genienlp 0.7.0a3

No project description provided

Genie NLP library

This library contains the NLP models for the Genie toolkit for virtual assistants. It is derived from the decaNLP library by Salesforce, but has diverged significantly.

The library is suitable for all NLP tasks that can be framed as Contextual Question Answering, that is, with 3 inputs:

• text or structured input as context
• text input as question
• text or structured output as answer

As the work by McCann et al. shows, many NLP tasks can be framed in this way. Genie primarily uses the library for semantic parsing, paraphrasing, translation, and dialogue state tracking, and this is what the models work best for.

Installation

genienlp is available on PyPi. You can install it with:

pip3 install genienlp

After installation, genienlp command becomes available.

Usage

Training a semantic parser

The general form is:

genienlp train --tasks almond --train_iterations 50000 --data <datadir> --save <model_dir> <flags>

The <datadir> should contain a single folder called "almond" (the name of the task). That folder should contain the files "train.tsv" and "eval.tsv" for train and dev set respectively.

To train a BERT-LSTM (or other MLM-based model) use:

genienlp train --tasks almond --train_iterations 50000 --data <datadir> --save <model_dir> \
  --model TransformerLSTM --pretrained_model bert-base-cased --trainable_decoder_embedding 50

To train a BART or other Seq2Seq model, use:

genienlp train --tasks almond --train_iterations 50000 --data <datadir> --save <model_dir> \
  --model TransformerSeq2Seq --pretrained_model facebook/bart-large --gradient_accumulation_steps 20

The default batch sizes are tuned for training on a single V100 GPU. Use --train_batch_tokens and --val_batch_size to control the batch sizes. See genienlp train --help for the full list of options.

NOTE: the BERT-LSTM model used by the current version of the library is not comparable with the one used in our published paper (cited below), because the input preprocessing is different. If you wish to compare with published results you should use genienlp <= 0.5.0.

Inference on a semantic parser

In batch mode:

genienlp predict --tasks almond --data <datadir> --path <model_dir> --eval_dir <output>

The <datadir> should contain a single folder called "almond" (the name of the task). That folder should contain the files "train.tsv" and "eval.tsv" for train and dev set respectively. The result of batch prediction will be saved in <output>/almond/valid.tsv, as a TSV file containing ID and prediction.

In interactive mode:

genienlp server --path <model_dir>

Opens a TCP server that listens to requests, formatted as JSON objects containing id (the ID of the request), task (the name of the task), context and question. The server writes out JSON objects containing id and answer. The server listens to port 8401 by default, use --port to specify a different port or --stdin to use standard input/output instead of TCP.

Calibrating a trained model

Calibrate the confidence scores of a trained model:

1. Calcualate and save confidence features of the evaluation set in a pickle file:

   genienlp predict --task almond --data <datadir> --path <model_dir> --save_confidence_features --confidence_feature_path <confidence_feature_file>
   

2. Train a boosted tree to map confidence features to a score between 0 and 1:

   genienlp calibrate --confidence_path <confidence_feature_file> --save <calibrator_directory> --name_prefix <calibrator_name>
   

3. Now if you provide --calibrator_paths during prediction, it will output confidence scores for each output:

   genienlp predict --tasks almond --data <datadir> --path <model_dir> --calibrator_paths <calibrator_directory>/<calibrator_name>.calib
   

Paraphrasing

Train a paraphrasing model:

genienlp train-paraphrase --train_data_file <train_data_file> --eval_data_file <dev_data_file> --output_dir <model_dir> --model_type gpt2 --do_train --do_eval --evaluate_during_training --logging_steps 1000 --save_steps 1000 --max_steps 40000 --save_total_limit 2 --gradient_accumulation_steps 16 --per_gpu_eval_batch_size 4 --per_gpu_train_batch_size 4 --num_train_epochs 1 --model_name_or_path <gpt2/gpt2-medium/gpt2-large/gpt2-xlarge>

Generate paraphrases:

genienlp run-paraphrase --model_name_or_path <model_dir> --temperature 0.3 --repetition_penalty 1.0 --num_samples 4 --batch_size 32 --input_file <input_tsv_file> --input_column 1

Translation

Use the following command for training/ finetuning an NMT model:

genienlp train --train_tasks almond_translate --data <data_directory> --train_languages <src_lang> --eval_languages <tgt_lang> --no_commit --train_iterations <iterations> --preserve_case --save <save_dir> --exist_ok --skip_cache --model TransformerSeq2Seq --pretrained_model <hf_model_name>

We currently support MarianMT, MBART, MT5, and M2M100 models.
To save a pretrained model in genienlp format without any finetuning, set train_iterations to 0. You can then use this model to do inference.

