Learning to Rank Short Text Pairs with Convolutional Deep Neural Networks

摘要: Learning a similarity function between pairs of objects is at the core learning to rank approaches. In information retrieval tasks we typically deal with query-document pairs, in question answering -- question-answer pairs. However, before can take place, such needs be mapped from original space symbolic words into some feature encoding various aspects their relatedness, e.g. lexical, syntactic and semantic. Feature engineering often laborious task may require external knowledge sources that are not always available or difficult obtain. Recently, deep approaches have gained lot attention research community industry for ability automatically learn optimal representation given task, while claiming state-of-the-art performance many computer vision, speech recognition natural language processing. this paper, present convolutional neural network architecture reranking short texts, where text relate them supervised way training data. Our takes only input, thus requiring minimal preprocessing. particular, consider elements pair sentences. We test our system on two popular TREC: Question Answering Microblog Retrieval. model demonstrates strong first beating previous systems by about 3\% absolute points both MAP MRR shows comparable results tweet reranking, enjoying benefits no manual additional parsers.