An Efficient Approach for Findings Document Similarity Using Optimized Word Mover’s Distance

Abstract

We introduce Optimized Word Mover’s Distance (OWMD), a similarity function that compares two sentences based on their word embeddings. The method determines the degree of semantic similarity between two sentences considering their interdependent representations. Within a sentence, all the words may not be relevant for determining contextual similarity at the aspect level with another sentence. To account for this fact, we designed OWMD in two ways: first, it decreases system’s complexity by selecting words from the sentence pair according to a predefined set of dependency parsing criteria; Second, it applies the word mover’s distance (WMD) method to previously chosen words. When comparing the dissimilarity of two text sentences, the WMD method is used because it represents the minimal “journey time” required for the embedded words of one sentence to reach the embedded words of another sentence. Finally, adding an exponent function to the inverse of the OWMD dissimilarity score yields the resulting similarity score, called Optimized Word Mover’s Similarity (OWMS). Using STSb-Multi-MT dataset, the OWMS measure decreases MSE, RMSE, and MAD error rates by $66.66\%$, $40.70\%$, and $37.93\%$ respectively than previous approaches. Again, OWMS reduces MSE, RMSE, and MAD error rates on Semantic Textual Similarity (STS) dataset by $85.71\%$, $62.32\%$, and $60.17\%$ respectively. For STSb-Multi-MT and STS datasets, the suggested strategy reduces run-time complexity by $33.54\%$ and $49.43\%$, respectively, compared to the best of existing approaches.