-
Flexible Generation of Natural Language Deductions
Kaj Bostrom, Xinyu Zhao, Swarat Chaudhuri, and Greg Durrett. Virtual Poster Session 2 (Monday 12:30pm AST). Talk page
We build a model to generate valid deductions in natural language: statements that represent useful logical inferences from a handful of input premises. This model is a crucial building block for systems that can do long-horizon, explainable reasoning in natural language; come talk to us about our goals with this!
-
Connecting Attributions and QA Model Behavior on Realistic Counterfactuals
Xi Ye, Rohan Nair, and Greg Durrett. Virtual Poster Session 2 (Monday 12:30pm AST). Talk page
Explanations should tell us about how models behave on counterfactuals; if the input were different in some way, how would the prediction change? But which counterfactuals? We present a way to analyze the validity of explanations for reasoning about a set of realistic counterfactuals, with a focus on QA. Commonly used attribution techniques fall short on this measure of validity, but we show some results suggesting a class of better techniques that might improve things!
-
[Findings] Can NLI Models Verify QA Systems' Predictions?
Jifan Chen, Eunsol Choi, and Greg Durrett. To be presented at MRQA
A combination of existing models for manipulating questions and passages in natural language can enable off-the-shelf NLI models to check the work of QA systems. We show that this can improve confidence in our systems and even in our datasets through an auditing process.
-
[Findings] Optimal Neural Program Synthesis from Multimodal Specifications
Xi Ye, Qiaochu Chen, Isil Dillig, and Greg Durrett.
We present a neural program synthesis model that takes natural language and examples as input. Unlike other language-to-code models, we use sophisticated pruning techniques based on program semantics to guide the search and find the optimal program according to the model. We show strong improvements on a regex synthesis task from our prior work.
