Skip to content

SandipanMajhi/Recursive-Reasoning

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

5 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Recursive-Reasoning

In this project, I have reimplemented the Maieutic Reasoning Paper(see citation at the end) using Google's FLAN-T5-XL and XXL. Maieutic Reasoning means generating explanations through repeated questioning to finally reach a decision. Reaching the decision is a satisfiability problem. We use RC2 MAXSAT in this to find out the satisfiability for the decisions and reach a binary decision of whether true or false.

Example generated by FLAN-T5-XL for the proposition = (If Becky is allergic to legumes, while Nancy is allergic to tree nuts, Nancy is more likely to eat peanuts than Becky.)

If Becky is allergic to legumes, while Nancy is allergic to tree nuts, Nancy is more likely to eat peanuts than Becky.
├── If one is allergic to tree nuts, one would not eat peanuts.
│   └── Peanuts and tree nuts are different types of tree Nut.
├── In fact, if Nancy is allergic to tree nuts she would only eat peanuts.
│   └── Tree nuts and peanuts are very different nut.
└── Nancy are allergic to tree nuts and Becky is not.
    ├── Nancy can't eat any kind of nut while Becky can be allergic.
    └── Tree nuts are common among the population and one only has to look online in order to see if they are allergic to certain foods.

For generating trees run the following code :

 > python3 main_gen2.py --dataset_name com2sense --mode normal --seed 21

For inference on the generated trees :

  > python3 main_inference2.py --dataset_name com2sense --mode normal --seed 21

Results

Dataset : Com2sense (dev set - 782 examples)

Generator+Verifier Generation-Seed Inference-Seed Type-of-method Max-Acc
FlanT5 + Roberta. 42 Default Method-1 61.5%
FlanT5 + Roberta. 42 Default Method-2 66.8%
FlanT5 + FlanT5. 42 42 Method-1 58.8%
FlanT5 + FlanT5. 42 15 Method-1 60.35%
FlanT5 + FlanT5. 42 42 Method-2 69.43%
FlanT5 + FlanT5. 42 21 Method-2 70.33%
GPT-3 + Roberta (paper) 72.5%

Dataset : CREAK (dev set - 1371 examples)

Generator+Verifier Generation Seed Inference Seed Type of method Max-Acc
FlanT5 + Roberta. 42 Default Method-1 70.02%
FlanT5 + Roberta. 42 Default Method-2 82.4%
FlanT5 + FlanT5. 42 42 Method-1 70.45%
FlanT5 + FlanT5. 42 42 Method-2 83.37%
FlanT5 + FlanT5. 42 21 Method-1 70.45%
FlanT5 + FlanT5. 42 21 Method-2 83.95%
GPT-3 + Roberta (paper) 85.2%

Dataset - StrategyQA (dev set - 229 examples)

Generator+Verifier Seed Value Type of Method Max Acc
Flan T5 + Roberta Verifier 42 Method-1 50.21 %
Flan T5 + Roberta Verifier 21 Method-1 50.21 %
Flan T5 + Roberta Verifier 42 Method - 2 51.96 %
Flan T5 + Roberta Verifier 21 Method - 2 51.09 %
Flan T5 + Flan T5 Verifier 42 Method - 1 49.34 %
Flan T5 + Flan T5 Verifier 21 Method - 1 48.03 %
Flan T5 + Flan T5 Verifier 42 Method - 2 51.52 %
Flan T5 + Flan T5 Verifier 21 Method - 2 51.52 %
GPT-3 (paper) 60.7%

Paper Citation:

@inproceedings{Jung2022MaieuticPL,
  title={Maieutic Prompting: Logically Consistent Reasoning with Recursive Explanations},
  author={Jaehun Jung and Lianhui Qin and Sean Welleck and Faeze Brahman and Chandra Bhagavatula and Ronan Le Bras and Yejin Choi},
  booktitle={Conference on Empirical Methods in Natural Language Processing},
  year={2022}
}

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors