective category, ac z represents the word vector of the seed word z belonging to perspective category k, c k represents the set of seed words under perspective category k, |c k | represents the total number of seed words under perspective category k, and z represents the seed word in perspective category k.

[0022] 2.3) Judge the perspective category of the perspective word. If the score of perspective word i and perspective category k is greater than the threshold σ, then connect the node of category k with perspective word i.

[0023] 2.4) Connect all the perspective category nodes with the head node root to enhance the connectivity of the graph.

[0024] In S03, the input text is enumerated to obtain the block hidden representation Sp i,j, where \(0\leq i < j\leq \text{length}(\text{input})\). Use a feed - forward neural network to calculate the probability \(\tau\in\{A, O, \text{null}\}\) of the block word label, where \(A\), \(O\), and \(\text{null}\) represent aspect words, opinion words, and others respectively. Filter out other results, retain the aspect word set and the opinion word set, and combine the elements in the two sets pairwise to obtain the opinion binary - tuple block combination \((\text{aspect}, \text{opinion})\), where \(\text{aspect}\) represents the aspect and \(\text{opinion}\) represents the opinion. Then calculate the sentiment - category attention score \(\alpha\) with the information in the event structure graph τ and the attention score \(\beta\) of the opinion binary - tuple j , and then obtain the evidence vector. After splicing with the opinion binary - tuple block combination representation, perform legality and sentiment judgment. The attention calculation formula is as follows:

[0025]

[0026]

[0027]

[0028]

[0029] where \(W\) τ is the training parameter when calculating the sentiment category \(\tau\), represents the cumulative sum of the concatenation of the aspect word and the opinion word of the opinion binary - tuple under the sentiment category \(\tau\) in the event structure graph, represents the tuple to be classified \([at\) i , ot\) i , and \(\eta\) is the hyper - parameter that adjusts the attention calculation of the aspect word and the opinion word here. represents concatenation, \(k\) represents the total number of opinion binary - tuples under this sentiment category, \(\tau'\) represents a single sentiment in the sentiment category set \(SE\), \(SE\) represents the sentiment category set, \(\alpha\) τ′ represents the attention score of the binary - tuple and the sentiment category \(\tau'\), \(W\) a represents the attention parameter of the aspect word to be learned, \(W\) o represents the attention parameter of the opinion word to be learned, \(K\) represents the total number of tuples under the sentiment category \(\tau\), \(j'\) represents a single tuple under the sentiment category \(\tau\), and \(\beta\) j′ represents the attention score of the tuple \(j'\) under the sentiment category \(\tau\).

[0030] Specifically, in step \(S03\), the text opinion triple extraction result is obtained through the following steps:

[0031] 3.1) Obtain the aspect word and the opinion word. Use the pre - trained model BERT to obtain the vector representation of each token in the input sequence, and use the method of enumeration to obtain the hidden representation \(S_p\) of the block composed of continuous tokens i,j ​, where \(0 \leq i \leq j \leq \text{length}(\text{input})\). The label probabilities of the blocks are calculated using a feed-forward neural network. Each sentiment label \(\tau \in \{A, O, \text{null}\}\) represents belonging to a perspective word, an opinion word, and others respectively.

[0032] 3.2) Generate evidence vectors based on the event structure graph, filter out the blocks labeled as null, and pairwise combine the perspective word intervals and opinion word intervals to form a candidate set \(S\) of sentiment analysis tuples. candidates For each pair of opinion binary tuple blocks \((a_t\) j , \(o_t\) j ) \in S\) candidates , calculate its attention score with the event structure graph and generate an evidence vector. The relevant formulas are as follows:

[0033]

[0034]

[0035] Where, W τ are the opinion tuples and attention parameters under sentiment \(\tau\) in the event structure graph respectively. Specifically, \(p\) τ is the sum of tuple representation vectors under the sentiment category in the graph, \(\alpha\) τ is the attention score of the candidate tuple with the sentiment category \(\tau\) of the event structure graph. For the opinion tuples under the sentiment category, the attention calculation formula is:

