metadata

license: mit
task_categories:
  - text-classification
language:
  - en
size_categories:
  - 1M<n<10M
annotations_creators:
  - no-annotation
multilinguality:
  - monolingual
pretty_name: UTCD

Universal Text Classification Dataset (UTCD)

Load dataset

from datasets import load_dataset
dataset = load_dataset('claritylab/utcd', name='in-domain')

Description

UTCD is a curated compilation of 18 datasets revised for Zero-shot Text Classification spanning 3 aspect categories of Sentiment, Intent/Dialogue, and Topic classification. UTCD focuses on the task of zero-shot text classification where the candidate labels are descriptive of the text being classified. TUTCD consists of ~ 6M/800K train/test examples.

UTCD was introduced in the Findings of ACL'23 Paper Label Agnostic Pre-training for Zero-shot Text Classification by Christopher Clarke, Yuzhao Heng, Yiping Kang, Krisztian Flautner, Lingjia Tang and Jason Mars. Project Homepage.

UTCD Datasets & Principles:

Sentiment
- GoEmotions introduced in GoEmotions: A Dataset of Fine-Grained Emotions
- TweetEval introduced in TWEETEVAL: Unified Benchmark and Comparative Evaluation for Tweet Classification (Sentiment subset)
- Emotion introduced in CARER: Contextualized Affect Representations for Emotion Recognition
- Amazon Polarity introduced in Character-level Convolutional Networks for Text Classification
- Finance Phrasebank introduced in Good debt or bad debt: Detecting semantic orientations in economic texts
- Yelp introduced in Character-level Convolutional Networks for Text Classification
Intent/Dialogue
- Schema-Guided Dialogue introduced in Towards Scalable Multi-Domain Conversational Agents: The Schema-Guided Dialogue Dataset
- Clinc-150 introduced in An Evaluation Dataset for Intent Classification and Out-of-Scope Prediction
- SLURP SLU introduced in SLURP: A Spoken Language Understanding Resource Package
- Banking77 introduced in Efficient Intent Detection with Dual Sentence Encoders
- Snips introduced in Snips Voice Platform: an embedded Spoken Language Understanding system for private-by-design voice interfaces
- NLU Evaluation introduced in Benchmarking Natural Language Understanding Services for building Conversational Agents
Topic
- AG News introduced in Character-level Convolutional Networks for Text Classification
- DBpedia 14 introduced in DBpedia: A Nucleus for a Web of Open Data
- Yahoo Answer Topics introduced in Character-level Convolutional Networks for Text Classification
- MultiEurlex introduced in MultiEURLEX -- A multi-lingual and multi-label legal document classification dataset for zero-shot cross-lingual transfer
- BigPatent introduced in BIGPATENT: A Large-Scale Dataset for Abstractive and Coherent Summarization
- Consumer Finance introduced in Consumer Complaint Database

In order to make NLP models more broadly useful, zero-shot techniques need to be capable of label, domain & aspect transfer. As such, in the construction of UTCD we enforce the following principles:

Textual labels: In UTCD, we mandate the use of textual labels. While numerical label values are often used in classification tasks, descriptive textual labels such as those present in the datasets across UTCD enable the development of techniques that can leverage the class name which is instrumental in providing zero-shot support. As such, for each of the compiled datasets, labels are standardized such that the labels are descriptive of the text in natural language.
Diverse domains and Sequence lengths: In addition to broad coverage of aspects, UTCD compiles diverse data across several domains such as Banking, Finance, Legal, etc each comprising varied length sequences (long and short). The datasets are listed above.

Structure

Data Samples

Each dataset sample contains the text, the label encoded as an integer, and the dataset name encoded as an integer.

{
    'text': "My favourite food is anything I didn't have to cook myself.",
    'labels': [215],
    'dataset_name': 0
}

Datasets Contained

The UTCD dataset contains 18 datasets, 9 in-domain, 9 out-of-domain, spanning 3 aspects: sentiment, intent and topic.

Below are statistics on the datasets.

In-Domain Datasets

Dataset	Aspect	#Samples in Train/Test	#labels	average #token in text in Train/Test
GoEmotions	sentiment	43K/5.4K	28	12/12
TweetEval	sentiment	45K/12K	3	19/14
Emotion	sentiment	16K/2K	6	17/17
SGD	intent	16K/4.2K	26	8/9
Clinc-150	intent	15K/4.5K	150	8/8
SLURP	intent	12K/2.6K	75	7/7
AG News	topic	120K7.6K	4	38/37
DBpedia	topic	560K/70K	14	45/45
Yahoo	topic	1.4M/60K	10	10/10

Out-of-Domain Datasets

Dataset	Aspect	#Samples in Train/Test	#labels	average #token in text
Amazon Polarity	sentiment	3.6M/400K	2	71/71
Financial Phrase Bank	sentiment	1.8K/453	3	19/19
Yelp	sentiment	650K/50K	3	128/128
Banking77	intent	10K/3.1K	77	11/10
SNIPS	intent	14K/697	7	8/8
NLU Eval	intent	21K/5.2K	68	7/7
MultiEURLEX	topic	55K/5K	21	1198/1853
Big Patent	topic	25K/5K	9	2872/2892
Consumer Finance	topic	630K/160K	18	190/189

Configurations

The in-domain and out-of-domain configurations has 2 splits: train and test.

The aspect-normalized configurations (aspect-normalized-in-domain, aspect-normalized-out-of-domain) has 3 splits: train, validation and test.

Below are statistics on the configuration splits.

In-Domain Configuration

Split	#samples
Train	2,192,703
Test	168,365

Out-of-Domain Configuration

Split	#samples
Train	4,996,673
Test	625,911

Aspect-Normalized In-Domain Configuration

Split	#samples
Train	115,127
Validation	12,806
Test	168,365

Aspect-Normalized Out-of-Domain Configuration

Split	#samples
Train	119,167
Validation	13,263
Test	625,911