UTILIZATION OF COMPLEX FEATURES BY EXPLOITATION OF EXTERNAL KNOWLEDGE ON ASPECT BASED SENTIMENT ANALYSIS

• Main Author: Zakhralativa Ruskanda, Fariska
• Format: Dissertations
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/57030
• Institution: Institut Teknologi Bandung

With the rapid development of the internet, aspect-based sentiment analysis (ABSA) has become a rapidly growing research topic. However, this topic, which is closely related to Natural Language Processing (NLP), still leaves some unresolved challenges. Some of these challenges include the representation of aspects in various words and phrases, various part-of-speech both nouns and non-nouns, the varying complexity of sentences, which are very difficult to handle with simple syntactic rules. Several approaches were taken to obtain a more complete and complex representation of the text. This complex representation, or for short: a complex feature, is a feature that contains a set of basic features that are used with certain rules in a tiered manner. This study aims to produce a complex feature extraction method in ABSA, which is generated by using external knowledge, as input for the sentiment analysis process based on supervised learning. Complex feature extraction is expected to be able to handle all appearance patterns of aspects, opinions, and the relationship between the two. Meanwhile, external knowledge is expected to provide consideration in identifying aspects and opinions. This research was conducted based on the hypothesis that (1) The use of complex features in the form of interactions between various basic features in the review sentence is more relevant for ABSA. Then, the complex features along with the aspect extraction rules obtained from the learning results can provide more accurate ABSA results compared to the hand-written extraction rules; and (2) Optimization of dependency feature on complex features dynamically combined with classifier and external knowledge can improve the extraction performance compared to a rule-based approach. This research contributes in producing an aspect extraction method in ABSA that optimizes the use of dependency features and the learning process automatically on cross-domain review sentence data. In addition, the proposed method is more flexible because it does not limit the word’s part-of-speech and the results are more accurate because it is able to handle expressions in the form of phrases. The specific contributions of this research are (1) A rule learning method for aspect extraction, namely Dependency – Sequential Covering, and (2) An aspect extraction method that optimizes the use of dynamic dependency features and pairwise classification in order to improve the performance of the extraction results. The Dependency-Sequential Covering rule learning method is a learning algorithm used to automatically construct extraction rules, which mainly uses the dependency syntax feature in the review sentence. The advantage of this method is that the aspect extraction rules are clear, easy to understand, and do not require a seed opinion and seed rule set. The rule set obtained from the learning outcomes is used in conjunction with external knowledge to extract the rule-based aspects of the cross-domain review dataset. The test results on 4 dataset reviews from 4 product domains, show that the Dependency-Sequential Covering method outperforms the baseline (Double Propagation and Aspectator) for the f-measure value with the highest f-measure value of 0.633. Aspect extraction method with dependency-based complex feature optimization and pairwise classification is the second method proposed. In this method, three new dependency features are introduced, namely: relation probability, focus node, and sentence clause. In addition, aspect confidence and opinion confidence scores are also used, which determine the appropriateness of aspect words and opinion words, involving the part-of-speech, external knowledge and sentiment scores. The proper aspect classification process is carried out by combining the existing pairwise classification algorithms and the existing binary classifiers. The test results on 2988 review sentences from 6 product domains show that this method is able to outperform the baseline (Dependency-Sequential Covering) by 19.3% on the f-measure value. The two methods produced by this dissertation research have also succeeded in proving this research hypothesis, namely: (1) Extraction of aspects in ABSA can be done by utilizing the syntactic characteristics of the language in the review sentence. The results of this study indicate that complex features in the form of dependency syntactic features combined with POS-tag and constituent parse trees features are more relevant for various domains of review sentences. This is indicated by its use of the extraction rules obtained from learning outcomes and provides more accurate results. (2) The use of complex features dynamically added by optimization of dependency features, along with pairwise classifiers and external knowledge can improve the performance of the extraction results compared to a rule-based approach.