Massive-scale Decoding for Text Generation using Lattices

Jiacheng Xu, Sid Reddy, Greg Durrett

TL;DR: a new search algorithm to construct lattices encoding many generation options; two key technical contributions: (1) best-first search, (2) path recombination.

Visualization

We provide a few examples in the vis folder and on my homepage. You need to download the html files to view and interact with the model outputs.

The complete set of outputs are available on Box.

Getting started

• model contains all of the methods, including baselines like beam search, nucleus sampling, and our methods.
• evaluation contains scripts for evaluation.
• command are the prompts and shells we use to run the experiment.

Beam Search:

PYTHONPATH=./ python src/recom_search/scripts/run_pipeline.py -nexample 100  -ngram_suffix 4  -beam_size 16 -min_len 10 -max_len 35   -model bs 

Best-first Search:

PYTHONPATH=./ python src/recom_search/scripts/run_pipeline.py -nexample 100  -ngram_suffix 4  -beam_size 16 -min_len 10 -max_len 35   -model astar_baseline

Best-first Search with Recomb:

PYTHONPATH=./ python src/recom_search/scripts/run_pipeline.py -nexample 100  -ngram_suffix 4 -beam_size 16 -min_len 10 -max_len 35 -model astar -merge imp  -avg_score 0.75  -dfs_expand 

Best-first Search with Zip:

PYTHONPATH=./ python src/recom_search/scripts/run_pipeline.py -nexample 100  -ngram_suffix 4 -beam_size 16 -min_len 10 -max_len 35 -model astar -merge zip  -avg_score 0.75  -dfs_expand 

More detailed instructions coming soon!

Citation

@misc{xu-etal-2022-massive,
    title={Massive-scale Decoding for Text Generation using Lattices},
    author = {Xu, Jiacheng and Jonnalagadda, Siddhartha Reddy and Durrett, Greg},
    year={2022},
    eprint={2112.07660},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

Contact

[email protected]