Neural Word Embedding Approach to Resolve Semantic Ambiguities for Amharic Language

Abstract

The ambiguity problem in a natural language is a common problem, that it needs to be solved to get the exact context or meaning of the sentence intended to be specified. Humans are naturally adept at word sense disambiguation and can recognize different senses in words via spoken language. Computers on the other hand have trouble recognizing accurate word meanings. We have different kinds of ambiguities in natural language, in this thesis we will consider the semantic ambiguities. The goal of this thesis is to describe a machine learning model based on supervised and unsupervised word embedding on neural networks. Basically we have used supervised and unsupervised neural word embedding techniques to solve the semantic lexical ambiguities seen in Amharic language. We have developed two approaches: the supervised and unsupervised word sense approach. In this supervised version the dataset or documents with ambiguous words are manually labeled with the support of language linguistics. In the second approach we have used the sense definition (gathered from Babelnet) as a sense vector so that the input sentence will be converted to context vector and their context similarities will be checked with the sense value. Amount of dataset used was more than 36,000 documents or 1 million amount in word level for testing and development. The F1 measure we have got is 92.5% and accuracy of 86.0%.