Prerpints.org logo
Preprint
Article

Enhanced Sentiment Analysis with Syntactic Dependency and Advanced Graph Convolution Model

Altmetrics

Downloads

104

Views

31

Comments

0

This version is not peer-reviewed

Multimodal Sentiment Analysis Based on Deep Learning Methods Such as Convolutional Neural Networks

Submitted:

28 November 2023

Posted:

29 November 2023

You are already at the latest version

Alerts
Abstract
This paper presents the Advanced Syntactic-Graph Convolutional Model (ASGCM), a pioneering approach in Aspect-Based Sentiment Analysis (ABSA) that integrates syntactic dependency features within a graph convolution framework. ASGCM stands out for its novel use of dependency edge encoding and tag-based graph convolutions, providing a fine-grained analysis of sentiments associated with specific aspects in text. This model meticulously captures the intricacies of syntactic structures, thereby offering enhanced precision in sentiment analysis. Notably, ASGCM incorporates a dual-layer graph convolution system: one layer processes syntactic dependencies (edges), while the other interprets semantic roles (tags), ensuring a comprehensive understanding of both structural and contextual elements in text. We rigorously tested ASGCM on multiple datasets, including both English and Chinese languages, and our findings reveal a significant improvement in sentiment classification accuracy compared to existing models. The versatility of ASGCM makes it a robust tool for diverse linguistic environments, setting a new standard for ABSA methodologies.
Keywords: 
Subject: Computer Science and Mathematics  -   Artificial Intelligence and Machine Learning

1. Introduction

Aspect-Based Sentiment Analysis (ABSA), a nuanced subfield of Sentiment Analysis (SA) [1,2,3,4], plays a pivotal role in opinion mining from textual data. Unlike traditional SA, ABSA delves into the sentiments expressed towards specific aspect words or phrases within a text, offering a more detailed sentiment understanding. For instance, in a review stating, "The camera quality is excellent, but the battery life is short", ABSA discerns the distinct positive sentiment towards the ’camera quality’ and negative sentiment towards ’battery life’. This granular analysis breaks free from the constraints of general language structures, tapping into a richer spectrum of sentiment information.
In ABSA, the syntactic structure of sentences, represented by dependency trees, holds critical information. These trees consist of dependency edges and tags that form the backbone of relationships between words, particularly linking aspect words to their corresponding sentiment expressions [4,9,24,25,26]. However, a significant challenge in ABSA is the effective utilization of these syntactic dependencies for sentiment classification. Traditional models often overlook the dual importance of both dependency edges and tags, leading to a disconnect between aspect words and their sentiment contexts.
To address these limitations, we introduce the Advanced Syntactic-Graph Convolutional Model (ASGCM), which innovat4ely employs a dependency graph convolutional network. This model not only utilizes the syntactic dependencies but also enhances the interplay between aspect and sentiment words through its advanced convolutional framework. The ASGCM model is unique in its dual-layer approach: one layer focuses on the structural relationships indicated by dependency edges, while the other layer interprets the semantic roles defined by dependency tags. This dual-layered approach ensures a comprehensive analysis of the text, capturing both the structural and semantic nuances.
The primary contributions of this paper include:
  • Introducing ASGCM, an advanced model that efficiently leverages syntactic dependencies for more accurate sentiment classification in ABSA.
  • Demonstrating the effectiveness of ASGCM through extens4e evaluations on eight d4erse datasets.

