Deep learning-based idiomatic expression recognition for the Amharic language.

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Machine Learning

Background:

• Idiomatic expressions are challenging for natural language processing (NLP) due to their non-literal meanings.
• Existing NLP models for Amharic often overlook idioms, impacting performance in tasks like machine translation and sentiment analysis.
• Idioms are prevalent in Amharic conversation and literature.

Purpose of the Study:

• To propose and evaluate a novel model for detecting idiomatic expressions in Amharic text.
• To address the limitations of current NLP models in handling Amharic idioms.
• To enhance the accuracy of various Amharic NLP applications by incorporating idiom detection.

Main Methods:

• Developed a convolutional neural network (CNN) model integrated with FastText word embeddings.
• Collected a dataset of 1700 idiomatic and 1600 non-idiomatic Amharic expressions from books.
• Trained and tested the model using an 80/10/10 data split for training, validation, and testing.

Main Results:

• The proposed CNN-FastText model achieved 98% learning accuracy on the training dataset.
• The model demonstrated 80% accuracy on the unseen testing dataset.
• Performance was compared favorably against traditional machine learning classifiers like KNN, SVM, and Random Forest.

Conclusions:

• The CNN-FastText model shows significant promise for accurate idiomatic expression detection in Amharic.
• This approach can improve the performance of downstream NLP tasks for the Amharic language.
• Further research can build upon this model to enhance Amharic language understanding in AI systems.