Braid: Weaving Symbolic and Neural Knowledge into Coherent Logical Explanations

Authors: Aditya Kalyanpur, Tom Breloff, David A Ferrucci10867-10874

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we evaluate Braid on the ROC Story Cloze test and achieve close to stateof-the-art results while providing frame-based explanations.
Researcher Affiliation Industry Aditya Kalyanpur, Tom Breloff, David A Ferrucci Elemental Cognition Inc. {adityak, tomb, davef} @ec.ai
Pseudocode Yes Detailed pseudo-code of the proof-graph builder algorithms are in the supplementary material (Appendix).
Open Source Code No The paper does not provide a direct link to the source code for Braid or explicitly state that it is being released.
Open Datasets Yes We evaluate Braid on the ROC Story Cloze test and achieve close to stateof-the-art results while providing frame-based explanations.
Dataset Splits Yes For our experiments, we focus on the Spring 2016 ROC dataset, which has a validation set and a test set of 1871 examples each. ... We used a 90/10 split for train/dev and found that the classifier accuracy to be 87.5% on the dev set.
Hardware Specification No All DRG models were trained on a single GPU with an effective batch size of 64, a learning rate of 1e-3 and warmup=100 steps. (This only mentions "a single GPU" without specifying the model or other hardware details).
Software Dependencies No The paper mentions software like "T5-base" and "BERT" but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes All DRG models were trained on a single GPU with an effective batch size of 64, a learning rate of 1e-3 and warmup=100 steps. Models for frame-detection (step 1) were trained for 100 epochs, while models for ending-prediction (step 2) were trained for 10 epochs.