Semantic similarity between short paragraphs using Deep Learning

Textual semantic similarity plays an increasingly important role in tasks such as information retrieval, text mining and text-based searches. Multiple approaches have been presented to enhance methods for information retrieval by understanding the underlying meaning of sentences. However, most of these focus on single line sentences. In this paper, we try to evaluate the effectiveness of these approaches to understand the semantic meaning of short paragraphs. We use an existing recurrent neural network architecture and train it using document embedding vectors to try and infer the meaning of small paragraphs consisting of one, two or three sentences. We use three different methods - Manhattan distance, Euclidean distance and cosine distance - to evaluate the performance and effectiveness of measuring the semantic similarity. The conclusion compares the performance of all three methods.