To produce translations for an eval/ test set run the following command:

genienlp predict --tasks almond_translate --data <data_directory> --pred_languages <src_lang> --pred_tgt_languages <tgt_lang> --path <path_to_saved_model> --eval_dir <eval_dir> --skip_cache --val_batch_size 4000 --evaluate <valid/test>  --overwrite --silent

If your dataset is a document or contains long examples, pass --translate_example_split to break the examples down into individual sentences before translation for better results.
To use alignment as described in our localization paper (cited below), use --replace_qp and --force_replace_qp which ensures the parameters between quotations marks in the sentence are preserved in the output. The alignment code has been updated and improved since 0.6.0 release, so if you wish to compare the results use genienlp <=0.6.0. However, we recommend using the newer version for higher translation quality.

Named Entity Disambiguation

First run a bootleg model to extract mentions, entity candidates, and contextual embeddings for the mentions.

genienlp bootleg-dump-features --train_tasks <train_task_names> --save <savedir> --preserve_case --data <dataset_dir> --train_batch_tokens 1200 --val_batch_size 2000 --database_type json --database_dir <database_dir> --min_entity_len 1 --max_entity_len 4 --bootleg_model <bootleg_model>

This command generates several output files. In <dataset_dir> you should see a prep dir which contains preprocessed data (e.g. data converted to memory-mapped format, several array to facilitate embedding lookup etc.) If your dataset doesn't change you can reuse the same files. It will also generate several files in <results_temp> folder. In eval_bootleg/[train|eval]/<bootleg_model>/bootleg_lables.jsonl you can see the examples, mentions, predicted candidates and their probabilities according to bootleg.

Now you can use the extracted features from bootleg in downstream tasks such as semantic parsing to improve named entity understanding and consequently generation:

genienlp train --train_tasks <train_task_names> --train_iterations <iterations> --preserve_case --save <savedir> --data <dataset_dir> --model TransformerSeq2Seq --pretrained_model facebook/bart-base --train_batch_tokens 1000 --val_batch_size 1000 --do_ned --database_dir <database_dir> --ned_retrieve_method bootleg --entity_attributes type_id type_prob --add_entities_to_text append --bootleg_model <bootleg_model>

See genienlp --help and genienlp <command> --help for more details about each argument.

Citation

If you use the MultiTask Question Answering model in your work, please cite The Natural Language Decathlon: Multitask Learning as Question Answering.

@article{McCann2018decaNLP,
  title={The Natural Language Decathlon: Multitask Learning as Question Answering},
  author={Bryan McCann and Nitish Shirish Keskar and Caiming Xiong and Richard Socher},
  journal={arXiv preprint arXiv:1806.08730},
  year={2018}
}

If you use the BERT-LSTM model (Identity encoder + MQAN decoder), please cite Schema2QA: High-Quality and Low-Cost Q&A Agents for the Structured Web

@InProceedings{xu2020schema2qa,
  title={{Schema2QA}: High-Quality and Low-Cost {Q\&A} Agents for the Structured Web},
  author={Silei Xu and Giovanni Campagna and Jian Li and Monica S. Lam},
  booktitle={Proceedings of the 29th ACM International Conference on Information and Knowledge Management},
  year={2020},
  doi={https://doi.org/10.1145/3340531.3411974}
}

If you use the paraphrasing model (BART or GPT-2 fine-tuned on a paraphrasing dataset), please cite AutoQA: From Databases to QA Semantic Parsers with Only Synthetic Training Data

@inproceedings{xu-etal-2020-autoqa,
    title = "{A}uto{QA}: From Databases to {QA} Semantic Parsers with Only Synthetic Training Data",
    author = "Xu, Silei  and Semnani, Sina  and Campagna, Giovanni  and Lam, Monica",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.emnlp-main.31",
    pages = "422--434",
}

If you use MarianMT/ MBART/ MT5 for translation task, or XLMR-LSTM model for Seq2Seq tasks, please cite Localizing Open-Ontology QA Semantic Parsers in a Day Using Machine Translation and the original paper that introduced the model.

@inproceedings{moradshahi-etal-2020-localizing,
    title = "Localizing Open-Ontology {QA} Semantic Parsers in a Day Using Machine Translation",
    author = "Moradshahi, Mehrad and Campagna, Giovanni and Semnani, Sina and Xu, Silei and Lam, Monica",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
    month = November,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.emnlp-main.481",
    pages = "5970--5983",
}