Task definition

Given a sentence containing n words, the objective of the ASTE task is to extract all the triplets of each sentence. is the number of triplets and is the kth triplet in the sentence. represent aspect term, opinion term, and sentiment polarity respectively. a and o consist of one or more words, and .

Feature extraction layer

Semantic encoder.

We first employ the pre-trained language model BERT to extract contextual semantic representation from the input sentence. Given that BERT utilizes subword tokenization, the sequence length of contextual feature representation obtained after BERT encoding may not be consistent with the original sentence length. To address this, an alignment mechanism is employed to aggregate subword-level hidden states into word-level representation, thereby maintaining dimensional consistency.

(1)(2)

where denotes the vector representation of the kth word after BERT coding, and denotes the sequence of subwords after subword tokenization.

After acquiring the contextual semantic representation of the input sentence, the multi-head attention mechanism with relative position information is then employed to construct the attention score matrix of the sentence. The attention scores can reflect the degree of association between words, with higher attention scores indicating a stronger association between two words, and relative position information can focus more on the neighboring words and strengthen the local dependency relationship. The process of constructing the attention score matrix as follows:

(3)(4)

where RMHA denotes the multi-head attention mechanism with relative position information, ContentScores denotes the attention scores between words, RelativeBias denotes the bias of the relative positions of two words, and denotes the dimension of each attention head.

Finally, the attention score matrix and the aggregated hidden state are input into the GCN to obtain a higher-order semantic representation. Fig 3 illustrates the GCN architecture, where A and X are the adjacency matrix and input feature vector respectively, is the normalized adjacency matrix, and are the weight matrix and bias of layer l respectively, Hidden represents the result of , and ReLu is the activation function. In semantic encoder, the whole process can be represented as:

(5)(6)

where and are the weight matrix and bias of the kth layer respectively, is the representation of the kth layer, denotes the representation of the (k+1)th layer, D is the degree matrix, N denotes the number of nodes, and denotes the activation function, and we choose function as the activation function.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. The architecture of GCN.

https://doi.org/10.1371/journal.pone.0340792.g003

The semantically enhanced contextual feature are finally obtained after the convolution operation.

Syntactic encoder.

It is evident that each word possesses a richness of meaning. Within the general domain, each word has a distinct meaning, while in the special domain, the same word assumes different meanings in different contexts. Furthermore, POS plays an important role in the ASTE task as a common form of external grammatical knowledge. In the ASTE task, nouns are frequently employed as potential aspect terms, adjectives are typically considered as potential opinion terms, and other POS also plays a significant role. For instance, degree adverbs can cause sentiment polarity to become entirely opposite. In order to enhance the model‘s capacity to comprehend sentences, a fine-grained character-level embedding is introduced at the initial embedding stage. Therefore, a quadruple embedding is employed to initialize the sentence embedding, which are general domain embedding, special domain embedding, POS embedding and character-level embedding. For POS embedding, five distinct POS tags are selected, namely nouns, verbs, adjectives, adverbs and others to construct a POS embedding table. This table is initialized randomly and subsequently updated during the training process. For character-level embedding, CNN is utilized to obtain character-level embedding for each word. Therefore, the final initial embeddings are as follows:

(7)

where , , , and denote the general domain embedding table, special domain embedding table, POS embedding table and character-level embedding table, respectively.

After obtaining the initial embeddings, the Bi-LSTM is employed to extract the contextual feature of the input sentence. The Bi-LSTM processes the input sequence in both forward and backward directions. The final result, , is obtained by concatenating the results from the forward and backward directions.

(8)(9)(10)

where denotes the dimension of the hidden state of the Bi-LSTM, indicates the concatenation operation.

