Back to the Future: Unsupervised Backprop-based Decoding for Counterfactual and Abductive Commonsense Reasoning (2010.05906v4)

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 LLMs (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 LLMs 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.

An Overview of DELOREAN: Unsupervised Backprop-based Decoding for Nonmonotonic Reasoning

The paper "Back to the Future: Unsupervised Backprop-based Decoding for Counterfactual and Abductive Commonsense Reasoning" addresses the challenge of incorporating both past and future contexts in nonmonotonic reasoning tasks using generative LLMs (LMs). It introduces DELOREAN (DEcoding for nonmonotonic LOgical REAsoNing), an unsupervised decoding algorithm that leverages backpropagation to incorporate future constraints into generation processes normally limited to past context conditioning.

Methodological Advances

DELOREAN uses off-the-shelf, unidirectional LMs to handle the reasoning tasks abductive text generation and counterfactual story rewriting. The method involves alternating between forward and backward propagations, without altering the LM parameters. The forward pass utilizes a standard left-to-right generation, while the backward pass integrates future constraints through backpropagation. This strategic alternation allows for a generated output that is coherent with both past and future contexts. The resultant output vector, derived from mixing forward and backward outputs, is ranked using an unsupervised BERT-based coherence measure. This ranking system employs a next-sentence prediction task to enhance overall coherence.

Empirical Validation

The DELOREAN approach was rigorously evaluated against unsupervised and some supervised baselines on two tasks: abductive reasoning, using the ART dataset, and counterfactual reasoning, using the TIMETRAVEL dataset. On both tasks, DELOREAN surpassed all unsupervised counterparts and even outperformed some supervised models, based on both automatic and human evaluations. For instance, in abductive reasoning, DELOREAN showcased enhanced performance on BLEU-4, ROUGE-L, and BERTScore metrics compared to its peers. Human evaluators also found that DELOREAN's outputs were more coherent with the specified contexts than those from supervised models.

Implications and Future Prospects

The capability demonstrated by DELOREAN to enhance nonmonotonic reasoning using unsupervised learning offers significant promise for future AI applications. It demonstrates a departure from traditional reliance on vast amounts of training data and fine-tuning for every specific task. Instead, by manipulating representations through unsupervised methods, DELOREAN opens pathways to broader applicability and efficiency in handling diverse narrative reasoning scenarios.

Furthermore, this approach suggests potential adaptations for other reasoning tasks where maintaining coherence across context shifts is crucial. The backpropagation-based method might extend to applications beyond the generative domain, such as improving coherence in dialogue systems or facilitating improved context-dependent translation systems.

Conclusion

The introduction and success of DELOREAN in abductive and counterfactual reasoning affirm the feasibility of nonmonotonic reasoning in unsupervised settings. The innovative backpropagation-based decoding methodology not only provides a powerful tool for current nonmonotonic reasoning challenges but also lays the groundwork for broader explorations within AI, warranting continued investigation and experimentation in unsupervised reasoning techniques.