Skip to content

Latest commit

 

History

History
105 lines (74 loc) · 3.47 KB

README.md

File metadata and controls

105 lines (74 loc) · 3.47 KB
Raw

HIRO: Hierarchical Indexing for Retrieval-Augmented Opinion Summarization

Tom Hosking, Hao Tang and Mirella Lapata

Setup

Create a fresh environment:

conda create -n hiroenv python=3.9
conda activate hiroenv

or

python3 -m venv hiroenv
source hiroenv/bin/activate

Then install dependencies:

pip install -r requirements.txt

Download data/models:

Tested with Python 3.9.

Generating Summaries for Space/Amasum

First, run the pre-processing eval recipe:

torchseq-eval --recipe opagg.hiro_pre --model ./models/20240130_183901_d671_space --test

Then, get generations from your preferred LLM (we used Mistral 7B Instruct v0.2), based on the prompts in the files eval/llm_inputs_piecewise_test.jsonl and eval/llm_inputs_oneshot_test.jsonl. You might want to use my TGI Client to run efficient batched inference through a HuggingFace model:

python tgi-client/runner.py --input runs/hiro/space/llm_inputs_oneshot_test.jsonl --output runs/hiro/space/llm_outputs_oneshot_test_mistaral7b.js
onl --model mistralai/Mistral-7B-Instruct-v0.2

Then run the post-LLM eval recipe to get the scores:

torchseq-eval --recipe opagg.hiro_post --model ./models/20240130_183901_d671_space --test

TODO: How to generate summaries for other datasets

Training on Space/Amasum

(Optional) Rebuild the datasets:

Run the dataset filtering scripts ./scripts/opagg_filter_space.py and ./scripts/opagg_filter_space_eval.py

# SPACE
python scripts/generate_opagg_posnegtriples.py --dataset space-25toks-1pronouns  --min_pos_score 0.75 --ignore_neutral --unsorted --min_overlap 0.3

# AmaSum
python scripts/generate_opagg_posnegtriples.py --dataset amasum-electronics-25toks-0pronouns  --min_pos_score 0.75 --ignore_neutral --unsorted --min_overlap 0.3
python scripts/generate_opagg_posnegtriples.py --dataset amasum-shoes-25toks-0pronouns  --min_pos_score 0.75 --ignore_neutral --unsorted --min_overlap 0.3
python scripts/generate_opagg_posnegtriples.py --dataset amasum-sports-outdoors-25toks-0pronouns  --min_pos_score 0.75 --ignore_neutral --unsorted --min_overlap 0.3
python scripts/generate_opagg_posnegtriples.py --dataset amasum-home-kitchen-25toks-0pronouns  --min_pos_score 0.75 --ignore_neutral --unsorted --min_overlap 0.3

Train:

torchseq --train --reload_after_train --validate --config ./configs/hiro_space.json

Training on a new dataset

Setting up to train on a new dataset can be a bit tricky - I'm happy to help run you through the process, just email me or raise an issueon Github.

  • Make a copy of your dataset in a format expected by the preprocessing scripts
  • Clean the training data and eval data
  • Generate training pairs
  • Modify one of the configs to point to your data
  • Train the model
torchseq --train --reload_after_train --validate --config ./configs/{YOUR_CONFIG}.json

Citation

@misc{hosking2024hierarchical,
      title={Hierarchical Indexing for Retrieval-Augmented Opinion Summarization}, 
      author={Tom Hosking and Hao Tang and Mirella Lapata},
      year={2024},
      eprint={2403.00435},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}