2. Related Work

Aspect-Based Sentiment Analysis (ABSA) has evolved significantly over the years, transitioning from traditional machine learning models to more advanced deep learning techniques. Traditional machine learning-based ABSA models relied heavily on the quality of feature engineering [1,2,3,4]. These models often required extensive manual effort to craft features that could effect4ely capture the nuances of sentiment in relation to specific aspects.
The advent of deep learning ushered in a new era for ABSA, with models that could automatically learn feature representations. This shift is exemplified by the work of Dong et al. [5], who introduced a model that adaptively transmitted sentiment information to target aspect words by modeling syntactic relations, thereby improving upon benchmark methods. Tang et al. [6] proposed the Target-Dependent Long Short-Term Memory (TD-LSTM) network, a novel approach that models the context before and after target aspect words, surpassing the performance of standard LSTM models. Ma et al. [7] took this further by introducing an attention network designed for interact4e learning between aspect words and their contexts, enhancing the model’s ability to focus on relevant text segments.
Chen et al. [8] developed the Recurrent Attention Network (RAM), which is built upon the output of bidirectional LSTM. RAM utilized multi-head attention to extract sentiment classification features, demonstrating impress4e performance across various datasets. However, a common limitation among these models, especially those based on attention mechanisms, was their tendency to overlook syntactic constraints, often leading to the consideration of irrelevant contexts when determining sentiment polarity.
Addressing this issue, Zhang et al. [9] introduced the Aspect-Specific Graph Convolutional Network (ASGCN), which efficiently captured sentiment semantic information related to target aspect words by encoding the dependency tree. This approach was significant as it leveraged syntactic structures to enhance sentiment analysis. Similarly, Chen et al. [10] employed a self-attention network to dynamically learn semantic graph information based on the dependency tree, offering a more nuanced understanding of sentence structures.
Further innovations in this domain were made by Huang et al. [11], who proposed the Target-Dependent Graph Attention Network (TD-GAT). TD-GAT excelled in capturing abstract sentiment features contained in syntactic structures. Hou et al. [12] introduced the Sentiment-Aware Graph Convolutional Network (SA-GCN), which achieved high correlation between aspect words and sentiment words through the graph convolutional layer-output features of the dependency tree. CoGAN, proposed by Chen et al. [13], modeled two different types of sentence-level sentiment information to obtain the final sentence representation, showcasing the model’s versatility.
Zhao et al. [14] developed the Syntactic Dependency Graph Convolutional Network (SD-GCN), which used a bidirectional attention network to construct the representation of each aspect word in context. This model was particularly adept at capturing the sentiment features of each aspect word in a sentence through a graph convolutional network. Jia et al. [15] combined syntactic features with multi-semantic fragment features to extract sentiment classification features based on dependency and structural attention, significantly improving the effectiveness of sentiment classification.
Innovations in reconstructing the dependency tree have also been pivotal. Wang et al. [16] built a dependency tree with aspect words as its root, constructing the Recurs4e Graph Attention Network (R-GAT) to further enhance model performance. BiGCN, proposed by Zhang et al. [9], optimized sentiment classification features by performing hierarchical interact4e convolution operations on syntactic and lexical graphs. Li et al. [17]’s DualGCN used two modules to capture syntactic and semantic information of sentences, employing a regularizer to constrain semantic relevance and mitigate errors in dependency analysis. Finally, Hou et al. [18]’s GraphMerge reconstructed different tree structures from various syntactic parsers and applied these to a graph neural network for sentiment classification, showcasing the potential of integrating d4erse syntactic analyses.
These advancements in ABSA models, especially those incorporating syntactic dependencies, have significantly contributed to the field, paving the way for more accurate and nuanced sentiment analysis. Our proposed model builds upon these foundations, introducing novel techniques and methodologies to further enhance the efficiency and accuracy of sentiment classification in ABSA.

3. Methodology

3.1. Overall Architecture

Our proposed Reconfigured Dependency Graph Convolutional Network (RDGCN) architecture incorporates three main components: the Edge-Enhanced Graph Convolutional Network (EEGCN), the Tag-Enriched Graph Convolutional Network (TEGCN), and a Biaffine attention mechanism. Drawing inspiration from Li et al. [17], we employ a dependency parser to create a dependency tree for each sentence. This tree allows us to ascertain dependency edges and tags between nodes, forming the basis for our edge and tag adjacency matrices. The aspect’s contextual word sequence is then input into a BiLSTM layer, which, when combined with the EEGCN and TEGCN modules, yields two distinct types of contextual word features. These features are subsequently synthesized by the Biaffine attention mechanism to predict sentiment polarity, finalized through concatenation and softmax operation.

3.2. Task Modeling

We consider a sentence S = { w 1 , w 2 , … , w n } , with n indicating the total number of words. For an aspect word a within S, we determine its associated sequence S a = { w 1 , w 2 , … , w m } , with m as the count of aspect-defining words. The context sequence S c includes all words except a. For simplicity, we set S c = S , masking the aspect word a. Our model’s goal is to predict the sentiment polarity y of the aspect a, where y ∈ { 0 , 1 , 2 } , denoting negat4e, neutral, and posit4e sentiments, respect4ely.

