This repository contains codes for the paper "Disentangled Retrieval and Reasoning for Implicit Question Answering".
Creating an Elasticsearch index of our corpus. Following StrategyQA: https://github.com/eladsegal/strategyqa/tree/main/elasticsearch_index
Our experiments were conducted in a Python 3.7 environment. To clone the repository and set up the environment, please run the following commands:
git clone https://github.com/eladsegal/strategyqa.git
cd strategyqa
pip install -r requirements.txt
The official StrategyQA dataset files with a detailed description of their format can be found on the dataset page.
To train our baseline models, we created a 90%/10% random split of the official train set to get an unofficial train/dev split: data/strategyqa/[train/dev].json.
Download link to our full corpus of Wikipedia paragraphs is available on the dataset page. A script for indexing the paragraphs into Elasticsearch is available here.
python Multi-view QueryGeneration.py
The attribute retriever is built following Sentence-Transformer. The retrieved topic-related documents and the data processing of attribute retriever will be released after acception.
python==3.8
torch==1.9.0
nltk==3.6.8
transformers==4.9.0
The weight model named weights.th of baseline should be in the path ./pretrained_model/6_STAR_ORA-p/, which could be downloaded and unzipped from here.
Run the model with default configuration
python main.py
Configuration can be edited in the file main.py or in the running command line, for example,
python main.py \
--num_workers 1 \
--load_pretrained true \
--epoch_num 20 \
--batch_size 16 \
--max_length 512 \
--reason_train ./data/reason/train_sents.pk \
--reason_dev ./data/reason/dev_sents.pk \
--reason_test ./data/reason/test_sents.pk \
--prediction_path test_predictions.json \
--model_path ./checkpoints/mymodel.th \
--model_class ReasoningPlain
The json files in the path ./classification/ describes several strategies for the definition and classification of operators, which are crucial components in our reasoning. In the paper, we adopt the 5-class strategy, that is, comparison, logical, entail, numerical and binary. To try another classification strategy, change the configuration --op_classification accordingly.