Deep Learning Based Causal Relations Extraction for Afan Oromo Text

Abstract

Causal relations are cause and effect connections where one occurrence a cause influences the development of another event an effect. Causal relations extraction are a method which extract unstructured causal texts into a meaningful cause-effect structured categories. There is an alarming requirement to extract causal information from the growing amount of data. Previously, many researchers have done causal relations extraction for Amharic and English using machine learning and deep learning approaches. Nevertheless, no study has been done to identify causal relations from Afan Oromo text. This study aims to extract a causal relations from Afan Oromo text using a deep learning approaches. We gathered Afan Oromo causal relations corpus from Afan Oromo news sources, Oromia Communication Bureau, BBC News Afaan Oromoo, OBN Afaan Oromoo, Ethiopian Press Agency /Bariisaa, Tajaajila Oduu Itoophiyaa, Kallacha Oromiyaa, Waaltaa TV Afaan Oromoo and FBC Afaan Oromoo. The collected data were annotated under four classes by an expert. Text preprocessing tasks applied on the collected data. We have also used word2Vec and FastText word embedding to detect word similarity for our dataset. We have applied the proposed three models, namely LSTM, BiLSTM and CNN. The performance of causal relations extraction for Afan Oromo text was assessed using evaluation metrics after selecting a hyper-parameter for our model. The experimental results demonstrate that among the proposed models, BiLSTM outperformed LSTM and CNN, scoring an accuracy of 93.37% compared to LSTM 92.55% and CNN 84.57%, respectively.