WANLI

This repository contains the code and data for WANLI: Worker and AI Collaboration for Natural Language Inference Dataset Creation.

WANLI (Worker-AI Collaboration for NLI) is a collection of 108K English sentence pairs for the task of natural language inference (NLI). Each example is created by first identifying a "pocket" of examples in MultiNLI that share a challenging reasoning pattern, then instructing GPT-3 to write a new example with the same pattern. The set of generated examples are automatically filtered to contain those most likely to aid model training, and finally labeled and optionally revised by human annotators. In this way, WANLI represents a new approach to dataset creation that combines the generative strength of language models and evaluative strength of humans.

Model

You can download a WANLI-trained RoBERTa-large model from HuggingFace models here. This also includes a small demo!

Data

Download the WANLI dataset here!

Other NLI datasets used in this work (including MultiNLI and the out-of-domain evaluation sets) can be found in this Google Drive folder.

Pipeline

Here are the steps to replicate the process of creating WANLI. Recall that the prerequisites of this pipeline are an existing dataset (we use MultiNLI) and a task model trained on this dataset (we finetune RoBERTa-large). The relevant scripts can be found in the scripts/ folder.

1. Train RoBERTa-large on MultiNLI with classification/run_nli.py. The MultiNLI data is stored in data/mnli/. The model will be saved as models/roberta-large-mnli.
2. Compute the training dynamics for each example in the training set, using the saved checkpoints from the trained model, with cartography/compute_training_dynamics.py. The training dynamics will be stored inside the model directory. These statistics are used to collect the seed dataset via the most ambiguous p% of the training set.
3. In order to retrieve nearest neighbors for each seed example, we will pre-compute CLS token embeddings for all MultiNLI examples relative to the trained model. Use representations/embed_examples.py to produce a numpy file called mnli.npy inside the model directory.
4. Use pipeline.py to generate examples stored as generated_data/examples.jsonl! The pipeline uses the ambiguous seed examples found in step (2) and nearest neighbors found via the pre-computed embeddings from step (3), in order to generate examples with challenging reasoning patterns. For this step, you will need access to the GPT-3 API.
5. Heuristically filter the generated examples with filtering/filter.py to get generated_data/filtered_examples.jsonl.
6. Now we filter based on the estimated max variability, in order to keep examples most likely to aid model training. To do this, estimate the "training dynamics" of the generated data with respect to our trained task model, using cartography/compute_train_dy_metrics.py. Then, filter the dataset to keep examples with the highest estimated max variability using filtering/keep_ambiguous.py to create the final unlabeled dataset called generated_data/ambiguous_examples.jsonl.
7. Recruit humans to annotate examples in the final data file! Use the processing scripts in creating_wanli/ to process AMT batch results and postprocess them into NLI examples.

Citation

@inproceedings{liu-etal-2022-wanli,
    title = "{WANLI}: Worker and {AI} Collaboration for Natural Language Inference Dataset Creation",
    author = "Liu, Alisa  and
      Swayamdipta, Swabha  and
      Smith, Noah A.  and
      Choi, Yejin",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2022",
    month = dec,
    year = "2022",
    address = "Abu Dhabi, United Arab Emirates",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.findings-emnlp.508",
    pages = "6826--6847",
}