Pretrained language models have been used in various natural language processing applications. In the mental health domain, domain-specific language models are pretrained and released, which facilitates the early detection of mental health conditions. Social posts, e.g., on Reddit, are usually long documents. However, there are no domain-specific pretrained models for long-sequence modeling in the mental health domain. This paper conducts domain-specific continued pretraining to capture the long context for mental health. Specifically, we train and release MentalXLNet and MentalLongformer based on XLNet and Longformer. We evaluate the mental health classification performance and the long-range ability of these two domain-specific pretrained models. Our models are released in HuggingFace.
翻译:预训练的语言模型已在各种自然语言处理应用中使用。在心理健康领域,预训练并发布特定领域的语言模型,以便提早检测心理健康状况。社交帖子(例如Reddit上的帖子)通常是长文档。然而,在心理健康领域中,没有专门的预训练模型用于长序列建模。本文进行了领域特定的持续预训练,以捕获心理健康领域的长上下文。具体而言,我们基于XLNet和Longformer训练和发布了MentalXLNet和MentalLongformer。我们评估了这两个领域特定预训练模型的心理健康分类性能和长距离能力。我们的模型在HuggingFace上发布。