README.md

metadata

task_categories:
  - text-classification
language:
  - en
dataset_info:
  features:
    - name: text
      dtype: string
    - name: label
      dtype: int64
    - name: sa
      dtype: int64
  splits:
    - name: train
      num_bytes: 128596235
      num_examples: 1613790
    - name: test
      num_bytes: 35731728
      num_examples: 448276
    - name: dev
      num_bytes: 14325121
      num_examples: 179310
  download_size: 93470968
  dataset_size: 178653084

The Moji dataset (Blodgett et al., 2016) (http://slanglab.cs.umass.edu/TwitterAAE/) contains tweets used for sentiment analysis (either positive or negative sentiment), with additional information on the type of English used in the tweets which is a sensitive attribute considered in fairness-aware approaches (African-American English (AAE) or Standard-American English (SAE)).

The type of language is determined thanks to a supervised model. Only the data where the sensitive attribute is predicted with a certainty rate above a given threshold are kept.

Based on this principle we make available two versions of the Moji dataset, respectively with a threshold of 80% and of 90%. The dataset's distributions are presented below.

Dataset with 80% threshold

	Positive sentiment	Negative Sentiment	Total
AAE	73 013	44 023	117 036
SAE	1 471 427	652 913	2 124 340
Total	1 544 440	696 936	2 241 376

To load this dataset, use the following code :

dataset = load_dataset("LabHC/moji", revision='moji_conf_08')

or by default the version is the dataset with 80% threshold

dataset = load_dataset("LabHC/moji")

Dataset with 90% threshold

	Positive sentiment	Negative Sentiment	Total
AAE	30 827	18 409	49 236
SAE	793 867	351 600	1 145 467
Total	824 694	370 009	1 194 703

To load this dataset, use the following code :

dataset = load_dataset("LabHC/moji", revision='moji_conf_09')

---

Demographic Dialectal Variation in Social Media: A Case Study of African-American English (Blodgett et al., EMNLP 2016)