[0036]

[0037]

[0038] \(\eta\) is a hyperparameter for adjusting attention. \(W\) a , \(W\) o are the perspective word attention parameter and the opinion word attention parameter respectively. \(\beta\) j is the attention score of the candidate tuple with a single tuple under the sentiment category. Finally, the evidence vector \(E\) can be obtained, and the calculation formula is:

[0039]

[0040] E = \sum τ∈SE E τ \cdot \alpha τ

[0041] 3.3) Tuple legality judgment and sentiment calculation. Concatenate the candidate tuples with the evidence vectors, and calculate the probability of the label \(r \in R \cup \text{null}\) through a feed-forward neural network. \(R\) is the predefined sentiment category, and \(\text{null}\) indicates that there is no any sentiment connection between tuple elements.

[0042] In S04, the present invention performs a word clustering algorithm based on similarity on the output result of S03 to obtain a candidate set of new viewpoints. First, perform k-means clustering on the existing viewpoints in the graph, and then calculate the cosine similarity between the candidate tuple words and each cluster in the clustering result. Determine whether the similarity value is less than the given threshold θ. If all are lower than the threshold, add them to the new viewpoint candidate set Φ, otherwise discard them. Hand the set Φ to an expert for judgment. If the judgment result is Y, it is regarded as a new viewpoint, otherwise discard it. The confirmed new viewpoints use the method described in S02 to obtain the perspective category and add it to the event structure diagram.

[0043] Specifically, in step S04, the updated event structure diagram is obtained through the following steps:

[0044] 4.1) Screen new viewpoints. Use the k-means method to cluster the perspective words and viewpoint words in the event structure diagram. The average value of the words in each cluster of the clustering result is used as the vector representation of the cluster. Calculate the cosine similarity between the extraction model result and each cluster, and transform it using the sigmoid function. If the similarity values of the tuple with all clusters are lower than the threshold θ, then add it to the new viewpoint candidate set Φ.

[0045] 4.2) Confirm new viewpoints. Hand the new viewpoint candidate set Φ to an expert for judgment. If the judgment result is Y, it is a new viewpoint, otherwise discard it.

[0046] 4.3) Update the event structure diagram. Add the new viewpoints to the event structure diagram, and add the corresponding nodes and connecting edges. And use the method described in step S03 to add the perspective category and the connecting edge of the perspective word.

[0047] In S05, use the updated event structure diagram in S04 to continue to complete the training and prediction tasks of the next stage of the corpus until the viewpoint extraction model in S03 reaches convergence.

[0048] A progressive viewpoint extraction system for hot topics adopting the above method, which includes:

[0049] A seed event structure diagram construction module, used to construct a seed event structure diagram based on prior knowledge. The diagram contains viewpoint information nodes and edges representing the relationships between viewpoints; the prior knowledge includes dividing the social texts under the hot topic into multiple stages, giving a set of perspective categories and some viewpoints corresponding to the topic, and adding a small number of seed words that can reflect the category information for each perspective category;

[0050] A viewpoint extraction model training and prediction module, used to combine the updated event structure diagram or the seed event structure diagram in the previous stage to train and predict the viewpoint extraction model for the current stage text stream, and give the prediction result of the current stage text stream after the training ends;

[0051] An event structure diagram update module based on human - machine collaboration is used to filter the prediction results of the opinion extraction model using a word clustering and filtering algorithm based on word similarity, filter out the opinions that already exist in the event structure diagram, submit the new opinions to experts for confirmation, screen the confirmation results returned by the experts, remove the opinions marked as unqualified, add the qualified opinions to the event structure diagram to update the event structure diagram, and then return it to the opinion extraction model training and prediction module again. This process is repeated until the opinion extraction model converges, and the prediction results of the converged opinion extraction model are used as the final opinion extraction results.

