Dataset Construction and Evaluation for Aspect-Opinion Extraction in Bangla Fine-Grained Sentiment Analysis

In the present age, increasingly more people and entities post their opinions on web-based platforms. With this increasing trend of web-based opinion content, it is quite impossible to digest it by hand. Consequently, a computerized system for opinion and sentiment analysis is required. Sentiment analysis builds systems that automatically extract useful information from massive online reviews. Conventional sentiment analysis simply classifies a document/text into positive/negative/neutral polarity on the basis of overall sentiment polarity. In this case, it is considered that one sentiment is carried for the whole text, which may not be the case in the practical scenario. It is possible to appear more than one aspect, opinion, and sentiment in a single text, especially when the text is long. This motivates the introduction of fine-grained sentiment analysis, in which one can identify what a person/entity is mentioning and how they feel about each aspect/target entity. Aspect-opinion term extraction is a crucial aspect-based sentiment analysis (ABSA). It is a kind of granular-level sentiment analysis. In this study, our target is aspect-opinion terms extraction in the Bangla language. However, no annotated Bangla dataset is available for this task. Hence, we obtained textual data from a publicly available Bangla restaurant dataset, annotated them, and prepared a novel dataset. We then performed the task as a sequence labeling technique by utilizing our dataset. For the experiment, we employed vanilla transformers-based models such as mBERT, BanglaBERT, and BanglishBERT with linear layers and conditional random field (CRF) on top of them. To enhance the models’ performance, we combined different feature embeddings of mBERT, BanglaBERT, and BanglishBERT. We then utilized the linear layers and CRF on top of them. The experimental results indicated that the combined feature embeddings technique significantly improves models’ performance, and (mBERT & BanglishBERT) with linear layer performed best: 0.5840 on the F1 score among all the tested models.