[Paper Review] Back to the Future: Unsupervised Backprop-based Decoding for Counterfactual and Abductive Commonsense Reasoning

Lianhui Qin, Vered Shwartz+6 more|arXiv (Cornell University)|Oct 12, 2020

Topic Modeling52 references17 citations

TL;DR

This paper introduces DeLorean, an unsupervised decoding algorithm that enables left-to-right language models to incorporate both past and future context for abductive and counterfactual reasoning by alternating forward and backward passes with gradient-based refinement of hidden representations. It outperforms unsupervised and some supervised methods on abductive text generation and counterfactual story rewriting, achieving state-of-the-art results without fine-tuning or supervision.

ABSTRACT

Abductive and counterfactual reasoning, core abilities of everyday human cognition, require reasoning about what might have happened at time t, while conditioning on multiple contexts from the relative past and future. However, simultaneous incorporation of past and future contexts using generative language models (LMs) can be challenging, as they are trained either to condition only on the past context or to perform narrowly scoped text-infilling. In this paper, we propose DeLorean, a new unsupervised decoding algorithm that can flexibly incorporate both the past and future contexts using only off-the-shelf, left-to-right language models and no supervision. The key intuition of our algorithm is incorporating the future through back-propagation, during which, we only update the internal representation of the output while fixing the model parameters. By alternating between forward and backward propagation, DeLorean can decode the output representation that reflects both the left and right contexts. We demonstrate that our approach is general and applicable to two nonmonotonic reasoning tasks: abductive text generation and counterfactual story revision, where DeLorean outperforms a range of unsupervised and some supervised methods, based on automatic and human evaluation.

Motivation & Objective

• To address the challenge of conditioning language models on both past and future contexts for nonmonotonic reasoning tasks such as abductive and counterfactual reasoning.
• To develop an unsupervised decoding method that leverages only off-the-shelf left-to-right language models without requiring fine-tuning or additional supervision.
• To enable coherent text generation that respects both left and right context constraints, particularly in narrative reasoning where causal coherence and minimal edits are essential.
• To overcome limitations of existing methods that rely on narrow text-infilling, fine-tuning, or supervised attribute discriminators.

Proposed method

• DeLorean alternates between forward (left-to-right) and backward (right-to-left) passes to refine hidden representations of the output sequence.
• During backward passes, gradients are backpropagated from the future context using a task-specific loss, updating only the output representation while keeping model parameters fixed.
• The forward pass maximizes likelihood conditioned on the past context, approximating P(Y|X), while the backward pass enforces coherence with the future, approximating P(Z|XY).
• Forward and backward hidden states are combined into a single context vector from which tokens are sampled during decoding.
• The method operates entirely within a pre-trained language model framework, requiring no fine-tuning or additional training.
• It is designed to be general and applicable to any conditional generation task requiring integration of left and right context.

Experimental results

Research questions

• RQ1 Can an unsupervised decoding method effectively incorporate both past and future context in language model generation without fine-tuning?
• RQ2 How well can a backpropagation-based refinement mechanism improve coherence in abductive and counterfactual reasoning tasks?
• RQ3 Can a method relying only on left-to-right language models achieve performance competitive with supervised baselines on nonmonotonic reasoning tasks?
• RQ4 To what extent does the alternating forward-backward pass mechanism improve factual consistency and minimal edit behavior in counterfactual story rewriting?

Key findings

• DeLorean outperforms a range of unsupervised and some supervised baselines on both abductive text generation and counterfactual story rewriting, according to automatic and human evaluations.
• On the TimeTravel dataset for counterfactual story rewriting, DeLorean successfully rewrites story endings with minimal edits, preserving coherent elements like 'She looked forward to wearing it' when the counterfactual condition changes from 'went to mall' to 'ordered a shirt online'.
• In abductive reasoning, DeLorean generates plausible intermediate events (e.g., 'she got hurt by the swing') that bridge observed past and future states, demonstrating strong causal inference capability.
• The method achieves state-of-the-art results without any fine-tuning or supervision, relying solely on off-the-shelf left-to-right language models.
• Human evaluation confirms that DeLorean outputs are more coherent and plausible than those of strong unsupervised baselines, particularly in maintaining narrative consistency under counterfactual conditions.