Abstract
Aspect-based sentiment analysis produces a list of aspect terms and their corresponding sentiments for a natural language sentence. This task is usually done in a pipeline manner, with aspect term extraction performed first, followed by sentiment predictions toward the extracted aspect terms. While easier to develop, such an approach does not fully exploit joint information from the two subtasks and does not use all available sources of training information that might be helpful, such as document-level labeled sentiment corpus. In this paper, we propose an interactive multi-task learning network (IMN) which is able to jointly learn multiple related tasks simultaneously at both the token level as well as the document level. Unlike conventional multi-task learning methods that rely on learning common features for the different tasks, IMN introduces a message passing architecture where information is iteratively passed to different tasks through a shared set of latent variables. Experimental results demonstrate superior performance of the proposed method against multiple baselines on three benchmark datasets.
Introduction
Aspect-Based Sentiment Analysis (ABSA) represents a fine-grained approach to sentiment analysis, identifying specific aspect terms (e.g., "food") and their associated sentiments (e.g., "great"). Traditionally, this has been approached as a two-step pipeline: Aspect Extraction (AE) followed by Aspect Sentiment classification (AS). However, this sequential approach suffers from error propagation and fails to capture the inherent correlations between identifying an aspect and determining its sentiment.
To address these limitations, researchers from the National University of Singapore and SAP proposed the Interactive Multi-Task Learning Network (IMN). This framework moves beyond rigid pipelines, allowing multiple tasks to interact and refine each other's predictions through an iterative message-passing mechanism.
The Interactive Multi-Task Learning Network (IMN)
The IMN addresses three critical challenges in ABSA: error propagation in pipelines, the lack of explicit interaction between subtasks, and the scarcity of labeled aspect-level data. It achieves this through a unified architecture that jointly learns token-level tasks (AE and AS) and document-level tasks (Document Sentiment, DS; and Domain Classification, DD).
Message Passing Architecture
At the heart of the IMN is a message-passing mechanism that allows information to flow iteratively between tasks. Instead of a single forward pass, the model updates shared latent representations over $T$ iterations (with $T=2$ found to be optimal). This allows the model to refine its understanding; for instance, identifying a strong opinion word can help the model "re-read" the sentence to find the associated aspect term.
Opinion Transmission
A key feature of the IMN is the explicit transmission of information from the AE task to the AS task. By calculating the probability that a token is an opinion word, the model can weight the self-attention layer in the sentiment classifier, focusing its "attention" on the most relevant descriptive words (like "delicious" or "slow").
Experimental Setup
The IMN was evaluated on three standard benchmark datasets: Restaurant14 (D1), Laptop14 (D2), and Restaurant15 (D3). To mitigate data scarcity, the researchers incorporated large-scale document-level data from Yelp and Amazon (30,000 instances each) to train the DS and DD components, which act as powerful regularizers for the fine-grained tasks.
Results and Analysis
The primary metric for evaluation was the Integrated F1 (F1-I), which requires both the aspect span and the sentiment to be correctly identified for a prediction to be marked as accurate.
As shown in the results table, the IMN consistently outperformed state-of-the-art baselines like DECNN and INABSA. On the D1 dataset, IMN achieved an F1-I of 69.54%, a significant improvement over the best baseline's 67.25%. Ablation studies confirmed that both the message-passing mechanism and the inclusion of document-level tasks were critical to this success. Furthermore, qualitative analysis showed that the IMN is particularly effective at identifying uncommon aspect terms by leveraging nearby sentiment cues.
Conclusion
The Interactive Multi-Task Learning Network represents a significant step forward for end-to-end ABSA. By treating aspect extraction and sentiment classification as mutually reinforcing tasks rather than independent steps, the IMN achieves superior accuracy and robustness. Future work may explore the application of this architecture to other complex multi-task problems, such as relation extraction or semantic role labeling.
Summary updated