[0052] Compared with the prior art, the positive effects of the present invention are as follows: For specific hot topics, historical information is cleverly used to assist in the opinion extraction of new texts, without the need for a large amount of labeled data and complex artificial features, effectively reducing the impact of imbalanced data on the model. Its main innovation points lie in the following two aspects:

[0053] (1) In order to utilize historical comment information under the same topic and thus more accurately mine opinion information, the concept of an event structure diagram is innovatively proposed. In the extraction module, a dual - attention mechanism is used to extract evidence vectors from the event structure diagram to assist the model in opinion extraction training and prediction.

[0054] (2) Inspired by the process of human understanding of events, a progressive framework based on the event structure diagram is proposed. First, a seed event structure diagram is constructed using prior knowledge provided by experts. Then, the extraction module uses the event structure diagram of the previous stage to train and predict the current text stream. Next, the event structure diagram update module based on human - machine collaboration uses the prediction results to update the event structure diagram. This training process is repeated until the extraction model converges. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is an example of the event structure diagram proposed by the present invention;

[0056] Figure 2 It is a schematic diagram of the progressive framework based on the event structure diagram proposed by the present invention;

[0057] Figure 3 It is a schematic diagram of the opinion triple extraction model structure in step S03;

[0058] Figure 4 It is a comparison chart of the effects of progressive training and ordinary training of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0059] In order to enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific examples, but this does not constitute a limitation to the present invention.

[0060] Figure 2 This is a schematic diagram of a progressive framework based on an event structure diagram according to an embodiment of the present invention. Referring to Figure 2 the content, this embodiment specifically includes the following steps:

[0061] S01. Provide prior knowledge. Divide the social texts under hot topics into multiple stages, give the set of perspective categories and partial viewpoints corresponding to the topics in each stage, and for each perspective category, add a small number of seed words that can reflect the category information. Taking the relevant information with restaurant reviews as an example, the following prior knowledge can be set:

[0062] Perspective category 1: Service, waiter, waitress, being served, tip, delivery

[0063] Perspective category 2: Food, wine, menu, sushi, dinner, plate

[0064] In the text regarding restaurant reviews, two perspective categories - service and food - are given as examples, and relevant seed words are given under each perspective category as descriptions of the category. Further, a small amount of labeled data is used to construct a seed event structure diagram. For example, (delivery, conveniently, positive), (waiter, slowly, negative), (food, okay, neutral), (cake, delicious, positive), (baking, memorably, positive), etc.

[0065] S02. Construct a seed event structure diagram. Construct a seed event structure diagram based on the above prior knowledge. The diagram contains view information nodes and edges representing the relationships between views.

[0066] Figure 1 This is an example of the event structure diagram proposed by the present invention. Figure 1 In it, the perspective category nodes are connected to the head node, the perspective word nodes are connected to the perspective category nodes, the opinion word nodes are connected to the perspective word nodes, and the sentiment word nodes are connected to the opinion word nodes. The perspective words, opinion words, and sentiment words form an opinion triple.

[0067] According to the perspective category discrimination method, the viewpoints in S01 can obtain the results shown in Table 1.

[0068] Table 1

[0069]

[0070] Set the threshold σ to 0.6, then the perspective categories corresponding to the opinion words can be obtained as service, service, food, food, food respectively. Connect the corresponding perspective category nodes with the perspective word nodes, and connect all the perspective category nodes with the head node.

[0071] S03. Training and prediction of the opinion extraction model. Assume there are two pieces of data in the current text stream, namely "The cake is very delicious and sweet." and "The waiter is noisy and careless.". The labeled tags are (cake, very delicious, positive), (cake, sweet, positive), (waiter, noisy, negative), and (waiter, careless, negative).

[0072] The process of model training and prediction is as Figure 3 shown and includes the following steps:

[0073] 1) Lexical extraction based on block representation.