In order to obtain higher-order syntactic representation, two types of external knowledge are introduced: SenticNet [28] and the syntactic dependency tree. A external knowledge-enhanced adjacency matrix is constructed and modeled using GCNs to capture and enhance the syntactic structure of the sentence. The syntactic dependency tree shows the dependencies between words. In the ASTE task, syntactic dependency analysis can facilitate the identification of the relationship between aspect term and opinion term. As illustrated in Fig 1, the dependencies between words are depicted in a syntactic dependency tree. In SenticNet, sentiment scores are employed to indicate the polarity and intensity of each word, each word possesses a sentiment score ranging from −1–1. Sentiment score that approach −1 for a word signifies that the word is more negative, sentiment score that approach 1 indicates that the word is more positive, and sentiment score of 0 for a word denotes that the word possesses a neutral sentiment polarity or indicates the absence of the word from the SenticNet. The sentiment information of aspect and opinion terms can be enhanced by introducing the SenticNet. An example of SenticNet is shown in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. An example of SenticNet.

https://doi.org/10.1371/journal.pone.0340792.t001

To fully leverage the syntactic information of the sentence, the construction of an external knowledge-enhanced adjacency matrix is required. Specifically, a syntactic dependency matrix is generated through syntactic dependency parsing. The construction rule is as follows: if a dependency relationship exists between two words, the corresponding position in the matrix is set to 1, as well as the diagonal elements. Conversely, if no relationship exists, the corresponding position in the matrix is set to 0. The construction procedure is as follows:

(11)

Following the acquisition of the syntactic dependency matrix, the sentiment intensity adjacency matrix is then constructed, and the sentiment information between two words is computed through summation of their individual sentiment scores. The construction rule for the sentiment intensity adjacency matrix is as follows:

(12)

The final external knowledge-enhanced adjacency matrix is derived through integration of the syntactic dependency matrix T with the sentiment intensity adjacency matrix Z.

(13)

Finally, external knowledge-enhanced adjacency matrix and the contextual representation derived from the Bi-LSTM are input into the GCN to obtain a higher-order syntactic representation, and the whole process can be represented as:

(14)(15)

where and are the weight matrix and bias of the th layer respectively, is the representation of the kth layer, denotes the representation of the (k+1)th layer, D is the degree matrix, N denotes the number of nodes, and denotes the activation function, and we choose function as the activation function.

The syntactically enhanced contextual feature are finally obtained after the convolution operation.

Feature fusion layer

Following the acquisition of syntactically enhanced contextual feature and semantically enhanced contextual feature, an interaction mechanism for feature fusion is designed. Firstly, a self-attention mechanism is applied to the syntactically enhanced contextual feature and the semantically enhanced contextual feature, respectively, in effects to more effectively capture the dependencies between the features. Subsequently, the cross-attention mechanism is employed to derive the syntax-enhanced semantic feature and the semantics-enhanced syntactic feature .

The self-attention mechanism is formulated as:

(16)

where denotes the input sequence, denote the weight matrixs, and denotes the dimension of the hidden state, respectively.

Therefore, self-attention operations are performed on and , respectively:

(17)

(18)

The cross-attention mechanism is formulated as:

(19)

where X,Y denote the input sequences, denote the weight matrixs, and denotes the dimension of the hidden state, respectively.

Therefore, cross-attention operations are performed on and , respectively:

(20)

(21)

Consequently, in order to prevent the loss of critical features, we consider the initially obtained semantic feature and the syntactic feature and design a gating mechanism to fuse the four features to obtain the final feature representation .

(22)

where denote and respectively.

The final fused vector is represented as:

(23)

Triplet extraction layer

The BDTF [10] method is utilized for the extraction of triplets. A distinguishing feature of the BDTF method is its construction of a table, identification of the regions by S and E tags, and subsequent application of sentiment judgement to the regions. This approach diverges from other table-filling methods. As illustrated in Fig 4, an example of BDTF is presented. The S tag denotes the intersection point between the starting positions of the aspect term and the opinion term, while the E tag represents the intersection point between their ending positions. The region determined by the S tag and E tag constitutes the relationship region.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. An example of BDTF.

https://doi.org/10.1371/journal.pone.0340792.g004

Firstly, it is necessary to construct the relation-level representation. The final feature representation, , is used to construct the relation-level representation. Contextual information between words is also significant, which is obtained by means of a maximum pooling operation. The representations of the two words of the word pair, the contextual representation, and the interaction representation of the two words are concatenated. The concatenated representations are then passed through a nonlinear projection and a gelu activation function [29] to obtain the final relation-level representation. The representation of each word pair can be expressed by the following equation:

(24)

(25)

(26)

where denotes the maximum pooling operation, denotes the interaction representation between two words, and is the tensor parameter.

