Abstract

Integrating distributed representations with symbolic operations is essential for reading comprehension requiring complex reasoning, such as counting, sorting and arithmetics, but most existing approaches rely on specialized neural modules and are hard to adapt to multiple domains or multi-step reasoning. In this work, we propose the Neural Symbolic Reader (NeRd), which includes a reader, e.g., BERT, to encode the passage and question, and a programmer, e.g., LSTM, to generate a program for multi-step reasoning. By using operators like span selection, the program can be executed over text to generate the answer. Compared to previous works, NeRd is more scalable in two aspects: (1) domain-agnostic, i.e., the same neural architecture works for different domains; (2) compositional, i.e., complex programs can be generated by compositionally applying the symbolic operators. Furthermore, to overcome the challenge of training NeRd with weak supervision, we apply data augmentation techniques and hard ExpectationMaximization (EM) with thresholding. On DROP, a challenging reading comprehension dataset requiring discrete reasoning, NeRd achieves 1.37%/1.18% absolute gain over the state-of-the-art on Exact-Match/F1 metrics. With the same architecture, NeRd significantly outperforms the baselines on MathQA, a math problem benchmark that requires multiple steps of reasoning, by 25.5% absolute gain on accuracy when trained on all the annotated programs, and more importantly, still beats the baselines even with only 20% of the program annotations.