3.3. Contextual Representation

Word vectors are initialized with GloVe embeddings [42] to form a word vector table T ∈ R 5 × d e , with 5 denoting the vocabulary size and d e the dimensionality of the word vectors. The context word sequence S c is transformed into a feature vector V c = { X 1 , X 2 , … , X n } , where X i = T ( w i ) . These vectors are input into a BiLSTM layer to obtain contextualized word features H = { h 1 , h 2 , … , h L } ∈ R L × 2 d h , where L is the sequence length and d h the hidden dimension of the LSTM. The feature vector h i ∈ R 2 d h is comprised of concatenated forward h → i and backward h ← i LSTM outputs.
h → i = LSTM ( X i , h → i − 1 )
h ← i = LSTM ( X i , h ← i − 1 )
h i = [ h → i ; h ← i ]
We compute context word features H c = { h 1 , h 2 , … , h L c } ∈ R L c × 2 d h as:
H c = BiLSTM ( V c )

3.4. Advanced Syntactic-Graph Convolutional Model (ASGCM)

The dependency tree yields edge information between nodes, which is used to construct the edge adjacency matrix A e . The GCN processes the context word features H c to integrate edge information, producing the edge-informed context word features H e . This is computed as:
h e i l = σ ∑ j = 1 n A e i j l W e l h e j l − 1 + b e l
The tag adjacency matrix A t is similarly der4ed, with the TEGCN generating tag-informed context word features from H c .
T i j = d i output if i ≠ j , A e i j = 1 none if i ≠ j , A e i j = 0 self _ loop if i = j
T i j ′ = d i input if i ≠ j , A t i j = 1 none if i ≠ j , A t i j = 0
h t i l = σ ∑ j = 1 n A t i j l W t l h t j l − 1 + b t l

3.5. Biaffine Decoding

The Biaffine [43] layer fuses the features H e and H t to facilitate sentiment information interaction, yielding the sentiment classification features H e _ p i e and H t _ p i e , defined by:
H e _ p i e = Softmax H e U e ( H t ) T H t
H t _ p i e = Softmax H t U t ( H e ) T H e

3.6. Inference

The final feature set H e n d is achieved by concatenating H e _ p i e and H t _ p i e . The sentiment polarity is then predicted by passing H e n d through a sentiment classifier:
H e n d = [ H e _ p i e ; H t _ p i e ]
p = Softmax ( W p H e n d + b p )

3.7. Training

During training of RDGCN, we incorporate an L2 regularization term into the loss function. The object4e loss function C ( θ ) is defined as:
C ( θ ) = − ∑ k = 1 D y k log ( p k ) + λ 2 θ 2 2
where D is the training dataset, y k the true sentiment label of the kth sample, p k the predicted sentiment probability of the kth sample, θ the set of parameters, and λ the regularization coefficient.

4. Experiment

This section elucidates the experimental outcomes along with a comprehensive analysis.

4.1. Datasets

In our comprehensive evaluation, eight diverse datasets were employed, encompassing five in English—namely Restaurant [19], Laptop [19], Twitter [5], MP3Player [20], and Television [21] and three in Chinese, which include CMPR [22], Camera [23], and Notebook [23]. The initial trio are standard benchmark datasets, while the latter quintet comprises electronic product reviews. Consistent with prior research, instances labeled as “conflict” within the Laptop and Restaurant datasets were omitted. Sentiment polarity distribution statistics for these datasets are detailed in Table 1, noting the absence of neutral samples in MP3Player, Camera, and Notebook datasets.

4.2. Experimental Setup

For sentiment parsing in the Restaurant, Laptop, and Twitter datasets, the Biaffine dependency parser [24] was used alongside GloVe [25]-initialized word embeddings of 300 dimensions. Conversely, for the MP3Player, Television, CMPR, Camera, and Notebook datasets, dependency parsing was conducted via Stanford CoreNLP [26], with word embeddings assigned randomly. Tag embeddings were initialized at random in accordance with our compiled tag dictionary. Adam optimizer was utilized to minimize the loss function. Table 2 encapsulates the detailed parameter settings for our ASGCM model.

