Twpipe is a pipeline toolkit that parses raw tweets into universal dependencies. Twpipe contains the following components:

• a joint sentence sentence segmentor and tokenizer;
• a POS tagger;
• a transition-based parser (with parser ensemble and distillation supports).

For technique details, please refer our NAACL 2018 paper (for pipeline without sentence segmentor) and ACL 2018 paper (for parser ensemble and distillation).

• GCC (version greater than 5): json.hpp requires syntax that is not supports by GCC 4.X.
• MSVC 18.x: I've successfully compiled twpipe on MSVC 18. If you got any problem, please let me know.
• boost: we use its program_options.

In the project root directory, run the following commands

git submodule init
git submodule update
mkdir build
cmake .. -DEIGEN3_INCLUDE_DIR=/path/to/your/eigen3/

First unzip the released model with

bunzip2 -k model/en_ewt_en_tweebank_train.model.json.bz2

Then go back to the root directory. Suppose your input tweets are stored in the `input_file' with each tweet in one line. Run the following commands:

./bin/twpipe --segment-and-tokenize --postag --parse \
    --model model/en_ewt_en_tweebank_train.model.json input_file

The conllu-formatted output dumped to stdout.

1. The postagger we shipped in twpipe is a naive bidirectional LSTM sequence tagger which performs worse than that of Owoputi et al. (2015). We suggest using theirs instead.
2. We found that also doing sentence segmentation leads to better parsing performance.
3. Specifying word embeddings with --embedding ./data/glove.twitter.27B.100d.txt will lead better performance.

./bin/twpipe \
    --dynet-seed 1 \
    --train \
    --heldout ./data/en-ud-tweebank-train.conllu \
    --train-parser true \
    --train-postagger true \
    --train-segmentor-and-tokenizer true \
    --optimizer adam \
    --optimizer-enable-clipping true \
    --model model.twpipe \
    --max-iter 10 \
    ./data/en-ud-tweebank-train.conllu

We found the transition-based parser training is sensitive to initialization. To this remedy, we can train several parsers, ensemble them, then doing knowledge distillation.

To generate training data for the distillation models, you need first generating a set of baseline parsers (by varying the random seed with --dynet-seed). Then generate the ensemble of the parsers as:

./bin/generate_parse_ensemble_data \
    --models ./model1.twpipe,./model2.twpipe./model3.twpipe \
    ./data/en-ud-tweebank-train.conllu > en-ud-tweebank-train.actions

where different models are separated by comma. It will dump a json formatted data into stdout. And you can learn a single parser from this data using the following commands.

./bin/twpipe \
    --dynet-seed 1 \
    --train \
    --heldout ./data/en-ud-tweebank-train.conllu \
    --train-distill-parser true \
    --parse-ensemble-data ./en-ud-tweebank-train.actions \
    --optimizer adam \
    --optimizer-enable-clipping true \
    --model model.twpipe \
    --max-iter 10 \
    ./data/en-ud-tweebank-train.conllu

Our results for the NAACL 2018 paper was obtained by concatenating the en-ud-ewt-train.conllu and en-ud-tweebank-train.conllu. Try it please!

• Shipping Owoputi's Tagger: as pointed in our paper, Owoputi's Tagger is the SOTA twitter parser. We would like to ship their tagger within twpipe in the future.

• Olutobi Owoputi, Brendan O’Connor, Chris Dyer, Kevin Gimpel, Nathan Schneider, and Noah A. Smith. 2013. Improved part-of-speech tagging for online conversational text with word clusters. In Proc. of NAACL.
• Yijia Liu, Yi Zhu, Wanxiang Che, Bing Qin, Nathan Schneider, and Noah A. Smith. 2018. Parsing Tweets into Universal Dependency. In Proc. of NAACL.
• Yijia Liu, Wanxiang Che, Huaipeng Zhao, Bing Qin, and Ting Liu. 2018. Distilling Knowledge for Search-based Structured Prediction. In Proc. of ACL.

Please see the LICENSE.md file.