NLProlog: Reasoning with Weak Unification for Question Answering in Natural Language

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6151–6161 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 6151 NLProlog: Reasoning with Weak Unification for Question Answering in Natural Language Leon Weber Humboldt-Universität zu Berlin weberple@hu-berlin.de Pasquale Minervini University College London p.minervini@ucl.ac.uk Jannes Münchmeyer GFZ German Research Center for Geoscience Potsdam munchmej@gfz-potsdam.de Ulf Leser Humboldt-Universität zu Berlin leser@informatik.hu-berlin.de Tim Rocktäschel University College London t.rocktaschel@cs.ucl.ac.uk Abstract Rule-based models are attractive for various tasks because they inherently lead to inter- pretable and explainable decisions and can eas- ily incorporate prior knowledge. However, such systems are difficult to apply to prob- lems involving natural language, due to its lin- guistic variability. In contrast, neural models can cope very well with ambiguity by learn- ing distributed representations of words and their composition from data, but lead to mod- els that are difficult to interpret. In this pa- per, we describe a model combining neural net- works with logic programming in a novel man- ner for solving multi-hop reasoning tasks over natural language. Specifically, we propose to use a Prolog prover which we extend to utilize a similarity function over pretrained sentence encoders. We fine-tune the representations for the similarity function via backpropagation. This leads to a system that can apply rule- based reasoning to natural language, and in- duce domain-specific rules from training data. We evaluate the proposed system on two dif- ferent question answering tasks, showing that it outperforms two baselines – BIDAF (Seo et al., 2016a) and FASTQA (Weissenborn et al., 2017b) on a subset of the WIKIHOP corpus and achieves competitive results on the MEDHOP data set (Welbl et al., 2017). 1 Introduction We consider the problem of multi-hop reason- ing on natural language data. For instance, con- sider the statements “Socrates was born in Athens” and “Athens belongs to Greece”, and the question “Where was Socrates born?”. There are two possi- ble answers following from the given statements, namely “Athens” and “Greece”. While the answer “Athens” follows directly from “Socrates was born in Athens”, the answer “Greece” requires the reader to combine both statements, using the knowledge that a person born in a city X, located in a coun- try Y , is also born in Y . This step of combin- ing multiple pieces of information is referred to as multi-hop reasoning (Welbl et al., 2017). In the literature, such multi-hop reading comprehension tasks are frequently solved via end-to-end differ- entiable (deep learning) models (Sukhbaatar et al., 2015; Peng et al., 2015; Seo et al., 2016b; Raison et al., 2018; Henaff et al., 2016; Kumar et al., 2016; Graves et al., 2016; Dhingra et al., 2018). Such models are capable of dealing with the linguis- tic variability and ambiguity of natural language by learning word and sentence-level representa- tions from data. However, in such models, explain- ing the reasoning steps leading to an answer and interpreting the model parameters to extrapolate new knowledge is a very challenging task (Doshi- Velez …

Content truncated to 500 words