Aspect emotion triple extraction method for implicit recognition enhanced table filling

Abstract

The invention discloses an aspect emotion triple extraction method for implicit recognition enhanced table filling, and belongs to the field of emotion analysis. Comprising the following steps: modeling an input text and an overall implicit expression thereof by using a pre-trained encoder to obtain a context expression, and deploying a bidirectional information aggregation module to aggregate forward and backward information of two words in a word pair based on the context expression to construct a table expression, all aspect emotion triples are extracted from the tabular representation of the sentence using a decoding module. According to the method disclosed by the invention, the identification of the implicit aspect terms is enhanced, so that the table filling method can extract the aspect emotion triad from the text containing the implicit expression. Besides, the aggregation module fusing forward and reverse information not only integrates the characteristics of a single word, but also integrates the forward and reverse relationships between two words in a word pair, thereby effectively improving the performance of aspect emotion triple extraction.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of sentiment analysis, and particularly relates to an implicit recognition enhanced table filling aspect sentiment triple extraction method. BACKGROUND

[0002] Analyzing users' sentiment tendency on commodities through Internet comments has become a key research direction in the field of natural language processing, and the purpose is to help enterprises improve products and services. Aspect-based sentiment analysis (ABSA) is a task in the field of natural language processing, including aspect term extraction (ATE) and opinion term extraction (OTE) subtasks. Aspect sentiment triple extraction (ASTE) is a more fine-grained subtask, which integrates multiple subtasks in ABSA. The goal of ASTE is to extract all aspect sentiment triples from text, each triple consisting of an aspect term, an opinion term and the sentiment polarity corresponding to the aspect term.

[0003] ASTE task and three datasets from the restaurant domain, one from the notebook computer domain are proposed. Existing ASTE methods have shown good performance on these datasets, thus promoting the development of ASTE task. However, these datasets still maintain a relatively simple nature, which is different from real-world user comments. These differences include comment length, expression diversity and domain distribution. In order to bridge the gap between existing datasets and real-world user comments, more diversified DMASTE datasets are proposed, which are characterized by expanding the domain range, longer user comments, more diverse expressions, and annotating triples with implicit and explicit aspect terms. Existing methods have certain limitations when facing more complex DMASTE datasets: 1) In the DMASTE dataset, there are triples containing explicit and implicit aspect terms. The triple containing implicit aspect terms has no corresponding aspect term in the sentence, which makes the existing table filling method invalid in this case; 2) When constructing table representation through sentence context representation, the representations of the two words in the word pair are connected, and then the relationship between them is modeled, which leads to insufficient aggregation of effective information between words, thus hindering the subsequent modeling process from obtaining the maximum effective information. SUMMARY

[0004] The application aims to solve the deficiencies in the prior art and provide an implicit recognition enhanced table filling aspect sentiment triple extraction method. The method enhances the recognition of implicit aspect terms and uses a bidirectional information fusion module to fuse the forward and reverse relationships between the two words in the word pair, thus improving the accuracy of aspect sentiment triple extraction.

[0005] The aspect sentiment triple extraction method for implicit recognition and enhanced table filling includes using a pre-trained encoder to model input text and its overall implicit expression to obtain context representation, deploying a bidirectional information aggregation module to aggregate forward and backward information of two words in a word pair based on the context representation to construct a table representation, and using a decoding module to extract all aspect sentiment triples from the table representation of the sentence.

[0006] The aspect sentiment triple extraction method for implicit recognition and enhanced table filling adds a special token representing the implicit expression feature of the entire sentence at the end of the input text, and then encodes the text into a BERT encoder to obtain context representation.

[0007] The aspect sentiment triple extraction method for implicit recognition and enhanced table filling fuses the context representation of each word in a word pair through the integration of forward and backward information. For diagonal positions, each word is fused with the representation of a special token representing the implicit expression feature of the entire sentence. This process generates a table representation based on the positions of the two words.

[0008] The aspect sentiment triple extraction method for implicit recognition and enhanced table filling uses a boundary detection decoding module to identify aspect sentiment triples with explicit aspect items and aspect sentiment triples with implicit aspect items using previously obtained table representations.

[0009] The aspect sentiment triple extraction method for implicit recognition and enhanced table filling has the following advantages: 1. A special token representing the implicit expression feature of the entire sentence is introduced in the sentence before encoding. This token is then used in the decoding stage to extract aspect sentiment triples containing implicit aspect items. 2. A bidirectional information aggregation module is designed to integrate forward and backward information of two words in a word pair to create a more comprehensive table representation, provide richer representation, and improve model performance. BRIEF DESCRIPTION OF DRAWINGS

[0010] Figure 1 A model diagram of an aspect sentiment triple extraction method for implicit recognition and enhanced table filling embodiment.

[0011] Figure 2 A structure diagram of a bidirectional information aggregation module in an aspect sentiment triple extraction method for implicit recognition and enhanced table filling embodiment. DETAILED DESCRIPTION

[0012] The following will further explain the content of the present application by combining embodiments.

[0013] As shown in Figure 1 An aspect sentiment triple extraction method for implicit recognition and enhanced table filling includes the following steps:

[0014] 1) Given a text with words , since the implicit aspect terms are not explicitly expressed in the sentence, their existence must be inferred from the context content. First, a special token [IE] representing the implicit expression features inherent in the whole sentence is defined in the BERT vocabulary and is attached to the end of the sentence. Then, the [CLS] token representing the beginning of the sentence and the [SEP] token representing the end of the sentence are added to the beginning and end of the sentence. Subsequently, the context representation of the sentence is obtained by capturing the context relationship through the bidirectional pre-training model BERT. This process can be briefly described as:

[0015] 2) The context representation of the sentence is input into the bidirectional information aggregation module to obtain the table representation. As shown in Figure 2 , the bidirectional information aggregation module includes:

[0016] 2.1) Extract the representation of the special token [IE] from the context representation of the sentence, and copy it n times as the implicit expression representation, and the remaining representation is used as the context representation of the sentence.

