Named Entity Recognition for Social Media Texts with Semantic Augmentation

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 137–150, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 137 Learning to Explain: Datasets and Models for Identifying Valid Reasoning Chains in Multihop Question-Answering Harsh Jhamtani 1 Peter Clark 2 1 School of Computer Science, Carnegie Mellon University 2 Allen Institute for AI jharsh@cs.cmu.edu, peterc@allenai.org Abstract Despite the rapid progress in multihop question-answering (QA), models still have trouble explaining why an answer is correct, with limited explanation training data avail- able to learn from. To address this, we intro- duce three explanation datasets in which ex- planations formed from corpus facts are an- notated. Our first dataset, eQASC, contains over 98K explanation annotations for the mul- tihop question answering dataset QASC, and is the first that annotates multiple candidate explanations for each answer. The second dataset eQASC-perturbed is constructed by crowd-sourcing perturbations (while pre- serving their validity) of a subset of explana- tions in QASC, to test consistency and gen- eralization of explanation prediction models. The third dataset eOBQA is constructed by adding explanation annotations to the OBQA dataset to test generalization of models trained on eQASC. We show that this data can be used to significantly improve explanation quality (+14% absolute F1 over a strong retrieval base- line) using a BERT-based classifier, but still behind the upper bound, offering a new chal- lenge for future research. We also explore a delexicalized chain representation in which re- peated noun phrases are replaced by variables, thus turning them into generalized reasoning chains (for example: ”X is a Y” AND ”Y has Z” IMPLIES ”X has Z”). We find that gener- alized chains maintain performance while also being more robust to certain perturbations.1 1 Introduction While neural systems have become remarkably adept at question answering (QA), e.g., (Clark and Gardner, 2018), their ability to explain those an- swers remains limited. This creates a barrier for deploying QA systems in practical settings, and 1Code and datasets can be found at https://allenai. org/data/eqasc Q: What can cause a forest fire? (1) rain (2) static electricity (3) microbes (4) ... A: static electricity Q+A (declarative): Static electricity can cause a forest fire. Explanation (reasoning chain): [positive (valid)] Static electricity can cause sparks // (from corpus) AND Sparks can start a forest fire → Static electricity can cause a forest fire// (from corpus) // (Q+A) Explanation (Generalized reasoning chain, GRC): X can cause Y AND Y can start Z → X can cause Z Figure 1: Our datasets contain annotated (valid and in- valid) reasoning chains in support of an answer, allow- ing explanation classifier models to be trained and ap- plied. We also find that using a variabilized version of the chains improves the models’ robustness. limits their utility for other tasks such as education and tutoring, where explanation plays a key role. This need has become particularly important with multihop question-answering, where multiple facts are needed to derive an answer. In this context, seeing a chain of reasoning leading to an …

Content truncated to 500 words