[0074] First, obtain the hidden representation of each token. Here, a pre-trained model or a long short-term memory network can be used. Based on the token representation, enumerate all possible block representations of the current sentence. Here, a block is an interval composed of a single or consecutive several tokens. When enumerating in the present invention, a maximum length limit l s is added. For example, for the first piece of data in the text stream, there are blocks such as "egg, cake, cake is very", etc. Use a feed-forward neural network to calculate the label probability, and the perspective word recognition results can be obtained as cake and waiter, and the opinion word recognition results are very delicious, sweet, noisy, and careless. Calculate the loss function loss of the lexical recognition module τ .

[0075] 2) Generation of evidence vectors based on the event structure graph.

[0076] Perform block pair combination on the lexical extraction results obtained in 1). When combining, pair the perspective word sets and opinion word sets in the same data pairwise, and at the same time, limit the block boundaries not to overlap when pairing. Concatenate the text between the two words as the context feature into the representation, as shown in the following formula:

[0077]

[0078] where f(at i ,ot j ) represents the maximum pooling result of the hidden representation of the tokens between the perspective word at i and the opinion word ot j . After the above process, we can obtain the opinion binary tuple block combinations (cake, very delicious), (cake, sweet), (waiter, noisy), and (waiter, careless). For the opinion binary tuple block combination (cake, very delicious), the process of calculating its corresponding evidence vector is shown in the following formula:

[0079]

[0080]

[0081] E = ∑τ∈SE E τ ·α τ

[0082] Among them, in order to better enrich the node information, the present invention uses a Graph Convolutional Network (GCN) on the event structure diagram and allows the existence of self-loops in the graph. is the node representation of node u at the l+1 layer, W l is the parameter to be learned at the l-th layer. In particular, at the 0-th layer uses the glove word vector as the initial node representation, Node u represents node u. α τ and β j are the sentiment category and tuple-level attention scores calculated by the dual attention mechanism defined above. E is the finally generated evidence vector. b l represents the bias coefficient at the l-th layer in the GCN, N(u) represents the set of neighbor nodes of node u, c u represents the normalization limit, A uv represents the value of the adjacency matrix uv, K represents the number of tuples under the sentiment category τ, τ represents the sentiment category, and SE represents the set of sentiment categories.

[0083] 3) Sentiment analysis based on the evidence vector.

[0084] After each candidate tuple obtains the corresponding evidence vector, it is concatenated as the final vector representation of the entire tuple, and the legitimacy label of the tuple is calculated using a feedforward neural network, as shown in the following formula:

[0085]

[0086]

[0087] represents the label probability of the block with start and end positions i and j respectively, W s represents the weight parameter, b s represents the bias parameter, and from this, the loss function loss of the sentiment module can be obtained e . Further, the two loss functions are added together for joint training, as shown in the following formula:

[0088] J = loss τ + loss e

[0089] 4) After the data in this stage is trained after multiple iterations, the training data is predicted to obtain the prediction results: (cake, very delicious, positive), (cake, very sweet, positive), (waiter, noisy, positive), (waiter, careless, negative).

[0090] S04. Confirmation of new viewpoints based on human-machine collaboration. The output results generated in the S03 stage are filtered using a similarity-based clustering method. The existing viewpoints in the event structure diagram are subjected to k-means clustering, resulting in four viewpoint clusters - the first three viewpoints form three clusters, and the last two viewpoints form one cluster. Taking (cake, very delicious) as an example, its similarity scores with the three clusters are 0.12, 0.15, 0.56, and 0.82 respectively. Setting the threshold θ to 0.6, it will be classified into the last cluster. After the above process, (waiter, noisy, positive) and (waiter, careless, negative) are retained as the new viewpoint candidate set and submitted to the expert for judgment. The expert's judgment result for (waiter, careless, negative) is Y, while (waiter, noisy, positive) is an incorrect viewpoint.