4.3. Results

To deepen the exploration of our model’s performance and to offer a more comprehensive analysis, our study meticulously scrutinized the efficacy and generalizability of the Dependency Relation Graph Convolutional Network (ASGCM) against a spectrum of established models. This examination spanned across prevalent datasets such as Restaurant, Laptop, and Twitter—encompassing traditional machine learning approaches like Support Vector Machines (SVM) [4], adapt4e neural networks such as AdaRNN [5], and syntactically-informed neural networks like PhraseRNN [27]. Our analysis extended to a suite of electronic product review datasets, where ASGCM was juxtaposed with models specifically designed for graph-based sentiment analysis, including Aspect-Specific Graph Convolutional Network (ASGCN) [31], Bidirectional Graph Convolutional Network (BiGCN) [9], and Deep Recurrent Sentiment Analysis Network (DRSAN) [15].
In our robust comparat4e framework, we adopted accuracy and Macro-F1 score as the principal metrics, recognizing their significance in capturing the balance between precision and recall, especially in datasets with uneven class distributions. The results, encapsulated in Table 3, Table 4 and Table 5, offer a transparent view of ASGCM’s comparative advantage. These tables highlight ASGCM’s consistent outperformance in sentiment feature extraction from context, which, in turn, substantially enhances sentiment classification efficacy.
Specifically, the data presented in Table 3 corroborate ASGCM’s heightened sentiment classification accuracy, illustrating the model’s proficiency in harnessing syntactic dependency information to amplify context understanding. This proficiency is not marginal; ASGCM transcends the capabilities of DRSAN and other baselines with a significant margin—often exceeding 1 percentage point—which in the domain of sentiment analysis, is both statistically and practically significant.
The subsequent Table 4 and Table 5 provide further evidence of ASGCM’s formidable generalization capabilities across various contexts. It is noteworthy how ASGCM asserts its dominance on the CMPR dataset, eclipsing other advanced models such as ASGCN, BiGCN, and DRSAN. These margins of improvement are not only numerically substantial but also indicative of the model’s adaptability to discern sentiment in complex, aspect-based scenarios. The enhancement in performance across diverse linguistic datasets underscores ASGCM’s utility as a versatile tool in the evolving landscape of sentiment analysis.
Through this indicative experimental analysis, ASGCM has established itself as a vanguard model, setting a new benchmark for sentiment analysis by adeptly capturing the intricate interplay between syntactic structures and sentiment expressions.

5. Conclusion

To address the underutilization of dependency data in current sentiment classification frameworks, this study introduces the Advanced Syntactic-Graph Convolutional Model (ASGCM). This model effectively leverages not only the relational structure within dependency trees but also the granular dependency tag details. Empirical evaluations conducted on standard benchmarks such as Restaurant and Laptop, social media corpora like Twitter, and a variety of electronic product review datasets demonstrate ASGCM’s superior performance in sentiment classification and its enhanced capacity for generalization compared to existing models. Furthermore, the results derived from varying the number of Graph Convolutional Network (GCN) layers indicate that ASGCM’s GCN layers possess a notable flexibility, adapting proficiently to cross-lingual datasets of electronic product reviews.

