AutoTemp / Shallow-Aggressive-Decoding

Codes for the paper "Instantaneous Grammatical Error Correction with Shallow Aggressive Decoding" (ACL-IJCNLP 2021)

Shallow-Aggressive-Decoding

Codes for the paper "Instantaneous Grammatical Error Correction with Shallow Aggressive Decoding" (ACL-IJCNLP 2021)

Results

The performance and online inference efficiency evaluation of baseline and our approach in CoNLL-14.

Model	P	R	F0.5	Speedup
Transformer-big (beam=5)	73.0	38.1	61.6	1.0x
Our approach (9+3)	73.3	41.3	63.5	10.3x
Our approach (12+2 BART-Init)	71.0	52.8	66.4	9.6x

For reference, the beam=1 and beam=5 results of the state-of-the-art 12+2 (BART-Init) are:

12+2 BART-Init	CoNLL-14			BEA-19
Beam	P	R	F0.5	P	R	F0.5
1	71.0	52.8	66.4	74.7	66.4	72.9
5	71.4	52.8	66.7	75.8	66.3	73.7

The above models are all single models without ensemble.

Installation

conda create -n SAD python=3.6
conda activate SAD
conda install pytorch==1.5.1 torchvision==0.6.1 cudatoolkit=10.2 -c pytorch
cd fairseq
pip install --editable .

Usage

This section explains how to decode in different ways.

PTPATH=/to/path/checkpoint*.pt # path to model file
BINDIR=/to/path/bin_data # directory containing src and tgt dictionaries  
INPPATH=/to/path/conll*.bpe.txt # path to eval file
OUTPATH=/to/path/conll*.out.txt # path to output file
BATCH=xxx
BEAM=xxx

Directly use fairseq's interactive.py to decode:

bash interactive.sh $PTPATH $BATCH $BEAM $INPPATH $BINDIR $OUTPATH

Simplified greedy decoding:

python inference.py --checkpoint-path $PTPATH --bin-data $BINDIR --input-path $INPPATH --output-path $OUTPATH --baseline 

Aggressive decoding:

python inference.py --checkpoint-path $PTPATH --bin-data $BINDIR --input-path $INPPATH --output-path $OUTPATH --aggressive