[0091] S05. Update of the event structure diagram. (waiter, careless, negative) is used as a new viewpoint to update the diagram. First, the perspective category classification method described in S02 is used to obtain its corresponding perspective category as service, and the careless node and the corresponding connecting edges are added. The updated event structure diagram replaces the old diagram in the previous stage, and then returns to step S03 to assist the viewpoint extraction in the next stage. This process is repeated until the extraction model converges.

[0092] To verify the effectiveness of the progressive framework and the extraction model, the present invention conducts relevant experiments on four public datasets respectively. The four public datasets are Res14, Res15, Res16, and Lap14, which come from the International Semantic Evaluation Conference SemEval 2014, 2015, and 2016 respectively. Res and Lap represent that the datasets come from the restaurant and laptop sales fields. The metrics used in this article are precision, recall, and F1 value.

[0093] The present invention is first compared with the method of removing the progressive update module, as Figure 4 shown. The results show that as more topic-related texts are analyzed and added to the event structure diagram, the model performance continuously improves and stabilizes when the scale of the event structure diagram reaches a certain level.

[0094] In addition, the present invention has searched for some recent representative related works as benchmark models, namely the stage-based opinion extraction method Kwhw (Reference: Peng H, Xu L, Bing L, et al. Knowing what, how and why: A near complete solution for aspect-based sentiment analysis [C] / / Proceedings of the AAAI Conference on Artificial Intelligence: volume 34. 2020: 8600-8607.), and the holistic opinion extraction method JET (Reference: Xu L, Li H, Lu W, et al. Position-aware tagging for aspect sentiment triplet extraction [J]. arXiv preprint arXiv:2010.02609, 2020.). The detailed data is shown in Table 2. As can be seen from Table 2, compared with the benchmark models, the model mentioned in the present invention achieves the best results on almost all datasets.

[0095] Table 2. Results (%) of the present invention and benchmark models on four public datasets, where Ours -g represents the model without using the event structure diagram

[0096]

[0097] Based on the same inventive concept, another embodiment of the present invention provides a progressive opinion extraction system for hot topics using the above method, which includes:

[0098] A seed event structure diagram construction module, which is used to construct a seed event structure diagram based on prior knowledge. The diagram includes opinion information nodes and edges representing the relationships between opinions. The prior knowledge includes dividing social texts under a hot topic into multiple stages, giving a set of perspective categories and some opinions for the corresponding topic, and adding a small number of seed words that can reflect the category information for each perspective category. An opinion extraction model training and prediction module, which is used to train and predict an opinion extraction model for the current stage text stream by combining the event structure diagram or seed event structure diagram updated in the previous stage. After the training is completed, the prediction result of the current stage text stream is given. An event structure diagram update module based on human-computer collaboration, which is used to filter the prediction result of the opinion extraction model using a word clustering filtering algorithm based on word similarity, filter out the opinions that already exist in the event structure diagram, submit the new opinions to experts for confirmation, screen the confirmation results returned by the experts, remove the opinions marked as unqualified, add the qualified opinions to the event structure diagram to update the event structure diagram, and then return to the opinion extraction model training and prediction module again. This process is repeated until the opinion extraction model converges, and the prediction result of the converged opinion extraction model is used as the final opinion extraction result.

[0099] For the specific implementation processes of each module, refer to the description of the method of the present invention above.

[0100] Based on the same inventive concept, another embodiment of the present invention provides an electronic device (such as a computer, a server, a smart phone, etc.), which includes a memory and a processor. The memory stores a computer program, and the computer program is configured to be executed by the processor. The computer program includes instructions for executing each step in the method of the present invention.

[0101] Based on the same inventive concept, another embodiment of the present invention provides a computer-readable storage medium (such as ROM / RAM, a disk, an optical disc), and the computer-readable storage medium stores a computer program. When the computer program is executed by a computer, each step of the method of the present invention is implemented.