[0017] 2.2) The context representation of the sentence is copied and expanded into two 2D tables, represented as and . The row elements of , The column elements of . For forward information aggregation, , the information of its neighborhood is aggregated through a 3x3 convolution layer to obtain the output feature . Then, is concatenated after and further aggregated through another 3x3 convolution layer to obtain the output feature . This process can be briefly described as:

[0018] For backward information aggregation, similarly, is passed through a 3x3 convolution layer to obtain the output feature , and then is concatenated after , and another 3x3 convolution layer is used to obtain the backward information aggregation feature . This process can be briefly described as: Then, the four output features are concatenated and processed by a feature fusion layer consisting of a multi-layer perceptron, a 3x3 convolutional layer and a 1x1 convolutional layer to smooth the concatenated features and obtain the table representation after bidirectional information aggregation . This process can be represented as: where is an activation function, is layer normalization.

[0019] 2.3) Since the diagonal elements in only aggregate the bidirectional information between each word and itself, it can lead to insufficient information utilization when identifying aspect-pair with implicit aspect terms in subsequent operations. Therefore, the same bidirectional aggregation operation is performed on the context representation and the implicit representation to obtain an enhanced 1D representation of the implicit expression .

[0020] Finally, the diagonal elements in the 2D representation are replaced by the 1D enhanced representation to obtain the final table representation .

[0021] 3) The boundary detection and classification of aspect-pair based on the table representation has superior performance in aspect sentiment triplet extraction. Therefore, the previously obtained table representation is used to identify the boundary region of aspect-pair with explicit aspect terms using a decoding module, and the boundary region of aspect-pair with implicit aspect terms is detected along the diagonal line, followed by aspect sentiment triplet extraction. The decoding module includes:

[0022] 3.1) For aspect-pair with explicit aspect terms, the boundary region is delimited by the start position S-E and the end position E-E. For each non-diagonal element in the table representation , two classifiers are used to calculate its probability as labels S-E and E-E. This process can be described as: where is an activation function, is a linear layer.

[0023] 3.2) For aspect-pair with explicit aspect term, the boundary region is delimited by the start position S-E and the end position E-E. For each diagonal element in the table representation , two classifiers are used to compute its probability S-I and E-I as labels. This procedure can be summarized as:

[0024] 3.3) After obtaining the boundary region, the S-E, E-E and max-pooling results of the relation matrix on this region are combined to form the feature representation of aspect-pair with explicit aspect term . The S-I, E-I and max-pooling results of the relation matrix on this region are combined to form the aspect-pair feature representation with implicit aspect term . Then the classifier is used to predict the sentiment polarity corresponding to the aspect-pair. This procedure can be summarized as:

[0025] 4) For triple with explicit aspect term, given the ground truth boundary label , the loss function of boundary region detection is as follows: For triple with implicit aspect term, given the ground truth boundary label , the loss function of boundary region detection is as follows: Given the type label of sentiment polarity , the loss of region classification is calculated as: The total loss of the joint training model is:

[0026] To verify the effectiveness of the implicit recognition enhanced aspect sentiment triple extraction method proposed in this embodiment, the Electronics, Beauty, Fashion and Home four data sets in the DMASTE data set are selected for verification, and the specific experimental results are shown in Table 3:​​​

[0027] Table 3 ASTE task experiment F1 score results

[0028] Model Electronics Beauty Fashion Home BMRC 41.95 38.57 44.87 41.18 BART-ABSA 43.38 41.13 43.89 40.56 GAS 47.10 44.32 47.80 47.22 Span-ASTE 47.86 46.46 50.38 49.14 The present example 48.15 47.46 50.60 49.65

[0029] From the experimental results in Table 3, compared with the optimal baseline method Span-ASTE, the F1 score of the method proposed in the embodiment is improved by 0.29, 1.00, 0.22 and 0.51 on the four data sets respectively. It shows the superior performance of the proposed implicit recognition aspect of enhancing table filling sentiment triple extraction method.

Claims

1. An implicit aspect sentiment triple extraction method for enhancing table filling, comprising using a pre-trained encoder to model input text and its overall implicit representation to obtain a context representation, deploying a bidirectional information aggregation module to aggregate the forward and backward information of two words in a word pair based on the context representation to construct a table representation, and using a decoding module to extract all aspect sentiment triples from the sentence from the table representation.

2. The aspect sentiment triple extraction method of implicit recognition enhanced table filling according to claim 1, wherein: The pre-trained encoder is used to model the input text and its overall implicit representation to obtain the context representation. Specifically, a special token representing the implicit representation feature of the entire sentence is added at the end of the input text, and then the text is input into the BERT encoder for encoding to obtain the context representation.

3. The aspect sentiment triple extraction method of implicit recognition enhanced table filling according to claim 1, wherein: A bidirectional information aggregation module is deployed to aggregate the forward and backward information of two words in a word pair based on the context representation to construct a table representation. The specific steps are to integrate the context representation of each word in a word pair through the integration of forward and backward information. For diagonal positions, each word is fused with the representation of the special token representing the implicit representation feature of the entire sentence. This process generates a table representation based on the positions of the two words.

4. The aspect sentiment triple extraction method of implicit recognition enhanced table filling according to claim 1, wherein: A decoding module is used to extract all aspect sentiment triples from the table representation of the sentence. Specifically, a boundary detection decoding module is used to identify aspect sentiment triples with explicit aspect items and aspect sentiment triples with implicit aspect items using the previously obtained table representation.

Images