DynaSent: Dynamic Sentiment Analysis Dataset

DynaSent is an English-language benchmark task for ternary (positive/negative/neutral) sentiment analysis. This dataset card is forked from the original DynaSent Repository.

Contents

Citation

Christopher Potts, Zhengxuan Wu, Atticus Geiger, and Douwe Kiela. 2020. DynaSent: A dynamic benchmark for sentiment analysis. Ms., Stanford University and Facebook AI Research.

  @article{potts-etal-2020-dynasent,
    title={{DynaSent}: A Dynamic Benchmark for Sentiment Analysis},
    author={Potts, Christopher and Wu, Zhengxuan and Geiger, Atticus and Kiela, Douwe},
    journal={arXiv preprint arXiv:2012.15349},
    url={https://arxiv.org/abs/2012.15349},
    year={2020}}

Dataset files

The dataset is dynasent-v1.1.zip, which is included in this repository. v1.1 differs from v1 only in that v1.1 has proper unique ids for Round 1 and corrects a bug that led to some non-unique ids in Round 2. There are no changes to the examples or other metadata.

The dataset consists of two rounds, each with a train/dev/test split:

Round 1: Naturally occurring sentences

  • dynasent-v1.1-round01-yelp-train.jsonl
  • dynasent-v1.1-round01-yelp-dev.jsonl
  • dynasent-v1.1-round01-yelp-test.jsonl

Round 1: Sentences crowdsourced using Dynabench

  • dynasent-v1.1-round02-dynabench-train.jsonl
  • dynasent-v1.1-round02-dynabench-dev.jsonl
  • dynasent-v1.1-round02-dynabench-test.jsonl

SST-dev revalidation

The dataset also contains a version of the Stanford Sentiment Treebank dev set in our format with labels from our validation task:

  • sst-dev-validated.jsonl

Quick start

This function can be used to load any subset of the files:

import json

def load_dataset(*src_filenames, labels=None):
    data = []
    for filename in src_filenames:
        with open(filename) as f:
            for line in f:
                d = json.loads(line)
                if labels is None or d['gold_label'] in labels:
                    data.append(d)
    return data

For example, to create a Round 1 train set restricting to examples with ternary gold labels:

import os

r1_train_filename = os.path.join('dynasent-v1.1', 'dynasent-v1.1-round01-yelp-train.jsonl')

ternary_labels = ('positive', 'negative', 'neutral')

r1_train = load_dataset(r1_train_filename, labels=ternary_labels)

X_train, y_train = zip(*[(d['sentence'], d['gold_label']) for d in r1_train])

Data format

Round 1 format

{'hit_ids': ['y5238'],
 'sentence': 'Roto-Rooter is always good when you need someone right away.',
 'indices_into_review_text': [0, 60],
 'model_0_label': 'positive',
 'model_0_probs': {'negative': 0.01173639390617609,
  'positive': 0.7473671436309814,
  'neutral': 0.24089649319648743},
 'text_id': 'r1-0000001',
 'review_id': 'IDHkeGo-nxhqX4Exkdr08A', 
 'review_rating': 1,
 'label_distribution': {'positive': ['w130', 'w186', 'w207', 'w264', 'w54'],
  'negative': [],
  'neutral': [],
  'mixed': []},
 'gold_label': 'positive'}

Details:

  • 'hit_ids': List of Amazon Mechanical Turk Human Interface Tasks (HITs) in which this example appeared during validation. The values are anonymized but used consistently throughout the dataset.
  • 'sentence': The example text.
  • 'indices_into_review_text': indices of 'sentence' into the original review in the Yelp Academic Dataset.
  • 'model_0_label': prediction of Model 0 as described in the paper. The possible values are 'positive', 'negative', and 'neutral'.
  • 'model_0_probs': probability distribution predicted by Model 0. The keys are ('positive', 'negative', 'neutral') and the values are floats.
  • 'text_id': unique identifier for this entry.
  • 'review_id': review-level identifier for the review from the Yelp Academic Dataset containing 'sentence'.
  • 'review_rating': review-level star-rating for the review containing 'sentence' in the Yelp Academic Dataset. The possible values are 1, 2, 3, 4, and 5.
  • 'label_distribution': response distribution from the MTurk validation task. The keys are ('positive', 'negative', 'neutral') and the values are lists of anonymized MTurk ids, which are used consistently throughout the dataset.
  • 'gold_label': the label chosen by at least three of the five workers if there is one (possible values: 'positive', 'negative', 'neutral', and 'mixed'), else None.

