Cross-Domain Review Generation for Aspect-Based Sentiment Analysis

Jianfei Yu, Chenggong Gong, Rui XiaACL 2021

TL;DR

This paper introduces Cross-Domain Review Generation (CDRG), a novel approach to tackle the challenge of limited labeled data in Aspect-Based Sentiment Analysis (ABSA) across different domains. CDRG generates synthetic target-domain reviews with fine-grained annotations from existing source-domain data, enabling effective cross-domain adaptation for ABSA tasks.

What Problem Does This Paper Address?

Aspect-Based Sentiment Analysis (ABSA) aims to identify specific aspects within text (e.g., "battery life" in a laptop review) and determine the sentiment expressed towards them (e.g., "positive" or "negative"). While supervised learning models have achieved great success in ABSA, their performance heavily relies on large amounts of labeled data. Acquiring such fine-grained annotations for every new domain (e.g., restaurant reviews, product reviews) is prohibitively expensive and time-consuming. This data sparsity issue is a major bottleneck for applying ABSA in real-world scenarios. To overcome this, unsupervised domain adaptation methods are crucial, aiming to transfer knowledge from a well-annotated source domain to a target domain with only unlabeled data. Existing domain adaptation methods for ABSA often struggle with interpretability and rely solely on source domain supervision, leading to suboptimal performance.

Methods

The paper proposes a two-step approach for Cross-Domain Review Generation (CDRG):

  1. Domain Generalization: This step aims to create a "domain-independent" review. It first identifies domain-specific attributes (aspect and opinion terms) in both source and target domains. Then, these source-specific attributes in the labeled source reviews are masked (replaced with a [MASK] token). This process effectively removes domain-specific jargon while preserving the original sentence structure and sentiment annotation.
  2. Domain Specification: This step converts the domain-independent review into a target-domain review. It utilizes a pre-trained BERT model that is further trained on the unlabeled target domain data using masked language modeling (MLM). This "Target-Domain MLM" (TD-MLM) is then used to predict words for the masked tokens. To improve accuracy and coherence, the authors introduce two constraints:
  3. Target-Specific Aspect Constraint: Limits predictions to known target-domain aspect terms.
  4. Target-Specific Opinion Constraint: Ensures predicted opinion terms align with the sentiment of the original source-domain opinion term.

The generated target-domain reviews, now with fine-grained annotations, can be used to train ABSA models for the target domain. The paper also proposes two training strategies: "Independent Training" (using only generated reviews) and "Merge Training" (combining generated reviews with labeled source reviews).

Datasets

The experiments were conducted on four benchmark datasets:

  • Laptop (L): From SemEval-2014 ABSA challenge.
  • Restaurant (R): A combination of datasets from SemEval ABSA challenges 2014, 2015, and 2016.
  • Device (D): Collected by Toprak et al. (2010).
  • Service (S): From web services, introduced by Hu and Liu (2004).

The paper uses 10 transfer pairs between these domains, excluding D→L and L→D due to their similarity.

Key Insights and Ideas

  • Generating Annotated Data is Key: The core idea that generating synthetic, annotated target-domain data can be more effective than traditional domain adaptation techniques that rely on unlabeled data or domain-invariant features.
  • Domain-Independent Review as a Bridge: The intermediate "domain-independent" review acts as a crucial bridge, allowing for word-to-word alignment and direct transfer of annotations from source to target.
  • Importance of Constraints: The TD-MLM alone is not enough; incorporating target-specific aspect and opinion constraints significantly improves the quality and relevance of the generated target reviews.
  • Interpretability: Unlike methods that learn abstract domain-invariant features, CDRG offers a more interpretable approach by explicitly transforming domain-specific terms.

Key Results

  • State-of-the-Art Performance: The CDRG-based approaches, particularly with "Merge Training," significantly outperform existing state-of-the-art domain adaptation methods for both cross-domain End-to-End ABSA (E2E-ABSA) and Aspect Extraction (AE).
  • Significant Improvements: For E2E-ABSA, improvements of up to 2.83% (Micro-F1) were achieved over the best UDA methods. For AE, improvements reached up to 4.47%.
  • BERT Baseline Boost: Even a baseline BERT model, when fine-tuned solely on the generated target-domain reviews, outperformed previous domain adaptation approaches.
  • Manual Evaluation: Manual evaluation showed that the TD-MLM-C approach (with constraints) generated better target reviews in a majority of cases compared to its ablations.

Limitations and Future Work

  • Quality of Generated Reviews: While improved by constraints, the generated reviews can still contain incorrect annotations or non-aspect/non-opinion terms, as noted in the manual evaluation examples.
  • Dependency on Source Data: The effectiveness of CDRG is still dependent on the availability of sufficient labeled data in the source domain.
  • Future Work: The paper suggests exploring more sophisticated methods for domain-specific attribute extraction and generation, as well as extending CDRG to other NLP tasks beyond ABSA. Further research could also focus on improving the robustness of the generated reviews against noisy source data.

Summary updated

BibTeX
@inproceedings{yu2021cross,
  author = {Jianfei Yu and Chenggong Gong and Rui Xia},
  title = {{Cross-Domain Review Generation for Aspect-Based Sentiment Analysis}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
}

Loading related content...

Figure