Multilingual Datasets for Augmentative and Alternative Communication (AAC)

Multilingual Datasets for Augmentative and Alternative Communication (AAC)

There is a lack of text resources for designing predictive AAC interfaces in many languages.

This repository contains automatically translated versions of four existing datasets. We also used Llama4 to correct errors and augment the original datasets. The datasets were chosen to match the conversational style text AAC users may want to produce in person-to-person conversations.

• Multilingual COMM2 - Statements and questions written in response to ten everyday communication situations. We created an additional concise version of each utterances using Llama4.
• Multilingual Turk - Six turn dialogues about different topics. We created an additional similar dialogue using Llama4.
• Multilingual Imagine - Messages written by people who were asked to imagine they were using a scanning-style AAC interface.
• Multilingual DailyDialog - Multi-turn dialogues about different topics. We translated the first 20% of the original dataset.

For details of how we developed these datasets, see our EMNLP 2026 workshop paper:

@inproceedings{vertanen_multilingual,
  author    = {Keith Vertanen},
  title     = {Multilingual Datasets for Augmentative and Alternative Communication},
  booktitle = {Proceedings of the 6th Multilingual Representation Learning Workshop},
  year      = {2026},
}

Additional files:

• Python scripts used to construct the datasets.
• Character inventories for each language based on the datasets.

Details:

• We used Google’s Cloud Translation API to translate into 33 different languages. We performed the translation in July of 2025. In the case of the dialogue datasets, we treated an entire dialogue as a single document.
• For COMM2, Turk, and Imagine, we used the Llama 4 Maverick large language model (LLM) to correct obvious spelling or grammar errors in the original datasets.
• For COMM2, we used the LLM to produce a second concise version of each statement or question.
• For Turk, we used the LLM to produce a second similar dialogue based on each of the original dialogues.
• Datasets are provided in JSON format.
• Each of the linked zip files contains a README file with further dataset specific details.
• The first three datasets are licensed under CC BY 4.0. The DailyDialog dataset retains the original authors’ CC BY-NC-SA 4.0 license.

This material is based upon work supported by the National Science Foundation under Grant No. IIS-1750193 and IIS-2402876.