PrAACT: Predictive Augmentative and Alternative Communication with Transformers

Augmentative and Alternative Communication (AAC) systems assist individuals with complex communication needs, allowing them to construct sentences by selecting communication cards. The facilitation of sentence construction primarily entails enabling efficient communication card selection. An approach to this involves using language models for communication card prediction. Nonetheless, the heterogeneity of the AAC population poses unique challenges due to diverse vocabulary needs, suggesting one-size-fits-all approaches may be insufficient. This study introduces PrAACT, a method that leverages large, transformer-based language models, such as BERT, for communication card prediction. This method focuses on easy adaptability to incorporate user-specific vocabularies into the model. The method involves adapting a text corpus to the AAC domain, fine-tuning a transformer-based language model, and replacing the language model weights with an encoded representation of the user's vocabulary. We conducted an assessment of PrAACT under both zero-shot and few-shot scenarios. The performance of the models produced using PrAACT was superior to those pre-trained for the task. In addition, the main advantage of PrAACT is that it allows to quickly adapt a transformer-based language model to communication card prediction according to the user's vocabulary.

Automated summary

This study introduces PrAACT, a transformer-based method that assists individuals with complex communication needs in sentence construction. It utilizes language models to predict communication cards and can easily adapt to user-specific vocabularies. Evaluation shows that PrAACT outperforms pre-trained models for this task. The main advantage of PrAACT is its ability to quickly adapt a language model according to the user's vocabulary.