SEMANTIC SIMILARITY MODELING BASED ON MULTI-GRANULARITY INTERACTION MATCHING

摘要

Determining whether two sentences are semantically equivalent is complicated by the ambiguity and variability of natural language expression. The most approaches have used classifiers employing hand engineered features derived from complex natural language processing pipelines to automatically recognize equivalence relations; thus, the performances of the models heavily rely on the features designing. To avoid specific assumptions about the underlying language, we propose a recurrent neural network model for semantic similarity. Interaction features and text representations on multiple levels of granularity are automatically learned using a conditional bidirectional long short-term memory encoder. We extend this model with a soft-alignment attention mechanism that encourages fine-grained reasoning over equivalence or contradiction of pairs of words and phrases. The sentence-pair encoding is input to an output layer to determine the classification. The effectiveness of our model is demonstrated using two tasks: paraphrase identification and semantic relatedness measurement. The results on MRPC and SICK datasets show that our model leads to significant quality improvement on tasks, exceeding the previous state-of-the-art without using any hand-crafted features.