This repository contains the code for the paper: "Large-Scale QA-SRL Parsing" by Nicholas FitzGerald, Julian Michael, Luheng He and Luke Zettlemoyer.

Data can be found at: http://qasrl.org

This library requires AllenNLP.

In order to use the models below, you must also:

• Download the pretrained GloVe embeddings and copy the file to data/glove/glove.6B.100d.txt.gz in the repository directory.
• Download the pretrained ElMO word representations and copy the files to data/elmo/elmo_2x4096_512_2048cnn_2xhighway_options.json and data/elmo/elmo_2x4096_512_2048cnn_2xhighway_weights.hdf5

Clone this repository and make sure you have installed prerequisites as listed above.

Download and unpack the pretrained model in the data folder from the command line :

cd scripts
./download_pretrained.sh

Alternatively, you can download the compressed archive manually from here and uncompress it in the data/qasrl_parser_elmo folder.

This model uses 8 LSTM-layers for both span detection and question generation, and a 4-layer LSTM generator. It also uses ELMo deep contextualized word representations. This model achieves 85.0% F1 span detection (with t=0.5) and 47.6% question prediction accuracy on the dense dev set (compared to 81.3 and 47.2 for the model reported in our paper).

To run predictions with this model on new text, prepare a JSON-lines document with one line for each sentence with the following format:

{"sentence": "John went to the store."}
{"sentence": "The man ate the burrito and threw the trash in the garbage"}

From the root of the repository, run prediction with the following command:

python -m allennlp.run predict ./data/qasrl_parser_elmo {$INPUT_FILE} --include-package nrl --predictor qasrl_parser --output-file {$OUTPUT_FILE}

Which will produce the following output:

{"words": ["John", "went", "to", "the", "store"], "verbs": [{"verb": "went", "qa_pairs": [{"question": "Who went somewhere?", "spans": ["John"]}, {"question": "Where did someone go?", "spans": ["to the store"]}], "index": 1}]}
{"words": ["The", "man", "ate", "the", "burrito", "and", "threw", "the", "trash", "in", "the", "garbage", "bin", "."], "verbs": [{"verb": "ate", "qa_pairs": [{"question": "Who ate something?", "spans": ["The man"]}, {"question": "What did someone eat?", "spans": ["the burrito"]}], "index": 2}, {"verb": "threw", "qa_pairs": [{"question": "Who threw something?", "spans": ["The man"]}, {"question": "What did someone throw?", "spans": ["the trash"]}, {"question": "Where did someone throw something?", "spans": ["in the garbage bin", "the garbage bin"]}], "index": 6}]}

Training the QA-SRL parser consists of three stages. The Span Detection and Question Prediction models are trained seperately. Then, a script is run in order to combine these two models into one.

To train the Span Detection models, run:

python -m allennlp.run train {$CONFIG_FILE} --include-package nrl -s {$SAVE_DIRECTORY}

Two config files are included with this repository. These must be modified in order to point the *_data_path fields to your data directory.

1. configs/train_span.json - which reproduces the Span-based model from our paper.
2. configs/train_span_elmo.json - which is the same model but includes ELMo word representations.

Training the Question Detector uses the same command as above but with these config files:

1. configs/train_question.json - which reproduces the Sequence model from our paper.
2. configs/train_question_elmo - which, like above, includes the ELMo word representation.

To combine these two trained models into one model which can then be run for prediction, run the following script:

python scripts/combine_models.py --span {$PATH_TO_SPAN_MODEL_DIRECTORY} --ques {$PATH_TO_QUESTION_MODEL_DIRECTORY} --out {$OUTPUT_TAR_GZ_FILE}

To evaluate on the densely labeled Dev set described in our paper, run the command:

env CUDA_VISIBLE_DEVICES=1 python -m allennlp.run evaluate {$MODEL_DIRECTORY}/model.tar.gz --evaluation-data-file {$DATA_DIRECTORY}/dense/dev.jsonl.gz --cuda-device 0 --include-package nrl --overrides '{"dataset_reader": {"question_sources": "turk", "min_answers": 6, "min_valid_answers": 5}}'