Neural Attention for Learning to Rank Questions in Community Question Answering

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1734–1745, Osaka, Japan, December 11-17 2016. Neural Attention for Learning to Rank Questions in Community Question Answering Salvatore Romeo, Giovanni Da San Martino, Alberto Barr´on-Cede˜no and Alessandro Moschitti Qatar Computing Research Institute, HBKU, Doha, Qatar {sromeo, gmartino, albarron, amoschitti}@qf.org.qa Yonatan Belinkov, Wei-Ning Hsu, Yu Zhang, Mitra Mohtarami and James Glass MIT Computer Science and Artificial Intelligence Laboratory Cambridge, MA 02139, USA {belinkov, wnhsu, yzhang87, mitram, glass}@mit.edu Abstract In real-world data, e.g., from Web forums, text is often contaminated with redundant or irrele- vant content, which leads to introducing noise in machine learning algorithms. In this paper, we apply Long Short-Term Memory networks with an attention mechanism, which can select impor- tant parts of text for the task of similar question retrieval from community Question Answering (cQA) forums. In particular, we use the attention weights for both selecting entire sentences and their subparts, i.e., word/chunk, from shallow syntactic trees. More interestingly, we apply tree kernels to the filtered text representations, thus exploiting the implicit features of the subtree space for learning question reranking. Our results show that the attention-based pruning allows for achieving the top position in the cQA challenge of SemEval 2016, with a relatively large gap from the other participants while greatly decreasing running time. 1 Introduction Previous work on modeling high-level semantic tasks, e.g., paraphrasing, recognizing textual entailment, and question answering (QA), has shown that syntactic and semantic structures are essential for boosting the accuracy of the applied machine learning algorithm. However, when dealing with real-world data, automatic parsers typically decrease their accuracy. Additionally, there is often a considerable amount of meaningless text for the task, contributing to generating noisy structures. This kind of phenomena can be clearly observed in new Web applications, such as community Question Answering (cQA), where the presence of informal, redundant, and often unrelated text constitutes a major challenge for learning the automatic detection of related topics. For example, given the original question: Which all places are there for tourists to Qatar? My nephew 18 years on visit. a cQA system has to infer whether the following question is related: What are the tourist places in Qatar? I’m likely to travel in the month of june. Just wanna know some good places to visit. This implies that the system has to (i) recognize the sentences, My nephew 18 years on visit and I’m likely to travel in the month of june, as irrelevant for the classification task; and (ii) deal with non- standard writing (e.g., wanna). If the above steps are carried out with good accuracy, complex semantic models can be applied to the selected text to achieve better performance. In particular, such models have to encode structural relations between the constituents of the two questions, e.g., places are there for tourists and some good places to visit, in the learning algorithm. This is a challenging task as we do not know which relations might be useful, whereas explicitly including …

Content truncated to 500 words