[0102] The above embodiments are only illustrative of the principles and effects of the present invention, and are not used to limit the scope of the present invention. Any person skilled in the art who is familiar with this technology can make modifications and changes to the embodiments without departing from the technical principles and spirit of the present invention. The protection scope of the present invention shall be subject to what is described in the claims.

Claims

1. A progressive opinion extraction method for hot topics, characterized in that Including the following steps: S01. Provide prior knowledge, including dividing social texts under hot topics into multiple stages, giving a set of perspective categories and partial viewpoints corresponding to the topic, and adding a small number of seed words that can reflect category information to each perspective category; S02. Construct a seed event structure diagram based on prior knowledge, where the diagram contains viewpoint information nodes and edges representing the relationships between viewpoints; S03. Combine the event structure diagram or seed event structure diagram updated in the previous stage to train and predict the viewpoint extraction model for the current stage text stream. After the training is completed, give the prediction result of the current stage text stream; S04. Use a word clustering filtering algorithm based on word similarity to filter the prediction result of the viewpoint extraction model, filter out the viewpoints that already exist in the event structure diagram, and submit the new viewpoints to experts for confirmation; S05. Screen the confirmation results returned by the experts, remove the viewpoints marked as unqualified, add the qualified viewpoints to the event structure diagram to update the event structure diagram, and then return to S03 again. Repeat this process until the viewpoint extraction model converges. Take the prediction result of the converged viewpoint extraction model as the final viewpoint extraction result; In step S02, the following steps are used to construct the edges between vocabulary nodes in the event structure diagram: 2.1) Construct the internal edges of viewpoints, taking the perspective words, opinion words, and sentiment words at in the viewpoint triples as nodes, and connecting at j , ot j , se j as nodes respectively, and connecting the edges of at j and ot j , and the edges of ot j and se j , where 1 ≤ j ≤ M, and M is the total number of viewpoints in the event structure diagram; 2.2) Calculate the similarity between the perspective word and the perspective category. For each perspective word at j , calculate its cosine similarity wordSim z with the seed word ac j,z in category k. Take the average of the similarities of all seed words as the similarity categorySim j,k between the perspective word and the perspective category, and transform all perspective category similarities using the sigmoid function. The similarity calculation formula is as follows: Among them, categoryScore j,k represents the similarity score between the perspective word and the perspective category, ac z represents the word vector of the seed word z belonging to the perspective category k, c k represents the set of seed words under the perspective category k, |c k | represents the total number of seed words under the perspective category k, and z represents the seed word in the perspective category k; 2.3) Judge the perspective word perspective category. If the score of perspective word i and perspective category k is greater than the threshold σ, then connect the category k node and perspective word i; 2.4) Connect all perspective category nodes to the head node root to enhance the connectivity of the graph; In step S03, the following steps are used to obtain the text viewpoint triple extraction result: 3.1) Obtain the perspective words and opinion words, use the pre-trained model BERT to obtain the vector representation of each token in the input sequence, and use the method of enumeration one by one to obtain the hidden representation Sp of the block composed of consecutive tokens i,j , where 0 ≤ i ≤ j ≤ length(input); use the feed-forward neural network to calculate the label probability of the block, and each sentiment label τ ∈ {A, O, null}, representing belonging to the perspective word, opinion word and others respectively; 3.2) Generate an evidence vector based on the event structure diagram, filter out the blocks marked as null, and form a candidate set S of sentiment analysis tuples by pairwise combination of the perspective word intervals and opinion word intervals candidates ; For each pair of opinion binary tuple blocks combination (at j , ot j ) ∈ S candidates , calculate its attention score with the event structure diagram and generate an evidence vector. The relevant formulas are as follows: Among which W τ is a training parameter when calculating the sentiment category τ, represents the cumulative sum of the perspective words and opinion words in the opinion binary tuple under the sentiment category τ in the event structure diagram, represents the tuple to be classified [at i , ot i , η is a hyperparameter that adjusts the attention calculation of perspective words and opinion words, represents concatenation, k represents the total number of opinion binary tuples under this sentiment category, τ′ represents a single sentiment in the sentiment category set SE, α τ is the attention score of the candidate tuple and the sentiment category τ in the event structure diagram, β j is the attention score of the candidate tuple and a single tuple under the sentiment category, α τ′ represents the attention score of the binary tuple and the sentiment category τ′, W a represents the perspective word attention parameter to be learned, W o represents the opinion word attention parameter to be learned, K represents the total number of tuples under the sentiment category τ, j′ represents a single tuple under the sentiment category τ, β j′ represents the attention score of the tuple j′ under the sentiment category τ; finally, the evidence vector E is obtained, and the calculation formula is: E = ∑ t∈SE E τ ·α τ 3.3) Perform meta - combination legality judgment and sentiment calculation, splice the candidate tuple with the evidence vector, and calculate the probability of label r ∈ R ∪ null through a feed - forward neural network, where R is the predefined sentiment category and null indicates that there is no any sentiment connection between tuple elements.

