Transformers as Soft Reasoners over Language

Authors: Peter Clark, Oyvind Tafjord, Kyle Richardson

IJCAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We train transformers to reason (or emulate reasoning) over these sentences using synthetically generated data. Our models, that we call Rule Takers, 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).
Researcher Affiliation Industry Peter Clark , Oyvind Tafjord and Kyle Richardson Allen Institute for AI, Seattle, WA {peterc,oyvindt,kyler}@allenai.org
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes A live demo and all our datasets are available at https://allenai.org/data/ruletaker
Open Datasets Yes Each dataset contains 100k examples (25k of each Type without/with negation)... A live demo and all our datasets are available at https://allenai.org/data/ruletaker
Dataset Splits Yes Data is randomly split 70/10/20 into train/dev/test partitions, ensuring no overlap of theories between each partition.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory, or cloud instance types) used for running the experiments.
Software Dependencies No The paper mentions models like RoBERTa and BERT, but does not specify software versions for libraries, frameworks, or programming languages (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup No We use fixed hyperparameters (learning rate etc), inheriting the settings from Ro BERTa on RACE [Liu et al., 2019].