Following the acquisition of the relation-level representation of each word pair, the relation-level representations of all word pairs are assembled to form a 3D relation table . There may be some potential dependencies between relation-level representations of word pairs, and to extract such dependencies, the L layer ResNet-style CNN [30] is adopted. The table representation obtained from the l layer CNN is denoted as :

(27)

(28)

(29)

(30)

where denotes Layer Normalization [31] and the initial input table .

The final table representation, , is obtained and subsequently two biaffine attention networks [24] are used to enhance the representations of S and E tags. In order to prevent overfitting, two multi-layer perceptrons are employed.

(31)

(32)

The formula of biaffine attention network as follows:

(33)

where and are trainable weights and biases, we can obtain and by performing biaffine operation on S and E tags, respectively, and then we can obtain the probability scores by performing sigmoid operation on and :

(34)

(35)

Subsequently, the probability scores obtained after the biaffine attention network are utilized to enhance the representations of the original S and E tags. Following the acquisition of the enhanced representations, two classifiers are employed to determine the candidate S and E tags. Thereafter, the top-k strategy is implemented to prune the candidate S and E tags, resulting in the identification of the final boundary S and E tags.

(36)

(37)

With the final boundary S and E tags, the relationship regions have been determined. Given and , if satisfy ,the relation region is a valid region. Once the relationship region is determined to be a valid region, then aspect terms and opinion terms can be extracted, since there exists the aforementioned correspondence a between S tags, E tags, aspect terms, and opinion terms. For sentence , where aspect terms are and opinion terms are . Fig 5 demonstrates an example of aspect term and opinion term extraction.For each relation region, the representation of the S tag, the representation of the E tag, and the maximal pooling result of the relation region are used to construct a representation of the relation region. Finally, a classifier is employed to predict the sentiment polarity of the relation region:

(38)

(39)

(40)

where .

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Correspondence between S tags, E tags, aspect terms, and opinion terms.

https://doi.org/10.1371/journal.pone.0340792.g005

Boundary-aware contrastive learning

In ASTE task, it is imperative to accurately identify S and E tags. Aspect terms, opinion terms and relationship regions are determined by S and E tags. To facilitate the model‘s ability to identify S and E tags, a boundary-aware contrastive learning module has been developed. This module employs contrastive learning on both S and E tags to optimize their respective representations, thereby enhancing the model‘s recognition capability.

Initially, each word pair representation in the complete table representation is represented as vector representations. After processing through the ResNet CNN, the complete table representation is obtained, which consequently the representation for each individual word pair has also been obtained. The table representation is denoted as , thus the representation of each word pair is denoted as .

After obtaining the representation of each word pair, the next is to construct positive and negative samples. For multi-word aspect terms and multi-word opinion terms, the model may bias the identification of S and E tags, resulting in inaccurate identification of the start and end positions of aspect terms and opinion terms. To address this, a methodology is employed that involves the selection of positive and negative samples. Specifically, the correct S tags within the same triplet are designated as positive samples, while the surrounding positions of the correct S tags are utilized as negative samples.The correct S tag within the same triplet is first retrieved and represented as , and and represent the correct S tag and S tag representation of the ith triplet, respectively. By superimposing Gaussian random noise on the correct S tag representation, the positive sample representation and anchor sample representation are obtained, where the noise follows a normal distribution with a mean of 0 and a standard deviation of 0.1. This process can be formally expressed as:

(41)

(42)

The surrounding positions of correct S tag are selected as negative samples, and negative sample representations can be represented as , , , . For the ith triplet, its negative sample representation is denoted as .