References

  1. Nazir, A.; Rao, Y.; Wu, L.; Sun, L. Issues and challenges of aspect-based sentiment analysis: A comprehensive survey. IEEE Transactions on Affective Computing 2020, 13, 845–863. [Google Scholar] [CrossRef]
  2. Fei, H.; Wu, S.; Li, J.; Li, B.; Li, F.; Qin, L.; Zhang, M.; Zhang, M.; Chua, T.S. LasUIE: Unifying Information Extraction with Latent Adaptive Structure-aware Generative Language Model. Proceedings of the Advances in Neural Information Processing Systems, NeurIPS 2022, 2022, 15460–15475. [Google Scholar]
  3. Kiritchenko, S.; Zhu, X.; Cherry, C.; Mohammad, S. Detecting aspects and sentiment in customer reviews. 8th International Workshop on Semantic Evaluation (SemEval), pp. 437–442.
  4. Dong, L.; Wei, F.; Tan, C.; Tang, D.; Zhou, M.; Xu, K. Adaptive recursive neural network for target-dependent twitter sentiment classification. Proceedings of the 52nd annual meeting of the association for computational linguistics (volume 2: Short papers), 2014, pp. 49–54.
  5. Fei, H.; Li, J.; Wu, S.; Li, C.; Ji, D.; Li, F. Global Inference with Explicit Syntactic and Discourse Structures for Dialogue-Level Relation Extraction. Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI, 2022, pp. 4082–4088.
  6. Wu, S.; Fei, H.; Ren, Y.; Ji, D.; Li, J. Learn from Syntax: Improving Pair-wise Aspect and Opinion Terms Extraction with Rich Syntactic Knowledge. Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, 2021, pp. 3957–3963.
  7. Xiang, C.; Zhang, J.; Li, F.; Fei, H.; Ji, D. A semantic and syntactic enhanced neural model for financial sentiment analysis. Information Processing & Management 2022, 59, 102943. [Google Scholar] [CrossRef]
  8. Fei, H.; Zhang, Y.; Ren, Y.; Ji, D. Latent Emotion Memory for Multi-Label Emotion Classification. Proceedings of the AAAI Conference on Artificial Intelligence, 2020, pp. 7692–7699.
  9. Tang, D.; Qin, B.; Feng, X.; Liu, T. Effective LSTMs for target-dependent sentiment classification. arXiv preprint 2015, arXiv:1512.01100. [Google Scholar] [CrossRef]
  10. Wu, S.; Fei, H.; Li, F.; Zhang, M.; Liu, Y.; Teng, C.; Ji, D. Mastering the Explicit Opinion-Role Interaction: Syntax-Aided Neural Transition System for Unified Opinion Role Labeling. Proceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence, 2022, pp. 11513–11521.
  11. Ma, D.; Li, S.; Zhang, X.; Wang, H. Interactive attention networks for aspect-level sentiment classification. arXiv preprint 2017, arXiv:1709.00893. [Google Scholar] [CrossRef]
  12. Fei, H.; Zhang, M.; Ji, D. Cross-Lingual Semantic Role Labeling with High-Quality Translated Training Corpus. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 2020, pp. 7014–7026.
  13. Chen, P.; Sun, Z.; Bing, L.; Yang, W. Recurrent attention network on memory for aspect sentiment analysis. Proceedings of the 2017 conference on empirical methods in natural language processing, 2017, pp. 452–461.
  14. Fei, H.; Zhang, M.; Li, B.; Ji, D. End-to-end Semantic Role Labeling with Neural Transition-based Model. Proceedings of the AAAI Conference on Artificial Intelligence, 2021, pp. 12803–12811.
  15. Zhang, M.; Qian, T. Convolution over hierarchical syntactic and lexical graphs for aspect level sentiment analysis. Proceedings of the 2020 conference on empirical methods in natural language processing (EMNLP), 2020, pp. 3540–3549.
  16. Wu, S.; Fei, H.; Ren, Y.; Li, B.; Li, F.; Ji, D. High-Order Pair-Wise Aspect and Opinion Terms Extraction With Edge-Enhanced Syntactic Graph Convolution. IEEE ACM Trans. Audio Speech Lang. Process. 2021, 29, 2396–2406. [Google Scholar] [CrossRef]
  17. Chen, C.; Teng, Z.; Zhang, Y. Inducing target-specific latent structures for aspect sentiment classification. Proceedings of the 2020 conference on empirical methods in natural language processing (EMNLP), 2020, pp. 5596–5607.
  18. Fei, H.; Wu, S.; Ren, Y.; Zhang, M. Matching Structure for Dual Learning. Proceedings of the International Conference on Machine Learning, ICML, 2022, pp. 6373–6391.
