Deeply Embedded Knowledge Representation & Reasoning For Natural Language Question Answering: A Practitioner’s Perspective

Prediction or Comparison: Toward Interpretable Qualitative Reasoning

“…Tafjord et al (2019a) solves such tasks using semantic parsing and a symbolic solver; this type of method naturally falls short on performance when compared to neural systems. Meanwhile, some works tackle tasks using a data-driven "black-box" model based on representations computed by language models, which achieves superior performance (Tafjord et al, 2019b;Mitra et al, 2019;Asai and Hajishirzi, 2020;Mitra et al, 2020). Our work enjoys the performance improvement that a neural network brings but also provides interpretability.…”

“…DeepEKR: Similar to Mitra et al (2019), this model replaces the parser with a neural network, softening the symbolic representation (Mitra et al, 2020).…”

“…However, the most current datasets on qualitative reasoning are much smaller than standard Question Answering datasets, making the task challenging and receiving limited attention. At present, in the two mainstream qualitative relationship question datasets, QuaRel (Tafjord et al, 2019a) and QuaRTz (Tafjord et al, 2019b), the existing methods can be divided into two categories, symbolic reasoning based on a semantic parser (Krishnamurthy et al, 2017;Tafjord et al, 2019a) and a "black-box" model based on representations (Mitra et al, 2019;Tafjord et al, 2019b;Asai and Hajishirzi, 2020;Mitra et al, 2020).…”