Subsequent to the construction of the positive and negative samples, the contrastive learning loss function is utilized to calculate the distance between the positive and negative samples and to update the model parameters. In our approach, Triplet Contrastive Loss is employed to minimize the distance between anchor samples and positive samples while maximizing the distance between anchor samples and negative samples, thereby learning discriminative feature representations. Thus, the contrastive loss for the S tag of the ith triplet is:

(43)

where N denotes the number of samples of the ith triplet, denotes the vector representation of the jth negative sample in the ith triplet.

Therefore, he contrastive loss for S tags can thus be expressed as:

(44)

where M denotes the number of triplets.

Similarly, the contrastive loss can be obtained for the E tags.

Training loss

The final loss consists of three components, the loss of boundary detection , the loss of region classification and the loss of contrastive learning .

(45)

For the loss of boundary detection :

(46)

(47)

(48)

where and are the true boundary tags, and are the predicted boundary tags.

For the loss of region classification :

(49)

where is the true region type.

For the loss of contrastive learning :

(50)

where and denote the contrastive loss of the S tags and the contrastive loss of the E tags, respectively.

Datasets

We conducted comprehensive evaluations of our proposed method on two benchmark datasets for ASTE task: ASTE-Data-V1 [10] and ASTE-Data-V2 [11]. Both datasets comprise four sub-datasets with distinct domains: three restaurant review datasets (14res, 15res, and 16res) and one laptop review dataset (14lap). Table 2 presents the detailed statistics of these datasets.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. The detailed statistics of ASTE-Data-V1 and ASTE-Data-V2 datasets.

https://doi.org/10.1371/journal.pone.0340792.t002

Experimental settings

In the semantic encoding module, “bert-base-uncased” is employed as the encoder, and the layers of GCN are two layers. function is used in GCN architecture. The feature dimensions for both input and output representations of each GCN layer are 768. And random-walk normalization is employed to preprocess the adjacency matrix. In the syntactic encoding module, Spacy is utilized for syntactic dependency parsing and NLTK is employed for POS tagging. The hidden layer dimension of the Bi-LSTM is 300. In the word embedding module, we use a 100-dimensional 1D convolutional kernel for CNN character embedding and 300-dimensional pre-trained word vectors from Glove [32] for general domain word embedding. We also use 500-dimensional word embeddings from the Amazon Product Review Corpus [33] for special domain word embedding and 100-dimensional random initial vectors for POS embeddings. The GCN uses two layers. The feature dimensions for both input and output representations of each GCN layer are 600. And symmetrically normalization is employed to preprocess the adjacency matrix. In GCN, both weight matrices and bias terms are initialized by sampling from a uniform distribution. The number of layers is set to 1 for both MLPs, each with an input dimension of 768 and an output dimension of 300. fuction is employed in MLPs. The version of SenticNet is SenticNet 6. AdamW [34] is used as the optimizer. The model is trained for 15 epochs using different random seeds, and the model that performs best on the validation set is selected to test on the testing set.

Baselines

In order to demonstrate the effectiveness of our proposed method, we compared it with several state-of-art ASTE methods. These ASTE methods can be divided into three categories: pipeline-based methods, span-based methods and joint extraction methods.

Pipeline-based methods

• Peng-two-stage: [10] proposed a two-stage framework for the ASTE. In the initial stage, candidate aspect terms, along with their corresponding sentiment polarity and candidate opinion terms are extracted. In the next stage, aspect terms, sentiment polarity and opinion terms are paired.

Span-based methods

• Span-ASTE: [15] was the first to explicitly model interactions between aspect spans and opinion spans in order to address sentiment inconsistent of word pair interactions.
• Literature [27]: [27] proposed a framework for bidirectional extraction that was based on spans. The design of two decoders from aspect to opinion and from opinion to aspect was achieved, and bi-directional extraction was realized through the modelling of the relationship between spans by means of a multi-head attention mechanism.

