Languages: en
Multilinguality: monolingual
Language Creators: found
Annotations Creators: crowdsourced

Dataset Card for sst

Dataset Summary

The Stanford Sentiment Treebank is the first corpus with fully labeled parse trees that allows for a complete analysis of the compositional effects of sentiment in language.

Supported Tasks and Leaderboards

  • sentiment-scoring: Each complete sentence is annotated with a float label that indicates its level of positive sentiment from 0.0 to 1.0. One can decide to use only complete sentences or to include the contributions of the sub-sentences (aka phrases). The labels for each phrase are included in the dictionary configuration. To obtain all the phrases in a sentence we need to visit the parse tree included with each example. In contrast, the ptb configuration explicitly provides all the labelled parse trees in Penn Treebank format. Here the labels are binned in 5 bins from 0 to 4.
  • sentiment-classification: We can transform the above into a binary sentiment classification task by rounding each label to 0 or 1.


The text in the dataset is in English

Dataset Structure

Data Instances

For the default configuration:

{'label': 0.7222200036048889,
 'sentence': 'Yet the act is still charming here .',
 'tokens': 'Yet|the|act|is|still|charming|here|.',
 'tree': '15|13|13|10|9|9|11|12|10|11|12|14|14|15|0'}

For the dictionary configuration:

{'label': 0.7361099720001221, 
'phrase': 'still charming'}

For the ptb configuration:

{'ptb_tree': '(3 (2 Yet) (3 (2 (2 the) (2 act)) (3 (4 (3 (2 is) (3 (2 still) (4 charming))) (2 here)) (2 .))))'}

Data Fields

  • sentence: a complete sentence expressing an opinion about a film
  • label: the degree of "positivity" of the opinion, on a scale between 0.0 and 1.0
  • tokens: a sequence of tokens that form a sentence
  • tree: a sentence parse tree formatted as a parent pointer tree
  • phrase: a sub-sentence of a complete sentence
  • ptb_tree: a sentence parse tree formatted in Penn Treebank-style, where each component's degree of positive sentiment is labelled on a scale from 0 to 4

Data Splits

The set of complete sentences (both default and ptb configurations) is split into a training, validation and test set. The dictionary configuration has only one split as it is used for reference rather than for learning.

Dataset Creation

Curation Rationale

[Needs More Information]

Source Data

Initial Data Collection and Normalization

[Needs More Information]

Who are the source language producers?

Rotten Tomatoes reviewers.


Annotation process

[Needs More Information]

Who are the annotators?

[Needs More Information]

Personal and Sensitive Information

[Needs More Information]

Considerations for Using the Data

Social Impact of Dataset

[Needs More Information]

Discussion of Biases

[Needs More Information]

Other Known Limitations

[Needs More Information]

Additional Information

Dataset Curators

[Needs More Information]

Licensing Information

[Needs More Information]

Citation Information

    title = "Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank",
    author = "Socher, Richard  and
      Perelygin, Alex  and
      Wu, Jean  and
      Chuang, Jason  and
      Manning, Christopher D.  and
      Ng, Andrew  and
      Potts, Christopher",
    booktitle = "Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing",
    month = oct,
    year = "2013",
    address = "Seattle, Washington, USA",
    publisher = "Association for Computational Linguistics",
    url = "",
    pages = "1631--1642",


Thanks to @patpizio for adding this dataset.

Models trained or fine-tuned on sst

None yet