Resolving Amharic Lexical Ambiguity using Neural Word Embedding
Gashaw Demlew, Dessalew Yohannes · 2022
Despite the fact that there are numerous information sources available online, the search for the exact match with the user’s needs has been hampered for a variety of reasons. One issue is linguistic ambiguity, which exists at various levels such as semantic, syntactic, lexical/word, and so on. This paper focuses on Amharic word-level ambiguity and proposes a method for resolving it. Words can have more than one distinct meaning, and many words can be interpreted in different ways depending on the context. Automatically determining the correct sense of an ambiguous word is critical for many applications such as machine translation and information retrieval. To date, the majority of Amharic sense disambiguation work has been unsupervised/ knowledge base and based on WordNet. However, this approach has limitations in that the full Amharic WordNet is not available and manually preparing takes time and resources. To address the aforementioned limitations, this paper proposes a novel joint supervised and unsupervised sense disambiguation approach based on distributional semantic space. Each approach, as well as the combined approach, is assessed using a total of 10 ambiguous words. The supervised approach achieved 82.3 % of the F1-score and 70% accuracy, while the unsupervised approach achieved 85.7 % of the F1-score and 60% accuracy. The joint approach achieved an F1-score of 92.5 % and an accuracy of 86 %.