Timezone: »

PINTO: Faithful Language Reasoning Using Prompt-Generated Rationales
Peifeng Wang · Aaron Chan · Filip Ilievski · Muhao Chen · Xiang Ren
Event URL: https://openreview.net/forum?id=axuiKw0IRC »

Neural language models (LMs) have achieved impressive results on various language-based reasoning tasks by utilizing latent knowledge encoded in their own pretrained parameters. To make this reasoning process more explicit, recent works retrieve a rationalizing LM's internal knowledge by training/prompting it to generate free-text rationales, which can be used to guide task predictions made by either the same LM or a separate reasoning LM. However, rationalizing LMs require expensive rationale annotation, without any assurance that the generated rationales improve LM task performance or faithfully reflect LM decision-making. In this paper, we propose PINTO, an LM pipeline that rationalizes via prompt-based learning, and learns to faithfully reason over rationales via counterfactual regularization. First, PINTO maps out a suitable reasoning process for the task input by prompting a frozen rationalizing LM to generate a free-text rationale. Second, PINTO's reasoning LM is fine-tuned to solve the task using the generated rationale as context, while regularized to output less confident predictions when the rationale is perturbed. Across four datasets, we show that PINTO significantly improves the generalization ability of the reasoning LM, yielding higher performance on both in-distribution and out-of-distribution test sets. Also, PINTO leverages the rationales more faithfully than competitive baselines do.

Author Information

Peifeng Wang (University of Southern California)
Peifeng Wang

Peifeng Wang is a PhD student working at USC-ISI with Dr. Pedro Szekely, Xiang Ren, Filip Ilievski, and Muhao Chen. He received his bachelor's and master's degrees from Sun Yat-sen University in China.

Aaron Chan (University of Southern California)
Filip Ilievski (USC/ISI)
Filip Ilievski

Filip Ilievski is Research Lead in the Center on Knowledge Graphs within the Information Sciences Institute (ISI) at the USC Viterbi School of Engineering and Assistant Professor of Computer Science at USC. Filip holds a Ph.D. in Natural Language Processing from the Vrije Universiteit (VU) in Amsterdam, where he also worked as a postdoctoral researcher before joining ISI. His research focuses on developing robust and explainable neuro-symbolic technology with positive real-world impact, based on neural methods and high-quality knowledge. Filip has made extensive contributions in identifying long-tail entities in text, performing robust and explainable commonsense reasoning, and managing large-scale knowledge resources. Over the past three years, he mentored around twenty Master’s and Ph.D. students, and has been collaborating with researchers at USC, CMU, Bosch Research, RPI, University of Amsterdam, and the University of Lyon. Filip has over 40 peer-reviewed publications in top-tier venues on commonsense reasoning, information extraction, and knowledge graphs. He has also been actively organizing workshops (AAAI’21), tutorials (AAAI’21, ISWC’20, ISWC’21, TheWebConf’22, KGC’22), symposiums (USC), and a special journal issue (Semantic Web Journal) on these topics.

Muhao Chen (University of Southern California)
Xiang Ren (University of Southern California)

More from the Same Authors