Automatic Identification of Amharic Text Idiomatic Expressions Using a Deep Learning Approach

Natural Language Processing (NLP) is a tract of artificial intelligence and linguistics devoted to making computers understand the statements or words written in human languages. Amharic, the most widely spoken language in Ethiopia, uses a lot of idiomatic expressions and proverbs to emphasize the message of the text. The meaning of an idiomatic phrase cannot be inferred from individual words. Developing a model to identify Amharic idiomatic terms is helpful for different NLP applications like machine translation, sentiment analysis, spam classification, intent recognition, and so on. Few studies have been conducted to identify idiomatic expressions using K-Nearest Neighbors (KNN) and Convolutional Neural Network (CNN) algorithms for the Amharic language. The KNN model was designed to identify only dual-word idioms by neglecting idioms with three or more words, and this study didn’t use the n-gram technique to identify idioms found in a sentence or paragraph. Like the former model, the CNN model did not use the n-gram technique, and the testing accuracy was not greater than 80%. Due to this gap, we need to develop a deep learning model that identifies Amharic idioms constructed from two or more words. To identify idioms found in a sentence or paragraph, we used the n-gram method to divide the sentence or the paragraph into phrases for enhancing the previous works. For designing our model, we used 4053 idiomatic phrases and 4051 literal Amharic phrases. After the data were collected, text preprocessing techniques were applied to the collected data, and FastText and Word2Vec models were used for word embedding purposes. We conducted experiments with 70:30 and 80:20 data split ratios with FastText and Word2Vec word embedding models along with long short-term memory (LSTM) and bidirectional LSTM (Bi-LSTM) with and without attention layer algorithms. Among those experiments, the highest accuracy of 98.95% was attained using an 80:20 train-test split ratio, Adamax optimizer, 64 batch sizes, and a 0.001 learning rate by using Bi-LSTM with an attention layer and FastText word embedding model.