whatsthatbook

This repository contains code for the paper Decomposing Complex Queries for Tip-of-the-tongue Retrieval

Data

Dataset is released on huggingface at: https://huggingface.co/datasets/nlpkevinl/whatsthatbook

Installation

Create conda environment

conda create -n whatsthatbook python=3.9 --yes
conda activate whatsthatbook

Install the requirements.txt file

pip install requirements.txt

Clue Generation

To generate subqueries, input a prompt, examples, and the original query file. Our example uses openai as the LLM decomposer, so first set the required environmental variables OPENAI_API_KEY.

Example query for cover prompts.

python clue_extraction.py --prompt_text_file prompts/cover.prompt.txt --examples_file prompts/examples/cover_clue_examples.jsonl  --input_file ./data/2022-05-30_14441_gold_posts.cover_clues.jsonl --output_file ./data/debug.2022-05-30_14441_gold_posts.cover_clues.jsonl  --max_examples 1

Text Model Finetuning

To finetune dense models on the text-based metadata, run:

  python finetuning.py --model_path bert-base-uncased \
  --eval_data <path to dev> \
  --train_data  <path to train> \
  --output_dir <path to output> \
  --total_steps 10000 \
  --save_freq 5000 \
  --per_gpu_batch_size 16

To evaluate, trained models models build indices following /baselines/contriever/README.md then run passage_retrieval_all.py to generate output files and metrics.

References

Please consider referencing our work if you find it useful for your work

@misc{lin-etal:2023:arxiv,
  author    = {Kevin Lin and Kyle Lo and Joseph Gonzalez and Dan Klein},
  title     = {Decomposing Complex Queries for Tip-of-the-tongue Retrieval},
  note      = {arXiv:2305.15053},
  year      = {2023}
}