A Simple yet Effective Framework for Few-Shot Aspect-Based Sentiment Analysis

What is it about?

Aspect-Based Sentiment Analysis (ABSA) has attracted a lot of interest due to its wide application, such as recommendation systems and product-related question answering. Given a review sentence, ABSA seeks to mine the fine-grained opinion targets in terms of entity or aspect (collectively called aspect) in the review and determine the sentiment polarity toward each aspect. For instance, for a customer review in a Restaurant domain “The pizza here is delicious.”, the expected output of End-to-End ABSA (E2E-ABSA) is (“pizza”, Positive), where “pizza” is the aspect and Positive is the sentiment polarity toward “pizza”.

Featured Image

Photo by Buzz Andersen on Unsplash

Photo by Buzz Andersen on Unsplash

Why is it important?

Pre-training and Fine-tuning has been the dominant paradigm in the field of ABSA in recent years. Although this paradigm has achieved state-of-the-art ABSA performance, it relies on a large amount of fine-grained “aspect-sentiment” annotations in the reviews. In the domains that lack sufficient labeled data, this paradigm is still challenging to achieve satisfactory performance. In the real world, there are so many domains of products, and manually labeling for each domain is almost impossible. Hence, few-shot ABSA, the task conducting ABSA with only a small number of labeled data, has become an important direction for ABSA that has not yet achieved effective performance. To address this issue, we introduce a simple yet effective framework called FS-ABSA, which involves domain-adaptive pre-training and text-infilling fine-tuning. We approach the End-to End ABSA task as a text-infilling problem and perform domainadaptive pre-training with the text-infilling objective, narrowing the domain and objective gaps between pre-training and fine-tuning, and consequently facilitating the knowledge transfer. Experiments show that the resulting model achieves more compelling performance than baselines under the few-shot setting while driving the state-of-the-art performance to a new level across datasets under the fully-supervised setting. Moreover, we apply our framework to two non-English low-resource languages to demonstrate its generality and effectiveness.