Table-Filling via Mean Teacher for Cross-domain Aspect Sentiment Triplet Extraction

TL;DR

This work addresses cross-domain ASTE by reframing table-filling as a two-stage, region-based extraction task akin to object detection. It introduces Table-Filling via Mean Teacher (TFMT), a mean-teacher framework that uses region-level consistency and an MMD-based cross-domain constraint to align source supervision with unlabeled target data without generating synthetic labels. The method achieves state-of-the-art performance with substantially fewer parameters and lower training costs, validating the benefits of the OD-inspired perspective for cross-domain transfer in fine-grained sentiment tasks. Overall, TFMT provides a concise, effective baseline and points to fruitful directions for applying mean-teacher and region-level alignment in ASTE and related tasks.

Abstract

Cross-domain Aspect Sentiment Triplet Extraction (ASTE) aims to extract fine-grained sentiment elements from target domain sentences by leveraging the knowledge acquired from the source domain. Due to the absence of labeled data in the target domain, recent studies tend to rely on pre-trained language models to generate large amounts of synthetic data for training purposes. However, these approaches entail additional computational costs associated with the generation process. Different from them, we discover a striking resemblance between table-filling methods in ASTE and two-stage Object Detection (OD) in computer vision, which inspires us to revisit the cross-domain ASTE task and approach it from an OD standpoint. This allows the model to benefit from the OD extraction paradigm and region-level alignment. Building upon this premise, we propose a novel method named \textbf{T}able-\textbf{F}illing via \textbf{M}ean \textbf{T}eacher (TFMT). Specifically, the table-filling methods encode the sentence into a 2D table to detect word relations, while TFMT treats the table as a feature map and utilizes a region consistency to enhance the quality of those generated pseudo labels. Additionally, considering the existence of the domain gap, a cross-domain consistency based on Maximum Mean Discrepancy is designed to alleviate domain shift problems. Our method achieves state-of-the-art performance with minimal parameters and computational costs, making it a strong baseline for cross-domain ASTE.

Figures (6)

• Figure 1: The ASTE model trained in the restaurant domain fails to produce accurate results in the laptop domain.
• Figure 2: The general architecture of two-stage Object Detection (OD) and region-level table-filling method for Aspect Sentiment Triplet Extraction (ASTE) share similar underlying principles. $B$ (Beginning) and $E$ (Ending) represent the top-left and bottom-right corners of a candidate region, respectively.
• Figure 3: Examples of (a) cell-level and (b) region-level methods. As the latter closely resembles two-stage object detection, our TFMT is based on the region-level paradigm.
• Figure 4: The architecture of TFMT. Our model consists of two identical architectures: 1) the teacher model, whose parameters are first pre-trained on the source domain and later frozen in each training step; 2) the student model, whose parameters are reinitialized. The student model is trained using source domain data and target domain pseudo-labels provided by the teacher model, and the teacher model parameters are updated using EMA after the training step.
• Figure 5: Error analysis of pseudo labels.