Improving Aspect Sentiment Quad Prediction via Template-Order Data Augmentation

Hu, Mengting; Wu, Yike; Gao, Hang; Bai, Yang; Zhao, Shiwan

doi:10.18653/v1/2022.emnlp-main.538

Cited by 10 publications

(3 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We compare the proposed method with several classification-based aspect-based sentiment analysis models, including, DP (Qiu et al, 2011), JET , TAS-BERT (Wan et al, 2020) and Extract-Classify (Cai et al, 2021). In addition, generative models are also compared, such as BARTABSA (Yan et al, 2021), GAS (Zhang et al, 2021b), Paraphrase (Zhang et al, 2021a),TODA (Hu et al, 2022), Seq2Path (Mao et al, 2022) and OTG (Bao et al, 2022). 4 .…”

Section: Resultsmentioning

confidence: 99%

Exploring Graph Pre-training for Aspect-based Sentiment Analysis

Bao,

Wang,

Zhou

2023

Findings of the Association for Computational Linguistics: EMNLP 2023

View full text Add to dashboard Cite

Existing studies tend to extract the sentiment elements in a generative manner in order to avoid complex modeling. Despite their effectiveness, they ignore importance of the relationships between sentiment elements that could be crucial, making the large pre-trained generative models sub-optimal for modeling sentiment knowledge. Therefore, we introduce two pre-training paradigms to improve the generation model by exploring graph pre-training that targeting to strengthen the model in capturing the elements' relationships. Specifically, We first employ an Element-level Graph Pretraining paradigm, which is designed to improve the structure awareness of the generative model. Then, we design a Task-level Graph Pre-training paradigm to make the generative model generalizable and robust against various irregular sentiment quadruples. Extensive experiments show the superiority of our proposed method, and validate the correctness of our motivation. Our code can be found in https://github.com/HoraceXIaoyiBao/ EGP4ABSA-EMNLP2023.

show abstract

Section: Resultsmentioning

confidence: 99%

Exploring Graph Pre-training for Aspect-based Sentiment Analysis

Bao,

Wang,

Zhou

2023

Findings of the Association for Computational Linguistics: EMNLP 2023

View full text Add to dashboard Cite

show abstract

“…Paraphrase [30]: Another generative approach that paraphrases the original sentence to better utilize the semantic knowledge in pre-trained language models. DLO [31]: Building on [30], it was found that combining multiple templates could improve ASQP tasks through data augmentation.…”

Section: Generative Methodsmentioning

confidence: 99%

“…They transformed the quadruple prediction task into a text generation problem by combining annotated sentiment elements with pre-established templates and using the resulting natural language sentences as target sequences, addressing it through a Seq2Seq modeling paradigm. Following this, Hu et al [31] built on Zhang et al [30], noting the influence of the sentiment elements' order in templates on quadruple extraction performance, proposing the combination of multiple templates to improve the ASQP task through data augmentation. Differing from previous studies, we propose a tree structure information-aware prompt tuning method, aiming to jointly detect all sentiment elements in a given review sentence's tree.…”

Section: Related Workmentioning

confidence: 99%

Opinion-Tree-aware Prompt Tuning for Aspect Sentiment Quadruple Prediction

Yang,

Li,

Liu

et al. 2024

Preprint

View full text Add to dashboard Cite

In recent years, Aspect Sentiment Quadruple Prediction (ASQP) has emerged as a prominent task in the field of aspect-based sentiment analysis. We observe complex dependencies among aspect sentiment elements, rendering traditional sequence-to-sequence (Seq2Seq) modeling for element extraction insufficiently rigorous. Additionally, despite the widespread application of Pre-trained Language Models (PLMs) in Aspect-Based Sentiment Analysis (ABSA), there is a significant gap when representing dependencies between sentiment elements, which involve complex structure-informed label classification, diverging from PLM's Masked Language Model (MLM) pre-training tasks. This gap limits the full potential of PLMs. To bridge this gap and better adapt to the structural information between sentiment elements, we propose a Opinion-Tree-aware Prompt Tuning (OTPT) method. This approach transforms traditional PLM Seq2Seq tasks into sequence-to-tree tasks, optimizing PLM potential. Specifically, we model sentiment elements as a tree structure, creating a dynamic virtual template and label words, integrating tree structure information as soft prompts, and conducting stepwise sentiment label classification. Extensive experiments show that our method achieves performance superior to baseline on two widely used datasets, demonstrating good robustness in handling cases of missing sentiment elements and complex multi-sentiment quadruples.

show abstract