A semantic parser for commands from the RoboCup@Home General Purpose Service Robot task.

• Utterance to λ-calculus representation parser
• Lexer/parser for loading the released command generation CFG
• Tools for generating commands along with a λ-calculus representation
• Crowd-sourcing interface for collecting paraphrases

If you use this code or data, consider citing our paper Neural Semantic Parsing for Command Understanding in General-Purpose Service Robots. The data collected for this paper is available separately.

Set up a virtual environment using at least Python 3.6:

python3.7 -m virtualenv venv
source venv/bin/activate
pip install -r requirements.txt

The latest grammar and knowledgebase files (pulled from the generator) are provided in the resources directory. The grammar format specification will clarify how to interpret the files.

To produce the dataset, see data/make_dataset.py.

We base our training on previous work using AllenNLP for seq2seq semantic parser training. All of our experiments are declaratively specified in the experiments directory.

You can run them with

allennlp train \
experiments/seq2seq.json \
-s results/seq2seq \
--include-package gpsr_command_understanding

You can monitor training with Tensorboard, just point it at the log directory.

The train_all_models script will train every config back to back.

./scripts/train_all_models gen_demo experiments -t data/gen/train.txt -v data/gen/val.txt

To see a model's output on a data file, use the predict command

allennlp predict --archive-path results/ --include-package gpsr_command_understanding

You can poke at a trained model through the browser using AllenNLP as well

python -m gpsr_command_understanding.demo.logging_server \
    --archive-path results/seq2seq/model.tar.gz \
    --predictor  command_parser\
    --include-package gpsr_command_understanding