2. The method according to claim 1, wherein In the step S01, a human expert gives a perspective category set S = {ac1, ac2, …, ac n} according to his own knowledge for a specific topic. For each perspective category ac i ∈S, a small set of seed words c k that reflect the category information is provided.

3. The method according to claim 1, wherein In step S04, the following steps are used to obtain the updated event structure diagram: 4.1) Screen new viewpoints. Use the k - means method to cluster the perspective words and viewpoint words in the event structure diagram. The average value represented by the words in each cluster of the clustering result is used as the vector representation of the cluster. Calculate the cosine similarity between the extraction model result and each cluster, and transform it using the sigmoid function. If the similarity value of the tuple with all clusters is lower than the threshold θ, then add it to the new viewpoint candidate set Φ; 4.2) Confirm new viewpoints. Intersect the new viewpoint candidate set Φ with experts for judgment. If the judgment result is Y, it is a new viewpoint, otherwise discard it; 4.3) Update the event structure diagram. Add the new viewpoints to the event structure diagram, add corresponding nodes and edges, and add the edges between perspective categories and perspective words.

4. The method according to claim 1, wherein In step S05, step S03 uses the event structure diagram updated in step S04 to continue to complete the training and prediction tasks of the next - stage corpus until the viewpoint extraction model in step S03 converges.

5. A progressive viewpoint extraction system for hot topics, which adopts the method described in any one of claims 1 to 4, characterized in that, Including: A seed event structure diagram construction module, used to construct a seed event structure diagram based on prior knowledge, where the diagram contains viewpoint information nodes and edges representing the relationships between viewpoints; Among them, the prior knowledge includes dividing the social texts under hot topics into multiple stages, giving the perspective category set and partial viewpoints corresponding to the topic, and adding a small number of seed words that can reflect the category information for each perspective category; The opinion extraction model training and prediction module is used to train and predict the opinion extraction model for the current stage text stream by combining the event structure diagram or seed event structure diagram updated in the previous stage. After the training is completed, the prediction result of the current stage text stream is given; The event structure diagram updating module based on human-computer collaboration is used to filter the prediction result of the opinion extraction model by using the word clustering filtering algorithm based on word similarity, filter out the opinions that already exist in the event structure diagram, submit the new opinions to the expert for confirmation, screen the confirmation results returned by the expert, remove the opinions marked as unqualified, add the qualified opinions to the event structure diagram to update the event structure diagram, and then return to the opinion extraction model training and prediction module again. This cycle is carried out until the opinion extraction model converges. The prediction result of the converged opinion extraction model is used as the final opinion extraction result.

6. An electronic device, characterized in that, It includes a memory and a processor. The memory stores a computer program, and the computer program is configured to be executed by the processor. The computer program includes instructions for executing the method described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and when the computer program is executed by a computer, the method described in any one of claims 1 to 4 is implemented.

Images