Joint extraction methods

• JET: [11] proposed an innovative position-aware tagging scheme for joint sentiment triplet extraction. The method is capable of more effective capture of interactions among sentiment elements.
• GTS: [12] proposed a grid tagging scheme and developed an inference strategy to explore the relationships between different opinion factors.
• EMC-GCN: [22] employed multi-channel graphs to capture dependency relationships between words and enhanced the model with linguistic features.
• BDTF: [16] employed relation regions in a 2D table to represent sentiment triplets, thereby converting the ASTE task into a detection and classification task for relation regions.
• SA-Transformer: [23] incorporated dependency types into graph neural networks and proposed AEA, which learns different representations and weights for each edge through dependency types.
• Literature [21]: [21] proposed a novel grid tagging scheme, representing an extension of the GTS. The addition of four tags, , to GTS is intended to enhance the boundary recognition of aspect and opinion terms.
• PBLUN: [24] proposed a POS-based label update module to provide more accurate tags of aspect terms and opinion terms for subsequent tasks, and utilized two biaffine attention networks to enhance the representation of S and E tags in the BDTF method.

Experimental results

As demonstrated in Table 3, the experimental results of our method and other benchmark models on the four datasets of ASTE-Data-V2 are presented. Similarly, Table 4 shows the experimental results of our method and other benchmark models on the four datasets of ASTE-Data-V1. The evaluation of the models employs three evaluation metrics: F1 score (F1), Precision (P) and Recall (R). The experimental results demonstrate the efficacy of the proposed method on all datasets.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Experimental results on ASTE-Data-V2.

https://doi.org/10.1371/journal.pone.0340792.t003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Experimental results on ASTE-Data-V1.

https://doi.org/10.1371/journal.pone.0340792.t004

As presented in Table 3, compared to the Peng-two-stage model, our proposed method achieves a , and enhancement in the F1 score metrics, respectively. In comparison with the latest span-based method [27], our method enhances the F1 scores by , and , respectively. When evaluated against the BDTF, our method shows enhancements of , , , on four datasets. In comparison with the PBLUN framework, our method demonstrates improvements of , and on the 14lap, 15res, and 16res datasets respectively.

As shown in Table 4, compared with the Peng-two-stage model, our method achieves F1 score improvements of 19.07%,24.07%,21.10%,21.43% on the four datasets, respectively. Relative to the table-filling based method BDTF, our method achieves improvements of 1.58%, 0.63%, 1.92%, 1.54% in F1 score.

The experimental results demonstrate the efficacy of the proposed method, substantiating its superior capacity to extract the deep syntactic and semantic information of sentences. Furthermore, it is evident that the efficacy of boundary-aware contrastive learning module in our approach is effective in enhancing the representation of S and E tags.

Ablation study

In order to evaluate the validity of the different modules in our method, we conducted an ablation study on four datasets from ASTE-Data-V2, and chose the F1 score as the evaluation metric. As shown in Table 5 and Fig 6, the experimental results of the ablation study are presented. “w/o SenticNet” indicates that no SenticNet is utilized to enhance the adjacency matrix, “w/o Double” indicates that only a single encoder is employed, and “w/o Fusion” indicates that the feature fusion module is not used and only the features are simply added. “w/o MHA-score” denotes the absence of the construction of an attention score matrix with relative position information, with the results of the BERT encoder being utilized directly instead. “w/o CL” indicates the absence of the use of boundary-aware contrastive learning module to enhance the representations of S and E tags.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. The results of ablation study on ASTE-Data-V2.

https://doi.org/10.1371/journal.pone.0340792.t005

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. The experimental results of ablation study on ASTE-Data-V2.

https://doi.org/10.1371/journal.pone.0340792.g006

In comparison to our proposed method, employing the BERT single encoder for encoding results in a significant decline in model performance, with a decrease in F1 scores by , and on the four datasets, respectively. The construction of the adjacency matrix without the utilization of the SenticNet resulted in a decline in the F1 scores across the four datasets, with a decrease of 1.98%, 1.54%, 3.03%, and 2.09%, respectively. This demonstrates that incorporating the sentiment information of individual words enhances the extraction of sentiment triplets. Without the feature fusion mechanism, simply features addition results in the loss of valuable features and consequently leads to inaccuracy in extracting sentiment triplets. The employment of features addition leads to a decline in F1 scores by 8.69%, 1.34%, 2.51% and 3.13%.