  19. Huang, L.; Sun, X.; Li, S.; Zhang, L.; Wang, H. Syntax-aware graph attention network for aspect-level sentiment classification. Proceedings of the 28th international conference on computational linguistics, 2020, pp. 799–810.
  20. Fei, H.; Li, F.; Li, B.; Ji, D. Encoder-Decoder Based Unified Semantic Role Labeling with Label-Aware Syntax. Proceedings of the AAAI Conference on Artificial Intelligence, 2021, pp. 12794–12802.
  21. Fei, H.; Ren, Y.; Zhang, Y.; Ji, D. Nonautoregressive Encoder-Decoder Neural Framework for End-to-End Aspect-Based Sentiment Triplet Extraction. IEEE Transactions on Neural Networks and Learning Systems 2023, 34, 5544–5556. [Google Scholar] [CrossRef] [PubMed]
  22. Hou, X.; Huang, J.; Wang, G.; He, X.; Zhou, B. Selective attention based graph convolutional networks for aspect-level sentiment classification. arXiv preprint 2019, arXiv:1910.10857. [Google Scholar] [CrossRef]
  23. Fei, H.; Li, J.; Ren, Y.; Zhang, M.; Ji, D. Making Decision like Human: Joint Aspect Category Sentiment Analysis and Rating Prediction with Fine-to-Coarse Reasoning. Proceedings of the ACM Web Conference 2022, WWW, 2022, pp. 3042–3051.
  24. Li, J.; Fei, H.; Liu, J.; Wu, S.; Zhang, M.; Teng, C.; Ji, D.; Li, F. Unified Named Entity Recognition as Word-Word Relation Classification. Proceedings of the AAAI Conference on Artificial Intelligence, 2022, pp. 10965–10973.
  25. Fei, H.; Ren, Y.; Zhang, Y.; Ji, D.; Liang, X. Enriching contextualized language model from knowledge graph for biomedical information extraction. Briefings in Bioinformatics 2021, 22. [Google Scholar] [CrossRef] [PubMed]
  26. Fei, H.; Ren, Y.; Ji, D. Boundaries and edges rethinking: An end-to-end neural model for overlapping entity relation extraction. Information Processing & Management 2020, 57, 102311. [Google Scholar] [CrossRef]
  27. Fei, H.; Ji, D.; Li, B.; Liu, Y.; Ren, Y.; Li, F. Rethinking Boundaries: End-To-End Recognition of Discontinuous Mentions with Pointer Networks. Proceedings of the AAAI Conference on Artificial Intelligence, 2021, pp. 12785–12793.
  28. Mukherjee, R.; Shetty, S.; Chattopadhyay, S.; Maji, S.; Datta, S.; Goyal, P. Reproducibility, replicability and beyond: Assessing production readiness of aspect based sentiment analysis in the wild. Advances in Information Retrieval: 43rd European Conference on IR Research, ECIR 2021, Virtual Event, March 28–April 1, 2021, Proceedings, Part II 43. Springer, 2021, pp. 92–106.
  29. Fei, H.; Chua, T.; Li, C.; Ji, D.; Zhang, M.; Ren, Y. On the Robustness of Aspect-based Sentiment Analysis: Rethinking Model, Data, and Training. ACM Transactions on Information Systems 2023, 41, 50:1–50:32. [Google Scholar] [CrossRef]
  30. Zhuang, L.; Fei, H.; Hu, P. Knowledge-enhanced event relation extraction via event ontology prompt. Inf. Fusion 2023, 100, 101919. [Google Scholar] [CrossRef]
  31. Zhang, C.; Li, Q.; Song, D. Aspect-based sentiment classification with aspect-specific graph convolutional networks. arXiv preprint 2019, arXiv:1909.03477. [Google Scholar] [CrossRef]
  32. Fei, H.; Wu, S.; Ren, Y.; Li, F.; Ji, D. Better Combine Them Together! Integrating Syntactic Constituency and Dependency Representations for Semantic Role Labeling. Findings of the Association for Computational Linguistics: ACL/IJCNLP 2021, 2021, 549–559. [Google Scholar] [CrossRef]
  33. Chen, X.; Sun, C.; Wang, J.; Li, S.; Si, L.; Zhang, M.; Zhou, G. Aspect sentiment classification with document-level sentiment preference modeling. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 2020, pp. 3667–3677.
  34. Li, J.; Xu, K.; Li, F.; Fei, H.; Ren, Y.; Ji, D. MRN: A Locally and Globally Mention-Based Reasoning Network for Document-Level Relation Extraction. Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, 2021, 1359–1370. [Google Scholar]
  35. Liu, J.; Fei, H.; Li, F.; Li, J.; Li, B.; Zhao, L.; Teng, C.; Ji, D. TKDP: Threefold Knowledge-enriched Deep Prompt Tuning for Few-shot Named Entity Recognition. CoRR 2023, abs/2306.03974. [Google Scholar] [CrossRef]
