Transformers as Soft Reasoners over Language

By Peter Clark, Oyvind Tafjord, Kyle Richardson. 2020

465 citations recorded by the source corpus.

Beginning with McCarthy's Advice Taker (1959), AI has pursued the goal of providing a system with explicit, general knowledge and having the system reason over that knowledge. However, expressing the knowledge in a formal (logical or probabilistic) representation has been a major obstacle to this research. This paper investigates a modern approach to this problem where the facts and rules are provided as natural language sentences, thus bypassing a formal representation. We train transformers to reason (or emulate reasoning) over these sentences using synthetically generated data. Our models, that we call RuleTakers, provide the first empirical demonstration that this kind of soft reasoning over language is learnable, can achieve high (99%) accuracy, and generalizes to test data requiring substantially deeper chaining than seen during training (95%+ scores). We also demonstrate that the models transfer well to two hand-authored rulebases, and to rulebases paraphrased into more natural language. These findings are significant as it suggests a new role for transformers, namely as limited "soft theorem provers" operating over explicit theories in language. This in turn suggests new possibilities for explainability, correctability, and counterfactual reasoning in question-answering.

Subfields: Artificial Intelligence, Natural Language Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Natural Language Processing

Citations in the corpus

Chain of Thought Prompting Elicits Reasoning in Large Language Models (2022, neutral): “Another line of work aims to enable language models to perform logical or formal reasoning, often by verablizing the rules in natural language formal rules using language (Clark et al., 2020; Saeed et al., 2021; Liang et al., 2021).”
Ethical-Advice Taker: Do Language Models Understand Natural Language Interventions? (2021, background): “1https://github.com/allenai/ethical-interventions impossible just a few years ago (Wang et al., 2019; Clark et al., 2020).”
RuleBERT: Teaching Soft Rules to Pre-Trained Language Models (2021, methodology): “Unlike previous approaches (Clark et al., 2020), we enable deductive reasoning for a large class of soft rules with binary predicates and an unrestricted number of variables.”
Explainable Multi-hop Verbal Reasoning Through Internal Monologue (2021, methodology): “We augment GPS with a SOTA sequence-to-sequence (Seq2Seq) model and apply this model re-cursively to achieve high interpretability and better generalization of compositionality for a synthetic QA task (Clark et al., 2020).”
Flexible Generation of Natural Language Deductions (2021, background): “Since then, there has been massive progress in automated formal reasoning (De Moura and Bjørner, 2011); in contrast, progress in automated The Green Bay Packers play in the NFL. NFL teams play from September to January.”
Selection-Inference: Exploiting Large Language Models for Interpretable Logical Reasoning (2022, methodology): “While the model proposed by Tafjord et al. (2020) is very impressive, it only works for “Prove this statement to be True/False” style questions, since it relies on enumerating all possible implications and checking whether the question statement or its negation are present in the inferred facts,…”
ProofWriter: Generating Implications, Proofs, and Abductive Statements over Natural Language (2020, methodology): “Recent work on RuleTaker (Clark et al., 2020) demonstrated a modern approach to this goal, in which transformers emulate deductive reasoning”
Thinking Aloud: Dynamic Context Generation Improves Zero-Shot Reasoning Performance of GPT-2 (2021, methodology): “Against this background, the novelty of our study is to show that GPT-2 has a strong zero-shot performance on a NLI task involving deductive reasoning over rules [Clark et al., 2020].”
PRover: Proof Generation for Interpretable Reasoning over Rules (2020, methodology): “Clark et al. (2020) propose RuleTakers, a ﬁne-tuned RoBERTa model (Liu et al., 2019b) to show that transformers can act as “soft theorem provers” by predicting the ﬁ-nal answer in such reasoning-based problems with high accuracy.”
Explanations for CommonsenseQA: New Dataset and Models (2021, background): “Clark et al. (2020) studied the logical reasoning capacity of transformer based language models on various RCQA tasks.”
COCO-LM: Correcting and Contrasting Text Sequences for Language Model Pretraining (2021, methodology): “Then the language semantics captured during pretraining are conveyed to downstream tasks via the pretrained Transformer parameters [5, 8, 44].”
Logic-LM: Empowering Large Language Models with Symbolic Solvers for Faithful Logical Reasoning (2023, background): “Recent works in adapting large language models for logical reasoning tasks can be broadly categorized into two groups: 1) fine-tuning approaches that optimize LLMs’ reasoning ability through fine-tuning or training specialized modules (Clark et al., 2020; Tafjord et al., 2022; Yang et al., 2022), an”
Leap-Of-Thought: Teaching Pre-Trained Models to Systematically Reason Over Implicit Knowledge (2020, methodology): “Recently [4], it has been shown that Transformers [5] are an effective architecture for this goal, as they can be trained to reason over knowledge expressed as natural language statements.”
Relational World Knowledge Representation in Contextual Language Models: A Review (2021, background): “However, as Clark et al. (2020) note, these results do not necessarily conﬁrm that LMs can “reason,” but rather that they can at least emulate soft reasoning—raising an open question about how to develop probes and metrics to verify whether LMs can actually reason compositionally.”
Knowledge-Aware Language Model Pretraining (2020, background): “Scaling up the transformer parameter sizes will lead to signiﬁcant improvements in the language model capability: with wider and deeper transformer layers, it is observed that trans-former language models start to output more complicated semantics beyond lexical and syntactic patterns [6, 8, 7].”
Exploring Generalization Ability of Pretrained Language Models on Arithmetic and Logical Reasoning (2021, methodology): “Clark et al. (2020) perform Transformer-based models on simple logic reasoning test, and their results show that the model can get quite promising results and the model is not sensitive to the question length.”
Neural Databases (2020, background): “Transformers have been shown to perform soft-reasoning when provided with simple propositional rules [9].”