Aspect-based Sentiment Analysis
Although a rating can summarize a whole review, it is really the vast amount of finer details matters a lot. After all, each person's need is quite different and we wish a personalized fit of a product (or service) to our own needs.
Aspect-based sentiment analysis (ABSA) aims to find fine-grained opinions from reviews. For example, in a Laptop domain, we may wish to see whether the screen, keyboard, etc. are good or not. Although there are unlimited amount of reviews with coarse-grained ratings, ABSA severely lacks supervision from humans (e.g., in the form of annotated data). The above ABSA problem can be decomposed into two important sub-tasks: aspect extraction (AE) and aspect sentiment classification (ASC).
Aspect Extraction
Given a review sentence, such as The retina display is beautiful., AE aims to find aspects retina display. In DL, it is mostly formalized as a sequence labeling problem: label "The retina display is great ." as "O B I O O O" so to extract "retina display" as an aspect. Obviously, the context of an aspect is important and an AI agent needs to have enough domain knowledge to support such extraction, such as A beautiful thing in Laptop could be an aspect. However, counting on the strong supervision from humans aspects-by-aspects for a particular domain is impossible. We show that a simple domain word embedding can boost the performance. More ideally, a language model can boost it further as whenever you see some aspects dropped, LM really encouraging the context words to recover that aspects. This is the first paper to use BERT for AE and ASC.
Aspect Sentiment Classification
Given an aspect retina display and a review sentence The retina display is great., ASC detects the polarity of that aspect positive. One challenge of ASC is to detect the polarity of opinion expressions and there could be unlimited amount of such expressions to annotate. Again language model can help this too as human tends to repeat their opinions in their writing so knowing one opinion may help to automatically understand the other. For example, in Terrible product. It could be the last thing I may consider to buy, we may infer the harder opinion of the second sentence from Terrible in the first one so to automatically learn unlimited expressions like that.