Sports Reporter

Python code for Learning to Select, Track, and Generate for Data-to-Text (Iso et al; ACL 2019).

Resources

Rotowire-modified dataset

Please refer to rotowire-modified repo.

Usage

Dependencies

• The code was written for Python 3.X and requires DyNet.
• Dependencies can be installed using requirements.txt.
• For running information extractor, you should install torch.

Preprocessing

Before starting an experiment, you should run our provided setup.sh.

./setup.sh

After that, you can make the annotation file for training data via information extractor:

cd ./data2text-1
cat ../rotowire_v2/train.json | python -c 'import sys, json, nltk; print("\n".join(" ".join(nltk.word_tokenize(" ".join(x["summary"]))) for x in json.load(sys.stdin)))' > ../rotowire_v2/train_summary.txt
python data_utils.py -mode prep_gen_data -gen_fi ../rotowire_v2/train_summary.txt -dict_pfx "rotowire-modified-ie" -output_fi train_gold.h5 -input_path "../rotowire_v2" -train
th extractor.lua -gpuid 1 -datafile rotowire-modified-ie.h5 -preddata train_gold.h5 -dict_pfx "rotowire-modified-ie" -just_eval

Then, you can see the annotation file train_gold.h5-tuples.txt and make a vocab file for training.

cd ..
VOCAB=<path to the vocablary file>
python make_data.py ./rotowire_v2 ./data2text-1/train_gold.h5-tuples.txt $VOCAB

Train model

python reporter.py train $VOCAB --valid_file ./rotowire_v2/valid.json

Decode

MODEL=<path to the trained model file>
python reporter.py decode $VOCAB $MODEL ./rotowire_v2/test.json

Updated Results for RotoWire-modified

without writer info	RG (P% / #)	CS (P% / R%)	CO	BLEU
Joint+Rec+TVD (B=5)	18.09 / 48.54	23.24 / 28/92	14.47	15.34
Conditional (B=5)	20.28 / 61.76	27.20 / 29.76	15.88	15.26
Puduppully+, AAAI'19	82.55 / 34.05	32.30 / 43.74	16.67	14.82
Puduppully+, ACL'19	91.13 / 32.41	37.05 / 43.06	20.62	15.23
Iso+, ACL'19	91.98 / 31.66	40.44 / 46.63	21.56	15.74

with writer info	RG (P% / #)	CS (P% / R%)	CO	BLEU
Puduppully+, AAAI'19	82.55 / 34.05	32.30 / 43.74	16.67	14.82
+ stage 1	85.54 / 30.26	42.33 / 49.38	21.26	18.01
+ stage 2	83.35 / 32.42	33.28 / 42.92	16.73	16.57
+ stage 1 & 2	84.09 / 28.16	43.63 / 47.75	21.96	18.57
Iso+, ACL'19	91.98 / 31.66	40.44 / 46.63	21.56	15.74
+ writer	93.32 / 29.44	51.76 / 55.21	24.97	20.62

License and References

This code is available under the MIT Licence, see LICENCE

When you write a paper using this code, please cite the followings.

@InProceedings{Iso2019Learning,
    author = {Iso, Hayate
              and Uehara, Yui
              and Ishigaki, Tatsuya
              and Noji, Hiroshi
              and Aramaki, Eiji
              and Kobayashi, Ichiro
              and Miyao, Yusuke
              and Okazaki, Naoaki
              and Takamura, Hiroya},
    title = {Learning to Select, Track, and Generate for Data-to-Text},
    booktitle = {Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL)},
    year = {2019}
  }

Author

@isomap