Here is some code one could use to augment a dataset, as loaded by load_dataset, with a field giving the full review text from the Yelp Academic Dataset:

import json

def index_yelp_reviews(yelp_src_filename='yelp_academic_dataset_review.json'):
    index = {}
    with open(yelp_src_filename) as f:
        for line in f:
            d = json.loads(line)
            index[d['review_id']] = d['text']
    return index

yelp_index = index_yelp_reviews()

def add_review_text_round1(dataset, yelp_index):
    for d in dataset:
        review_text = yelp_index[d['text_id']]
        # Check that we can find the sentence as expected:
        start, end = d['indices_into_review_text']
        assert review_text[start: end] == d['sentence']
        d['review_text'] = review_text
    return dataset

Round 2 format

{'hit_ids': ['y22661'],
 'sentence': "We enjoyed our first and last meal in Toronto at Bombay Palace, and I can't think of a better way to book our journey.",
 'sentence_author': 'w250',
 'has_prompt': True,
 'prompt_data': {'indices_into_review_text': [2093, 2213],
  'review_rating': 5,
  'prompt_sentence': "Our first and last meals in Toronto were enjoyed at Bombay Palace and I can't think of a better way to bookend our trip.",
  'review_id': 'Krm4kSIb06BDHternF4_pA'},
 'model_1_label': 'positive',
 'model_1_probs': {'negative': 0.29140257835388184,
  'positive': 0.6788994669914246,
  'neutral': 0.029697999358177185},
 'text_id': 'r2-0000001',
 'label_distribution': {'positive': ['w43', 'w26', 'w155', 'w23'],
  'negative': [],
  'neutral': [],
  'mixed': ['w174']},
 'gold_label': 'positive'}

Details:

  • 'hit_ids': List of Amazon Mechanical Turk Human Interface Tasks (HITs) in which this example appeared during validation. The values are anonymized but used consistently throughout the dataset.
  • 'sentence': The example text.
  • 'sentence_author': Anonymized MTurk id of the worker who wrote 'sentence'. These are from the same family of ids as used in 'label_distribution', but this id is never one of the ids in 'label_distribution' for this example.
  • 'has_prompt': True if the 'sentence' was written with a Prompt else False.
  • 'prompt_data': None if 'has_prompt' is False, else:
    • 'indices_into_review_text': indices of 'prompt_sentence' into the original review in the Yelp Academic Dataset.
    • 'review_rating': review-level star-rating for the review containing 'sentence' in the Yelp Academic Dataset.
    • 'prompt_sentence': The prompt text.
    • 'review_id': review-level identifier for the review from the Yelp Academic Dataset containing 'prompt_sentence'.
  • 'model_1_label': prediction of Model 1 as described in the paper. The possible values are 'positive', 'negative', and 'neutral'.
  • 'model_1_probs': probability distribution predicted by Model 1. The keys are ('positive', 'negative', 'neutral') and the values are floats.
  • 'text_id': unique identifier for this entry.
  • 'label_distribution': response distribution from the MTurk validation task. The keys are ('positive', 'negative', 'neutral') and the values are lists of anonymized MTurk ids, which are used consistently throughout the dataset.
  • 'gold_label': the label chosen by at least three of the five workers if there is one (possible values: 'positive', 'negative', 'neutral', and 'mixed'), else None.

To add the review texts to the 'prompt_data' field, one can extend the code above for Round 1 with the following function:

def add_review_text_round2(dataset, yelp_index):
    for d in dataset:
        if d['has_prompt']:
            prompt_data = d['prompt_data']
            review_text = yelp_index[prompt_data['review_id']]
            # Check that we can find the sentence as expected:
            start, end = prompt_data['indices_into_review_text']
            assert review_text[start: end] == prompt_data['prompt_sentence']
            prompt_data['review_text'] = review_text
    return dataset

