2020/09/01 by Tushar Khot, Khot, Tushar, Daniel Khashabi +7 · 1 citation
Computer Science · #Artificial Intelligence (cs.AI) #Computation and Language (cs.CL) #Explainable Artificial Intelligence (XAI) #FOS: Computer and information sciences #Natural Language Processing Techniques #Topic Modeling
paper · pdf · doi:10.48550/arxiv.2009.00751
openalex publication_date 2020/09/01 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
We propose a general framework called Text Modular Networks(TMNs) for\nbuilding interpretable systems that learn to solve complex tasks by decomposing\nthem into simpler ones solvable by existing models. To ensure solvability of\nsimpler tasks, TMNs learn the textual input-output behavior (i.e., language) of\nexisting models through their datasets. This differs from prior\ndecomposition-based approaches which, besides being designed specifically for\neach complex task, produce decompositions independent of existing sub-models.\nSpecifically, we focus on Question Answering (QA) and show how to train a\nnext-question generator to sequentially produce sub-questions targeting\nappropriate sub-models, without additional human annotation. These\nsub-questions and answers provide a faithful natural language explanation of\nthe model's reasoning. We use this framework to build ModularQA, a system that\ncan answer multi-hop reasoning questions by decomposing them into sub-questions\nanswerable by a neural factoid single-span QA model and a symbolic calculator.\nOur experiments show that ModularQA is more versatile than existing explainable\nsystems for DROP and HotpotQA datasets, is more robust than state-of-the-art\nblackbox (uninterpretable) systems, and generates more understandable and\ntrustworthy explanations compared to prior work.\n