Fariska Zakhralativa Ruskanda
Aspect term extraction is one of the main subtasks in aspect-based sentiment analysis. In aspect extraction task, we aim to extract the aspect/opinion target/feature/topic from the review text. One of the examples of aspect extraction method is the Sequential Covering algorithm by Ruskanda et al. In the said method, a list of aspect and opinion words from the annotated review datasets are used to aid the process of extracting aspect and opinion expressions. By calculating the similarity between each extracted word with words in the list and only taking words which similarity value is above a certain bound, the method managed to improve its precision. However, if the aspect and opinion of each review are not yet annotated, the word list would have to be crafted manually, costing a significant amount of time and effort. To tackle this downside, we proposed a way to automatically build aspect and opinion list by using word embedding. In our work, we expanded the scope of the word list to create domain-specific lexicon which contains aspect and opinions that are relevant to a domain. We chose restaurant, handphone, and digital camera domain based on the review datasets that would be used to evaluate the performance of aspect extraction. We have adapted the works of Grefenstette and Muchemi to automatically build domainspecific lexicons. The resulting domain-specific lexicons were used in the modified version of Sequential Covering algorithm for aspect extraction. Even though the recalls obtained from the modified method are still lower than the recalls obtained from baseline Sequential Covering, this method managed to yield better F1 scores than the baseline method.
Penerapan Karya Tulis
The aspect and opinion of each review are not yet annotated, the word list would have to be crafted manually, costing a significant amount of time and effort.