This is a project for unsupervised detection of semantic slots in natural language sentences. The code can be found in the induction/ directory. It requires python 3.6 with installed libraries listed in requirements.txt.

The method can be run in several stages.

1. Data preparation:

We prepared the code to work with three datasets:

CamRest676 https://github.com/WING-NUS/sequicity/tree/master/data/CamRest676
MultiWOZ https://github.com/budzianowski/multiwoz/tree/master/data
CamrestSLU https://aspace.repository.cam.ac.uk/handle/1810/248271

Our pipeline works with preprocessed data that are stored in the python pickle format. There are 3 data specific readers in induction/dataset.py file. To load the data and do initial preprocessing run:

python -m induction.syntactic_parse_data --data_fn "dowloaded_dataset_root" --domain [camrest|carslu|woz-hotel] --data_type raw --srl_parse --dep_parse --output "stored_dataset.pickle"

2. Frame semantic parsing

We use two frame semantic parsers SEMAFOR (http://www.cs.cmu.edu/~ark/SEMAFOR/) and Open-sesame(https://github.com/swabhs/open-sesame). To obtain the individual utterances run:

python -m induction.data_overview --data_fn "stored_dataset.pickle" --data_type pickle --output_type user

Then the utterances must be parsed using the respective parser on the side. To augment the data with semantic parses, run:

python -m induction.add_semantic_parse --data_fn "stored_dataset.pickle" --domain [camrest|carslu|woz-hotel] --data_type pickle --output "stored_dataset.pickle" --semantic_parse "parser_output_file" --type [semafor|sesame]

3. Slot candidates selection

Now we can run the main pipeline selection process:

python -m induction.cluster_utterances --data_fn "stored_dataset.pickle" --data_type pickle --domain [camrest|carslu|woz-hotel] --embedding_file data/fasttext_embs.pkl --corpus "resulted_annotated_corpus.pickle" --work_dir "experiment_directory"

This step creates the slot candidates and stores them as annotated corpus object in the "experiment_directory".

4. Model training and prediction

python -m induction.ner --exp_dir "experiment_directory" --data_dir "experiment_directory/data" --dataset_path "stored_dataset.pickle" --model_dir "directory_to_save_model" --type cnn --threshold THRESHOLD --hidden_size HIDDEN --dropout DROPOUT --train_size SIZE --corpus "experiment_directory/corpus-final.pkl" --domain [camrest|carslu|woz-hotel] --result_file "result_file" [--predict]

5. Evaluation

echo "Average precision"
python -m induction.evaluators --domain [camrest|carslu|woz-hotel] --corpus "experiment_directory/corpus-final.pkl" --result_file "result_file"
echo "Clustering evaluation"
python -m induction.evaluate-merging.py "experiment_directory/corpus-final.pkl" [camrest|carslu|woz-hotel]

Running the example model

We provide a pretrained model compressed in the file example-model.zip. To try out, unzip the zip file and run