Context-based model for domain-specific lexicon building / Tasnim M. A. Zayet

Tasnim , M. A. Zayet (2024) Context-based model for domain-specific lexicon building / Tasnim M. A. Zayet. PhD thesis, Universiti Malaya.

	PDF (The Candidate's Agreement) Restricted to Repository staff only Download (228Kb)
	PDF (Thesis PhD) Restricted to Repository staff only until 31 December 2026. Download (1766Kb)

Abstract

With the numerous amounts of users’ generated data over the internet and the need to analyse it and get their opinions regarding a service, sentiment analysis has emerged. To perform sentiment analysis, many approaches have emerged. The easiest one is sentiment lexicons. There are two types of sentiment lexicons: general and domain-specific. General lexicons have a static public sentiment of the words, while domain lexicons have different sentiments of a word depending on the context. Domain lexicons usually extract the opinion pairs formed from the main domain noun and its corresponding opinion word. Sentiment will be assigned to the pair. One of the popular methods for this aim is frequency-based approaches. Frequency-based approaches are widely used in the field due to their easy implementation. However, these approaches suffer from the ambiguity problem. To overcome this problem, we proposed a new frequency-based model. The new model is a contextual-aware domain lexicon generation model. In the proposed model, a new contextual-aware frequency-based equation was proposed. It considers the nouns and their cooccurrences in the score calculation to extract the top n nouns and filter the candidate opinion pairs. The model has three main modules: a domain terms identification module, a context-based lexicon construction module and a sentiment assignment module. The proposed model considers the verb and noun sentiment during the sentiment assignment process. A two-step evaluation was done to evaluate the model, first to evaluate the proposed equation and then to evaluate the model. The equation was compared with other popular equations in building domain lexicons term frequency-inverse document frequency (TF-IDF) and pointwise mutual information (PMI), while the model was compared with the performance of other general lexicons. The evaluation was conducted using five datasets from Amazon reviews datasets: Sport & Outdoors, CDs & Vinyl, Fashion, Electronics and Appliances. The equation was evaluated on different percentages of top n, mainly 20%, 40%, and 60%. Both equation and model proved their efficiency and outperformed other approaches regarding recall and precision in most cases. The results exposed the importance of “context” in lexicon building and in decreasing the effect of ambiguity problems besides the significance of the sentiment of the nouns and verbs leading to effective domain-specific lexicons. Domain-specific lexicons result in more efficient classification of the sentiment compared to general ones besides holding sentiment regarding specific features and aspects rather than general static sentiment of the opinion words.

Actions (For repository staff only : Login required)