metadata

license: apache-2.0

Dataset Card for Hmar Language Conversations

This dataset contains anonymized WhatsApp group chat data, primarily in Hmar, a language spoken by the Hmar community. The dataset has been cleaned to remove sensitive information and includes messages exchanged in various WhatsApp groups. The data is intended for research in Natural Language Processing (NLP) and machine learning tasks focused on Hmar language analysis.

Dataset Details

Dataset Description

This dataset includes cleaned conversations from multiple WhatsApp groups. The conversations are mainly in Hmar, with some occasional use of other languages. The dataset has been processed to ensure the removal of sensitive information such as phone numbers, deleted messages, and media links.

Curated by: [Hmar Language Dataset Project]
Funded by [optional]: [self]
Shared by [optional]: [self]
Language(s) (NLP): Hmar
License: Apache 2.0

Dataset Sources

Repository: [Repository Link or "Coming Soon"]
Paper [optional]: [Link to Paper]
Demo [optional]: [Link to Demo or example]

Uses

Direct Use

This dataset is designed for research purposes, particularly for projects involving:

Language Modeling for Hmar
Sentiment Analysis on Hmar conversations
Machine Translation between Hmar and other languages
Chatbot Development for Hmar speakers

Out-of-Scope Use

The dataset is not suitable for:

Any use that violates the privacy of individuals (e.g., identifying users or using the data for advertising purposes)
Use cases that involve real-time monitoring or surveillance
Misuse involving the exploitation of group members or personal information

Dataset Structure

Each cleaned conversation file contains two columns:

user: The anonymized identifier of the user (e.g., @user).
message: The content of the message exchanged in the group.

Example:

user	message
@user1	i dam hlak maw ruola?
@user2	aw,ka dam e,nang i dam mo?

The dataset is organized into subfolders by group name, allowing easy access to each group’s conversations.

Dataset Creation

Curation Rationale

This dataset was created to support research in NLP for Hmar, a lesser-researched language. By using cleaned WhatsApp chat data, it provides real-world conversational data for researchers working on machine learning models and linguistic studies related to Hmar.

Source Data

The dataset consists of WhatsApp group chat data collected from various groups. The conversations primarily involve users who speak Hmar, with occasional use of other languages. The data is provided in CSV format, with each file representing a specific group chat.

Data Collection and Processing

The chat data was exported directly from WhatsApp and cleaned using a custom script. The cleaning process involves:

Removing media-related content (<Media omitted>).
Anonymizing phone numbers (replacing @+91xxxxxxxxxx with @user).
Retaining only the final versions of edited messages.
Removing system messages such as "This message was deleted."

Who are the source data producers?

The source data was generated by participants in various WhatsApp groups, primarily speakers of Hmar. The group members have consented to the use of their conversations for research purposes, but their identities have been anonymized in this dataset.

Annotations [optional]

This dataset does not contain manual annotations, but automatic preprocessing steps (e.g., removal of personal information) were carried out to ensure the privacy of the participants.

Annotation process

The dataset was processed programmatically, with automatic removal of personal data and non-message content (e.g., system messages). No manual annotations were performed on the conversations.

Who are the annotators?

As the dataset was cleaned programmatically, no human annotators were involved in this process.

Personal and Sensitive Information

The dataset was anonymized to remove any personal and sensitive information. This includes:

Anonymizing phone numbers.
Removing system messages and media links.
Ensuring only the final version of edited messages is included.

Bias, Risks, and Limitations

Bias

The dataset may be biased towards the conversational style and topics of WhatsApp groups, which could differ from other forms of text (e.g., formal writing or news articles). Additionally, since the data was collected from specific groups, it may not fully represent all possible linguistic variations within the Hmar community.

Risks and Limitations

While every effort was made to anonymize the data, users should ensure compliance with local data privacy laws (e.g., GDPR) when using the dataset.
The dataset may not be comprehensive enough for building a fully robust model for Hmar language processing, as it only represents a small subset of WhatsApp group conversations.

Language(s) (NLP): Hmar

Glottocode: The Glottocode for Hmar is not standardized across all resources, and entries such as "hmar1245" have been used in some databases. Given the variability, we recommend focusing on Hmar as the primary language identifier for this dataset.

Recommendations

Users are encouraged to combine this dataset with other sources of Hmar language data for more comprehensive NLP tasks. Researchers should also be aware of potential biases in the data and exercise caution when applying it to sensitive contexts.

Citation [optional]

BibTeX:

@misc{hmar_conversations,
  author = {Hmar Language Dataset Project},
  title = {Hmar Language Conversations Dataset},
  year = {2024},
  publisher = {Hugging Face},
  url = {https://huggingface.co./datasets/hmar-lang/hmar-language-conversations}
}