ASPECT AND SENTIMENT EXPRESSIONS RELATION EXTRACTION FOR INDONESIAN HOTEL REVIEWS
Customers’ opinion about a product can be concluded from their review by using aspect-based sentiment analysis. It is useful for product provider and future customer as consideration to make decision. One of the tasks involved in aspect-based sentiment analysis is aspect and sentiment expressions re...
Saved in:
| Main Author: | |
|---|---|
| Format: | Final Project |
| Language: | Indonesia |
| Online Access: | https://digilib.itb.ac.id/gdl/view/43557 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Institut Teknologi Bandung |
| Language: | Indonesia |
| Summary: | Customers’ opinion about a product can be concluded from their review by using aspect-based sentiment analysis. It is useful for product provider and future customer as consideration to make decision. One of the tasks involved in aspect-based sentiment analysis is aspect and sentiment expressions relation extraction for Chinese e-commerce reviews, researched by Chen et al. (2018). For researches with Indonesian corpus, there are no such task yet. This final project adapts the research by Chen et al. (2018) for Indonesian hotel reviews.
Expressions relation extraction in this final project is conducted in three steps, such as generation of aspect and sentiment expression pair candidates from the review, pair candidate feature extraction, and classification of pair candidates into valid or invalid classes. Classification is done by using gradient boosting with additional handling procedure for training and prediction of cases where a review has exactly one sentiment expression. Best features that are used are statistical, positional, semantic, and semantic similarity feature groups. Values of gradient boosting hyperparameters are determined with greedy experiment scheme. The training data consists of 4000 unique reviews with 27894 pair candidates in them.
Testing procedure is done using F1 score metric with data of 1000 reviews with 7085 pair candidates in them. The resulting macro F1 score is 0.934. |
|---|