SST-dev format

{'hit_ids': ['s20533'],
 'sentence': '-LRB- A -RRB- n utterly charming and hilarious film that reminded me of the best of the Disney comedies from the 60s.',
 'tree': '(4 (2 (1 -LRB-) (2 (2 A) (3 -RRB-))) (4 (4 (2 n) (4 (3 (2 utterly) (4 (3 (4 charming) (2 and)) (4 hilarious))) (3 (2 film) (3 (2 that) (4 (4 (2 (2 reminded) (3 me)) (4 (2 of) (4 (4 (2 the) (4 best)) (2 (2 of) (3 (2 the) (3 (3 Disney) (2 comedies))))))) (2 (2 from) (2 (2 the) (2 60s)))))))) (2 .)))',
 'text_id': 'sst-dev-validate-0000437',
 'sst_label': '4',
 'label_distribution': {'positive': ['w207', 'w3', 'w840', 'w135', 'w26'],
  'negative': [],
  'neutral': [],
  'mixed': []},
 'gold_label': 'positive'}

Details:

  • 'hit_ids': List of Amazon Mechanical Turk Human Interface Tasks (HITs) in which this example appeared during validation. The values are anonymized but used consistently throughout the dataset.
  • 'sentence': The example text.
  • 'tree': The parsetree for the example as given in the SST distribution.
  • 'text_id': A new identifier for this example.
  • 'sst_label': The root-node label from the SST. Possible values '0', '1' '2', '3', and '4'.
  • 'label_distribution': response distribution from the MTurk validation task. The keys are ('positive', 'negative', 'neutral') and the values are lists of anonymized MTurk ids, which are used consistently throughout the dataset.
  • 'gold_label': the label chosen by at least three of the five workers if there is one (possible values: 'positive', 'negative', 'neutral', and 'mixed'), else None.

Models

Model 0 and Model 1 from the paper are available here:

https://drive.google.com/drive/folders/1dpKrjNJfAILUQcJPAFc5YOXUT51VEjKQ?usp=sharing

This repository includes a Python module dynasent_models.py that provides a Hugging Face-based wrapper around these (PyTorch) models. Simple examples:

import os
from dynasent_models import DynaSentModel

# `dynasent_model0` should be downloaded from the above Google Drive link and 
# placed in the `models` directory. `dynasent_model1` works the same way.
model = DynaSentModel(os.path.join('models', 'dynasent_model0.bin'))

examples = [
    "superb",
    "They said the experience would be amazing, and they were right!",
    "They said the experience would be amazing, and they were wrong!"]

model.predict(examples)

This should return the list ['positive', 'positive', 'negative'].

The predict_proba method provides access to the predicted distribution over the class labels; see the demo at the bottom of dynasent_models.py for details.

The following code uses load_dataset from above to reproduce the Round 2 dev-set report on Model 0 from the paper:

import os
from sklearn.metrics import classification_report
from dynasent_models import DynaSentModel

dev_filename = os.path.join('dynasent-v1.1', 'dynasent-v1.1-round02-dynabench-dev.jsonl')

dev = load_dataset(dev_filename)

X_dev, y_dev = zip(*[(d['sentence'], d['gold_label']) for d in dev])

model = DynaSentModel(os.path.join('models', 'dynasent_model0.bin'))

preds = model.predict(X_dev)

print(classification_report(y_dev, preds, digits=3))

For a fuller report on these models, see our paper and our model card.

Other files

Analysis notebooks

The following notebooks reproduce the dataset statistics, figures, and random example selections from the paper:

  • analyses_comparative.ipynb
  • analysis_round1.ipynb
  • analysis_round2.ipynb
  • analysis_sst_dev_revalidate.ipynb

The Python module dynasent_utils.py contains functions that support those notebooks, and dynasent.mplstyle helps with styling the plots.

Datasheet

The Datasheet for our dataset:

Model Card

The Model Card for our models:

Tests

The module test_dataset.py contains PyTest tests for the dataset. To use it, run

py.test -vv test_dataset.py

in the root directory of this repository.

Validation HIT code

The file validation-hit-contents.html contains the HTML/Javascript used in the validation task. It could be used directly on Amazon Mechanical Turk, by simply pasting its contents into the usual HIT creation window.

License

DynaSent has a Creative Commons Attribution 4.0 International License.

Models trained or fine-tuned on dynabench/dynasent

None yet