  36. Pontiki, M.; Galanis, D.; Papageorgiou, H.; Androutsopoulos, I.; Manandhar, S.; AL-Smadi, M.; Al-Ayyoub, M.; Zhao, Y.; Qin, B.; De Clercq, O.; others. Semeval-2016 task 5: Aspect based sentiment analysis. ProWorkshop on Semantic Evaluation (SemEval-2016). Association for Computational Linguistics, 2016, pp. 19–30.
  37. Fei, H.; Li, F.; Li, C.; Wu, S.; Li, J.; Ji, D. Inheriting the Wisdom of Predecessors: A Multiplex Cascade Framework for Unified Aspect-based Sentiment Analysis. Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI, 2022, pp. 4096–4103.
  38. Wang, F.; Li, F.; Fei, H.; Li, J.; Wu, S.; Su, F.; Shi, W.; Ji, D.; Cai, B. Entity-centered Cross-document Relation Extraction. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, 2022, pp. 9871–9881.
  39. Fei, H.; Ren, Y.; Ji, D. Retrofitting Structure-aware Transformer Language Model for End Tasks. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, 2020, pp. 2151–2161.
  40. Jia, Y.; Wang, Y.; Zan, H.; Xie, Q. Syntactic information and multiple semantic segments for aspect-based sentiment classification. International Journal of Asian Language Processing 2021, 31, 2250006. [Google Scholar] [CrossRef]
  41. Cao, H.; Li, J.; Su, F.; Li, F.; Fei, H.; Wu, S.; Li, B.; Zhao, L.; Ji, D. OneEE: A One-Stage Framework for Fast Overlapping and Nested Event Extraction. Proceedings of the 29th International Conference on Computational Linguistics, 2022, pp. 1953–1964.
  42. Pennington, J.; Socher, R.; Manning, C.D. Glove: Global vectors for word representation. Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), 2014, pp. 1532–1543.
  43. Dozat, T.; Manning, C.D. Deep biaffine attention for neural dependency parsing. arXiv preprint 2016, arXiv:1611.01734. [Google Scholar] [CrossRef]
Table 1. SENTIMENT POLARITY DISTRIBUTION STATISTICS
Table 1. SENTIMENT POLARITY DISTRIBUTION STATISTICS
Dataset/Lang. Posit4e Negat4e Neutral
training test training test training test
Restaurant/EN 2164 728 805 196 633 196
Laptop/EN 987 341 866 128 460 169
Twitter/EN 1561 173 1560 173 3127 346
MP3Player/EN 305 108 204 58 0 0
Television/EN 2540 618 919 257 287 67
CMPR/ZH 1624 571 497 190 117 35
Camera/ZH 1117 406 482 174 0 0
Notebook/ZH 305 111 160 44 0 0
Table 2. PARAMETER SETTINGS
Table 2. PARAMETER SETTINGS
Name Value
Tag embedding dimension d tag 300
Word embedding dimension d e 300
LSTM hidden layer dimension d h 300
GCN hidden layer dimension d g 200
GCN layer l 2
Initializing weights U(-0.01,0.01)
Initializing bias 0
Regularization coefficient λ 10 − 4
Learning rate 10 − 3
Dropout 0.2
Table 3. EXPERIMENT RESULTS ON RESTAURANT, LAPTOP AND TWITTER
Table 3. EXPERIMENT RESULTS ON RESTAURANT, LAPTOP AND TWITTER
Model Restaurant (%) Laptop (%) Twitter (%)
Accuracy Macro-F1 Accuracy Macro-F1 Accuracy Macro-F1
SVM 80.16 - 70.49 - 63.40 63.30
ASGCM 83.10 73.58 77.01 73.74 75.68 74.03
Table 4. EXPERIMENT RESULTS ON MP3PLAYER AND TELEVISION
Table 4. EXPERIMENT RESULTS ON MP3PLAYER AND TELEVISION
Model MP3Player (%) Television (%)
Accuracy Macro-F1 Accuracy Macro-F1
ASGCN 72.28 71.35 81.74 63.07
BiGCN 72.79 71.21 83.19 65.42
DRSAN 73.49 72.92 86.30 66.72
ASGCM 74.31 73.72 87.18 66.58
Table 5. EXPERIMENT RESULTS ON CMPR, CAMERA AND NOTEBOOK
Table 5. EXPERIMENT RESULTS ON CMPR, CAMERA AND NOTEBOOK
Model CMPR (%) Camera (%) Notebook (%)
Accuracy Macro-F1 Accuracy Macro-F1 Accuracy Macro-F1
ASGCN 86.11 62.59 81.55 76.68 75.74 72.65
BiGCN 87.42 65.13 83.25 78.47 77.62 75.17
DRSAN 88.07 68.02 85.68 81.53 79.35 76.12
ASGCM 89.81 69.45 87.78 83.46 80.95 77.84
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.
Prerpints.org logo

Preprints.org is a free preprint server supported by MDPI in Basel, Switzerland.

facebook logotwitter logolinkedin logo
MDPI Initiatives
Important Links
Subscribe

Disclaimer

Terms of Use

Privacy Policy

© 2024 MDPI (Basel, Switzerland) unless otherwise stated