A Sensitivity Analysis of (and Practitioners’ Guide to) Convolutional Neural Networks for Sentence Classification

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 253–263, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP A Sensitivity Analysis of (and Practitioners’ Guide to) Convolutional Neural Networks for Sentence Classification Ye Zhang Dept. of Computer Science University of Texas at Austin yezhang@utexas.edu Byron C. Wallace College of Computer and Information Science Northeastern University byron@ccs.neu.edu Abstract Convolutional Neural Networks (CNNs) have recently achieved remarkably strong performance on the practically important task of sentence classification (Kim, 2014; Kalchbrenner et al., 2014; Johnson and Zhang, 2014; Zhang et al., 2016). How- ever, these models require practitioners to specify an exact model architecture and set accompanying hyperparameters, includ- ing the filter region size, regularization pa- rameters, and so on. It is currently un- known how sensitive model performance is to changes in these configurations for the task of sentence classification. We thus conduct a sensitivity analysis of one-layer CNNs to explore the effect of architecture components on model performance; our aim is to distinguish between important and comparatively inconsequential design decisions for sentence classification. We focus on one-layer CNNs (to the exclu- sion of more complex models) due to their comparative simplicity and strong empiri- cal performance, which makes it a modern standard baseline method akin to Support Vector Machine (SVMs) and logistic re- gression. We derive practical advice from our extensive empirical results for those interested in getting the most out of CNNs for sentence classification in real world settings. 1 Introduction Convolutional Neural Networks (CNNs) have re- cently been shown to achieve impressive results on the practically important task of sentence cate- gorization (Kim, 2014; Kalchbrenner et al., 2014; Wang et al., 2015; Goldberg, 2015; Iyyer et al., 2015; Zhang et al., 2016, 2017). CNNs can cap- italize on distributed representations of words by first converting the tokens comprising each sen- tence into a vector, forming a matrix to be used as input (e.g., see Fig. 1). The models need not be complex to realize strong results: Kim (2014), for example, proposed a simple one-layer CNN that achieved state-of-the-art (or comparable) results across several datasets. The very strong results achieved with this comparatively simple CNN ar- chitecture suggest that it may serve as a drop-in replacement for well-established baseline models, such as SVM (Joachims, 1998) or logistic regres- sion. While more complex deep learning models for text classification will undoubtedly continue to be developed, those deploying such technolo- gies in practice will likely be attracted to simpler variants, which afford fast training and prediction times. Unfortunately, a downside to CNN-based mod- els – even simple ones – is that they require prac- titioners to specify the exact model architecture to be used and to set the accompanying hyperpa- rameters. In practice, tuning all of these hyper- parameters is simply not feasible, especially be- cause parameter estimation is computationally in- tensive. Emerging research has begun to explore hyperparameter optimization methods, including random search (Bengio, 2012), and Bayesian op- timization (Yogatama and Smith, 2015; Bergstra et …

Content truncated to 500 words