The study of the boundary-aware contrastive learning approach demonstrates that the absence of the boundary-aware contrastive learning module results in a decline in F1 scores on the four datasets, with a decrease of 1.98%, 1.99%, 1.59%, and 1.13%, respectively. The incorporation of the boundary-aware contrastive learning module facilitates the model‘s capacity to accurately identify boundary S and E tags, thereby enhancing the effect of triplets extraction.

The investigation into the attention score matrix reveals that not constructing the attention score matrix with relative position information leads to a decline in F1 scores by 2.61%, 1.33%, 3.87%, and 1.74%, respectively, when compared with our method on the four datasets. This suggests that the attention score matrix with relative position information reflects the different importance of the interactions between the words and makes the model more capable of identifying pairs of words with stronger relations, thereby facilitating the recognition of aspect terms and opinion terms.

Effect of the number of GCN layers

In order to explore the effect of the number of GCN layers on the performance of the method, experiments were conducted on the four datasets of ASTE-Data-V2. The number of GCN layers was increased from 1 to 5, and the F1 score was used as the evaluation metric. As demonstrated in Fig 7, it is evident that the maximum F1 score for each dataset is obtained when the number of GCN layers is configured to 2. When the number of GCN layers is set to 1, the feature extraction is inadequate, resulting in the failure to extract useful features. When the number of GCN layers exceeds 2 layers, it will cause overfitting problem, and it is also easy to introduce redundant information that will cause model performance degradation. Consequently, the number of GCN layers was selected 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. The experimental results of different number of GCN layers.

https://doi.org/10.1371/journal.pone.0340792.g007

Case study

A case study was conducted to demonstrate the effectiveness of the proposed method. We selected seven examples from the ASTE-Data-V2 dataset for the purpose of comparing our method with the BDTF method. The seven sentences can be categorized into six distinct types based on failure types: sentiment polarity, comparative constructions, overlapping aspect terms, boundary identification, prepositional phrases, and long-range coordination. The results of the study are presented in Table 6.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Case study.

https://doi.org/10.1371/journal.pone.0340792.t006

Two examples are provided to illustrate errors in sentiment polarity prediction. In the first instance, the opinion term is “longer” and the aspect term is “baterry” which should correspond to a “positive” sentiment polarity. However, the BDTF method incorrectly predicts the sentiment polarity as “negative.” In the second example, both our method and BDTF successfully extract two triplets. Nevertheless, BDTF method incorrectly predicts the sentiment polarity of the triplet (Windows 7, crashed, negative) as “neutral,” potentially due to its failure to adequately incorporate contextual semantic information.

Regarding the example containing a comparative construction, the sentence contains two triplets. While the BDTF method accurately predicts both triplets, our method identifies only one. Specifically, our method fails to predict the triplet (decor, creative, positive) due to its inability to recognize the comparative construction “as…as…”. This limitation may be attributed to insufficient semantic comprehension of this specific type of comparative construction.

In the case of overlapping aspect terms, both our method and BDTF accurately identify all triplets within the sentence, including the count of triplets, the recognition and pairing of aspect and opinion terms, and the prediction of sentiment polarity.

For the example concerning boundary identification, which contains three triplets. Both our method and BDTF successfully identify the number of triplets. However, our method achieves perfect accuracy across all triplets, whereas BDTF predicts one erroneous triplet due to the incorrectly opinion terms extraction. This result demonstrates the effectiveness of the boundary-aware contrastive learning module in our approach, confirming its capability to improve boundary detection.

