Experiments with Convolutional Neural Network Models for Answer Selection

摘要

In recent years, neural networks have been applied to many textrnprocessing problems. One example is learning a similarity functionrnbetween pairs of text, which has applications to paraphrasernextraction, plagiarism detection, question answering, and ad hocrnretrieval. Within the information retrieval community, the convolutionalrnneural network model proposed by Severyn and MoschiŠirnin a SIGIR 2015 paper has gained prominence. Œis paper focusesrnon the problem of answer selection for question answering: wernaŠempt to replicate the results of Severyn and MoschiŠi using theirrnopen-source code as well as to reproduce their results via a de novorn(i.e., from scratch) implementation using a completely di‚erentrndeep learning toolkit. Our de novo implementation is instructivernin ascertaining whether reported results generalize across toolkits,rneach of which have their idiosyncrasies. We were able to successfullyrnreplicate and reproduce the reported results of Severynrnand MoschiŠi, albeit with minor di‚erences in e‚ectiveness, butrnarming the overall design of their model. Additional ablation experimentsrnbreak down the components of the model to show theirrncontributions to overall e‚ectiveness. Interestingly, we €nd thatrnremoving one component actually increases e‚ectiveness and thatrna simpli€ed model with only four word overlap features performsrnsurprisingly well, even beŠer than convolution feature maps alone