ericxsun / emnlp2016

LSTM Parse Reranker

Overview

We release the code and models used in our paper:

• Do Kook Choe and Eugene Charniak. "Parsing as Language Modeling." Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2016.

Prerequisites

bllipparser

tensorflow

Data Preprocessing

Run the following commands to preprocess the WSJ data:

git clone https://github.com/cdg720/emnlp2016.git && cd emnlp2016
./prepare.sh wsj-train wsj-dev

wsj-train and wsj-dev should have one tree per line:

Tree 1
Tree 2
...

If you have a license for the 5th Gigaword and want the tri-training data we use in the paper, send me an email with the license at [email protected]. To preprocess the tri-training data, change lines 38 and 48 in prepare.sh appropriately and run prepare.sh.

The code is based on ptb_word_lm.py and reader.py of the tensorflow RNN tutorial.

Training

mkdir -p models/wsj && python train.py --data_path=wsj --model_path=models/wsj/model

mkdir -p models/semi && python tri_train.py --data_path=semi --model_path=models/semi/model

Due to stochasticity, different runs of training produce models with slightly different results from what we report in the paper. We have retrained a few models after the submission of our paper and we list their F1s and numbers of training epochs here:

wsj	Paper	1	2	3	4	5
F1 (dev)	91.62	91.72	91.55	91.59	91.49	91.50
# epochs	37	47	38	41	31	49

semi	Paper	1	2	3	4	5
F1 (dev)	92.46	92.33	92.52	92.45	92.42	92.42
# epochs	26	13	21	31	21	26

Note that the F1s reported here are about 0.05-0.1 lower than they actually are. Between training epochs, we evaluate models with batch size 200, which allows faster but less accurate evaluation. Evaluating with batch size 10 returns correct F1s.

Reranking

python rerank.py --data_path=wsj --model_path=models/wsj/model --nbest_path=nbest

nbest has the following format:

n1
Tree 1.1
Tree 1.2
...
Tree 1.n1

n2
Tree 2.1
Tree 2.2
...
Tree 2.n2

...

Models

wsj model

semi-supervised model

The models we use in our paper are not compatible with tensorflow 0.11.0rc0 and we provide new models retraiend with the code in this repository.