The example involving prepositional phrase contains one triplet. BDTF predicts a single triplet, while our method predicts two triplets. However, all predicted triplets from both methods are incorrect. In this sentence, the aspect term should be “chicken in the salads.” However, the BDTF method identifies only “chicken” as the aspect term, while our approach erroneously extracts both “chicken” and “salads” as aspect terms. The model fails to recognize that “in the salads” is the attribute of “chicken,” leading to incorrect aspect term identification. The cause may lie in the model‘s limited ability to understand complex syntactic structures.

In the case involving long-range coordination, the sentence contains three triplets with overlapping opinion terms. Our method successfully extracts all triplets, but BDTF identifies only two triplets. The BDTF method fails to capture the aspect-opinion relationship following the conjunction “and” likely due to its limitations in parsing long-range coordination and its incomplete understanding of global syntactic dependencies.

In summary, the BDTF method demonstrates insufficient understanding of the semantic information of sentences, and fails to leverage the syntactic information of sentences and external knowledge to enhance the model‘s extraction of potential syntactic and semantic features of sentences. Compared with the BDTF, we use the external knowledge of syntactic dependency tree, part of speech and SenticNet to enhance the model‘s ability to understand sentences and extract triplets more accurately. However, through the above examples, it can be found that our model also has some limitations in dealing with some more complex syntactic structures. Future work should focus on incorporating deep syntactic analysis and fine-grained semantic reasoning to improve the robustness and performance of triplet extraction.

Computational efficiency and resource analysis

We conducted experiments on the 14lap and 14res datasets of ASTE-Data-V2, and reported the training and inference times, parameter counts and GPU memory usage. The results are shown in the Table 7. The experiments were conducted on 4090-24G GPU. The training time refers to the time required to complete one training epoch, while the memory usage refers to the peak GPU memory usage during one training epoch.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 7. The comparison of training time,inference time, parameter counts and memory usage.

https://doi.org/10.1371/journal.pone.0340792.t007

Compared with BDTF method, our method has increased the training and inference times and parameter counts. This is attributed to the incorporation of external knowledge such as syntactic dependency tree, part of speech and SenticNet, along with the design of dual encoders and a boundary-aware contrastive learning module, which collectively contribute to a more complex model architecture. However, our method achieves a slight reduction in GPU memory usage relative to the BDTF method. In future work, we will prioritize improving computational efficiency and inference time of our method.

Significance tests

To further validate the effectiveness of our proposed method, we conducted significance tests on the four datasets of ASTE-Data-V2. The experiments employed the F1 score as the evaluation metric, reporting the average F1 scores and their standard deviations obtained under five different random seeds. We selected BDTF and the PBLUN model based on BDTF as comparative models, and the specific experimental results are shown in Table 8.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 8. Experimental results of significance tests on ASTE-Data-V2.

https://doi.org/10.1371/journal.pone.0340792.t008

Experimental results demonstrate that, compared to BDTF, our method achieves average F1 score improvements of 1.4%, 0.38%, 1.97%, and 1.75% on the four datasets, respectively. When compared to PBLUN, our method achieves improvements in average F1 score of 0.63%, 1.21%, and 1.13% on the 14lap, 15res and 16res datasets, respectively, while experiencing a slight decline of 0.3% on the 14res dataset. The standard deviations obtained across the four datasets are 0.48, 0.32, 0.20, and 0.53, respectively. The relatively large standard deviations observed on 14lap and 16res indicate more pronounced variability in the single-run experimental results on these two datasets, whereas the results on 14res and 15res remain relatively stable.

The results indicate that the proposed method demonstrates superior average performance, suggesting its enhanced capability in extracting the features of sentences effectively. However, significant fluctuations observed in two datasets may be attributed to the highly complex architecture of the model. Consequently, in subsequent research, we will consider implementing lightweight modifications to our proposed model to improve its stability.