Capturing Discriminative Attributes Using Convolution Neural Network Over ConceptNet Numberbatch Embedding

A semantic representation of text helps us understand the lexical association between words. Capturing these associations becomes an integral part of perceiving any language. One such fundamental property that expresses these associations is ‘similarity.’ This property shows consistency in comprehending similar words in the same manner. However, seldom it falls short on specific tasks where the lexical similarity in itself is not sufficient enough to validate the semantic representations. In this paper, the objective is to capture such semantic distinctions. It is based on the shared task ‘capturing discriminative attributes’ conducted in SemEval-2018. Our team participated in the task and held an F1 score of 0.658 with GloVe representation. An extension to this work is taken up in this paper where a new embedding known as ConceptNet Numberbatch is explored. The ConceptNet word embedding in overall showed improvement in the scores with the previous rule-based feature representation. The model is further tuned over certain hyperparameters, which improved the score to as much as 6%. A comparison is also put forth here with another prominent embedding like FastText. The ConceptNet model achieved a near par score with the state of the art, based only on a simple ensemble of features as representation.