Textual Inference Identification in the Malayalam Language Using Convolutional Neural Network

Abstract

AbstractNatural language inference (NLI), earlier known as textual entailment, is an important task related to the semantic matching of natural language sentences. Systems and methodologies that can identify inferences are helpful for most language processing tasks like document summarization and question answering systems. There is active research in NLI for English and other foreign languages. Considering Indian languages like Malayalam, there are very few works done. Here, we focus on identifying inferences in the Malayalam language using one-dimensional convolutional neural network (CNN), multichannel CNN, and CNN architecture on matching sentences over interaction space with fastText-based embeddings. This work is an attempt to apply convolutional neural networks for NLI in Malayalam without any hand-engineered features. This approach contributes with a recall of 0.66% for binary and 0.51% for multiclass classification. This work also contributes to the language resources community.KeywordsMalayalamNatural language inferenceText entailmentDeep learningCNN