Rationalizing Neural Predictions

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 107–117, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Rationalizing Neural Predictions Tao Lei, Regina Barzilay and Tommi Jaakkola Computer Science and Artificial Intelligence Laboratory Massachusetts Institute of Technology {taolei, regina, tommi}@csail.mit.edu Abstract Prediction without justification has limited ap- plicability. As a remedy, we learn to extract pieces of input text as justifications – ratio- nales – that are tailored to be short and co- herent, yet sufficient for making the same pre- diction. Our approach combines two modu- lar components, generator and encoder, which are trained to operate well together. The gen- erator specifies a distribution over text frag- ments as candidate rationales and these are passed through the encoder for prediction. Ra- tionales are never given during training. In- stead, the model is regularized by desiderata for rationales. We evaluate the approach on multi-aspect sentiment analysis against manu- ally annotated test cases. Our approach out- performs attention-based baseline by a signif- icant margin. We also successfully illustrate the method on the question retrieval task.1 1 Introduction Many recent advances in NLP problems have come from formulating and training expressive and elabo- rate neural models. This includes models for senti- ment classification, parsing, and machine translation among many others. The gains in accuracy have, however, come at the cost of interpretability since complex neural models offer little transparency con- cerning their inner workings. In many applications, such as medicine, predictions are used to drive criti- cal decisions, including treatment options. It is nec- essary in such cases to be able to verify and under- 1Our code and data are available at https://github. com/taolei87/rcnn. the$beer$was$n’t$what$i$expected,$and$i‘m$not$sure$it’s$“true$ to$ style“,$ but$ i$ thought$ it$ was$ delicious.$ a" very" pleasant" ruby"red-amber"color$with$a$rela9vely$brilliant$finish,$but$a$ limited$amount$of$carbona9on,$from$the$look$of$it.$aroma$is$ what$ i$ think$ an$ amber$ ale$ should$ be$ -$ a$ nice$ blend$ of$ caramel$and$happiness$bound$together. Review Ratings Look: 5 stars Smell: 4 stars Figure 1: An example of a review with ranking in two cate- gories. The rationale for Look prediction is shown in bold. stand the underlying basis for the decisions. Ide- ally, complex neural models would not only yield improved performance but would also offer inter- pretable justifications – rationales – for their predic- tions. In this paper, we propose a novel approach to in- corporating rationale generation as an integral part of the overall learning problem. We limit ourselves to extractive (as opposed to abstractive) rationales. From this perspective, our rationales are simply sub- sets of the words from the input text that satisfy two key properties. First, the selected words represent short and coherent pieces of text (e.g., phrases) and, second, the selected words must alone suffice for prediction as a substitute of the original text. More concretely, consider the task of multi-aspect senti- ment analysis. Figure 1 illustrates a product review along with user rating in terms of two categories or aspects. If the model in this case predicts five star rating for color, it should also identify the phrase ”a very pleasant …

Content truncated to 500 words