MASTER: A Machine Led Solution to Amharic Arithmetic Word Problem

Abstract

Arithmetic Word problem solving is a challenging yet exciting task requiring an accurate understanding of a given problem text to solve the underlying arithmetic equation. Several research works have been proposed to solve arithmetic word problems in English and other languages. The Amharic language has fewer natural language processing (NLP) resources and has morphologically complex verbing with a unique sentence structure. Though the existing machine translation approaches have shown tremendous progress, translating the contextual meaning within a given sentence or phrase still needs to be completed. Moreover, arithmetic word problems contain contextual meanings which need to be interpreted precisely within the context of the language structure, which the current machine translation practices could not help much. Hence, a strategy based on the linguistic structure of the language to understand and solve a given Amharic arithmetic word problem is needed. By far, previous research has yet to be conducted in the area. This paper proposes a novel strategy using a schema-based machine-led solution to Amharic arithmetic word problems (AWP) as a step towards enabling comprehension in mathematics and teaching problem- solving for children in the elementary grades. The proposed Amharic AWP solver involves four stages to a complete solution: NLP task of preprocessing on a given problem text, simplification of the problem text into more straightforward sentences, knowledge representation using Amharic schema to understand the discourse structure in addition to extracting relevant concepts that led to a systematic solution and finally, generation of a natural language answer based on the propositions made under knowledge representation phase. The learning component in the proposed strategy uses schema production rules to define and incorporate a new schema that allows for handling new varieties of problems. To validate the proposed approach, a prototype has been developed using python, and experimental results have shown an accuracy of 88.81% on a large corpus of Amharic arithmetic word problems.