Introduction

This repository contains code for the paper Few-Shot Adaptation for Parsing Contextual Utterances with LLMs.

Data

The data is under the directory: release_data. The non-contextual utterance files with *-find_event.*.jsonl are the data used to train and evaluate the model before finetuning with contextual utterances. For contextual utterances, the files that are used are *-find_event_revise.*.proportional_split.jsonl for all paradigms in the paper except for the Parse-Then-Resolve paradigm which use *-find_event_revise.*.proportional_split.edit_fragment_plan..jsonl. Each line in the train/validation files contains one example.

Environment Setup

All experiments are run with a modified version of the codebase from the BenchCLAMP codebase under semantic_parsing_with_constrained_lm/. To set up the environement, follow the instructions under semantic_parsing_with_constrained_lm/README.md.

Grammar Files

To generate the grammar files, first convert the files into the original SMCalFlow format by running scripts/convert_to_smcalflow_format.py. An example of the arguments to apss int are in scripts/convert_to_smcalflow_format.sh.

With the original SMCalFlow formatted files, we can now generate the grammar files for the contextual utterances. To generate python src/semantic_parsing_with_constrained_lm/domains/lispress_v2/create_benchclamp_data.py in the semantic_parsing_with_constrained_lm directory.

Fine-Tuning Experiments

For the finetuning experiments, we first train the LLMs on the non-contextual utterances. The exp-name-pattern argument controls the model, data, input format, and learning rate for a training run. As an example, to produc e the base model trained on the non-contextual utterances, run the following command:

python -m semantic_parsing_with_constrained_lm.finetune.lm_finetune \
--config-name semantic_parsing_with_constrained_lm.configs.benchclamp_config \
--exp-name-pattern 't5-base-lm-adapt_calflowfindevent_no_context_all_0.0001'

Then, to evaluate the model run, model-loc passing in the model from the previous step.

python -m semantic_parsing_with_constrained_lm.run_exp \
        --config-name semantic_parsing_with_constrained_lm.configs.benchclamp_config \
        --exp-names 't5-base-lm-adapt_calflowfindevent_no_context_all_0.0001_10000_dev_eval' \
        --model-loc '/trained_models/1.0/t5-base-lm-adapt_calflowfindevent_no_context_all_0.0001/checkpoint-1000
0'

To finetune the model on the contextual utterances, change the exp-name-pattern to vary what data to train on, and what context is presented to the model. For example, to finetune the model with the Parse-With-Reference-P rogram paradigm on the contextual utterances, run the following command:

python -m semantic_parsing_with_constrained_lm.finetune.lm_finetune \
--config-name semantic_parsing_with_constrained_lm.configs.benchclamp_config \
--exp-name-pattern 't5-base-lm-findevent_calflowfindeventrevise_last_plan_low_0_0.0001'

The reproduce all other paradigms in the fine-tuning section of Table 1, change the last_plan to last_utteran ce for the Parse-With-Last-Utterance-History, and rewritten_utterance for the Rewrite-Then-Parse approarc h. For the Parse-Then-Resolve paradigm, change the data from calflowfindeventreviseeditfragment. Evaluation is done in the same way as the non-contextual utterances.

In-context Learning Experiments

To run the in-context learning experiments in Table 1, first set the environemtal variables by OPENAI_GPT3_ENGI NE to the model name in the OpenAI API. Our experiments use text-davinci-003. Also set SM_OPENAI_API_KEY to the OpenAI API key. Note: SM_OPENAI_KEY is an internal engine. Then, run the semantic_parsing_with_constrained _lm.run_exp with the the GPT3 config. For example, to run the Parse-With-Reference-Program paradigm, run the following command:

python -m semantic_parsing_with_constrained_lm.run_exp \
    --config-name semantic_parsing_with_constrained_lm.configs.benchclamp_gpt3_config \
    --exp-name-pattern 'text-davinci-003_calflowfindeventrevise_last_plan_low_0_2_dev_eval_constrained_bs_5'

Binary Classifier

To get the results for Table 2, we additionally need to train a binary classifer for deciding when to run the co ntextual parsing model vs non-contextual parsing model. To train finetune the model used in the paper, run scr ipts/finetune_revision_classifier.py finetunesroberta-base on examples a balanced set of examples from the r elease_data contextual and non-contextual utterance.

Contributing

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.

When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact [email protected] with any additional questions or comments.

Trademarks

This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.