Pankaj8922/query-classification

Dataset Card for Query Classification

Dataset Summary

Query Classification is a dataset of 21,627 English queries categorized into 19 distinct classes. The dataset was created by translating Chinese search queries into English using machine translation, making it suitable for training and evaluating text classification models for query intent detection.

Supported Tasks and Leaderboards

• Text Classification: Classify queries into one of 19 categories.
• Intent Detection: Identify the user's intent behind a search query.

Languages

The dataset is in English (translated from Chinese).

Dataset Structure

Data Instances

Each instance consists of a query and its corresponding label:

{
  "Query": "How to promote hair growth",
  "Label": "Medical care"
}

Data Fields

• Query (string): The English query text
• Label (string): The category label (one of 19 classes)

Data Splits

Split	Size
Train	21,627

Label Distribution

Label	Count
Education and Training	4,273
Medical care	2,656
Daily Life and Welfare	2,369
Books and novels	1,668
Products	1,342
Movies, TV shows, and anime	1,311
Characters	1,222
Software tools	919
Transportation and Tourism	908
Social Sciences and Technology	867
News	863
Game	705
Government Affairs	502
Production and manufacturing	487
Finance	436
Audio and performances	345
Companies and corporate hiring	267
Real estate decoration	258
Law	229

Dataset Creation

Curation Rationale

This dataset was created to provide a multi-class query classification benchmark for natural language processing tasks. The queries span diverse topics, making it useful for training models to understand user intent in search and conversational AI applications.

Source Data

The original data consisted of Chinese search queries that were translated to English using machine translation (Tencent/Hy-MT2-1.8B model).

Data Collection and Processing

1. Chinese queries and labels were collected from a filtered query classification dataset
2. Labels were translated once using batched machine translation
3. Queries were translated in batches of 96 using machine translation
4. Quality control: Queries where the translation still contained Chinese characters were flagged and excluded from the final dataset
5. Only successfully translated queries with consistent English labels were retained

Who are the source language producers?

The original queries were produced by Chinese-speaking users submitting search queries.

Annotations

The labels were originally in Chinese and translated to English using machine translation. Label consistency was verified to ensure no duplicate or variant translations existed for the same category.

Annotation process

• Labels were translated using machine translation (Tencent/Hy-MT2-1.8B)
• Translation quality was verified by checking for remaining Chinese characters
• Label consistency was manually verified to ensure all 19 categories have unique, consistent names

Who are the annotators?

Machine translation was performed using the Tencent/Hy-MT2-1.8B model.

Personal and Sensitive Information

The dataset contains search queries that may include names of public figures, but no private personal information.

Considerations for Using the Data

Social Impact of Dataset

This dataset can help improve query classification systems, enabling better search experiences and intent understanding in various applications.

Discussion of Biases

The dataset reflects the interests and search behaviors of Chinese-speaking users, which may introduce cultural and topical biases. The machine translation process may also introduce subtle semantic shifts from the original queries.

Other Known Limitations

• Machine translation quality may vary across query types
• Some nuanced queries may have lost subtle meaning during translation
• The dataset contains only one split (train) without validation or test sets

Licensing Information

This dataset is licensed under the Apache License 2.0.