Cross-language few-shot intent recognition via prompt-based tuning

Cross-language intent recognition is a fundamental task in cross-language understanding. Recently, this task has been addressed by pretrained cross-language language models. Existing approaches typically augment pretrained language models with additional data, such as annotated parallel corpora. However, these additional data are scarce in practice, especially for low-resource languages. Inspired by the recent effective results of prompt learning, this paper proposes a new framework for enhancing cross-language few-shot intent recognition methods based on prompt tuning (CIRP). The proposed method converts the cross-language intent recognition task into a masked language modelling problem by designing prompt templates. To make the proposed model more generalizable, and avoid templates and label words dependent on a specific language, the method encodes the prompt templates into language-independent embedding representations via the multilingual pretrained language models, and initializes the label words into soft label words by averaging the [mask] vector values from different utterances of the same label, which reduces the distance between label word embeddings and encoder outputs of the [mask] to increase the accuracy of cross-language intent recognition. The experimental results on the few-shot cross-language MultiATIS++, MIvD benchmark dataset show that, compared with the four baseline models, the CIRP performs remarkably well in terms of intent recognition accuracy. Notably, when the sample sizes are set to 1 and 8 shots, the cross-language intent recognition accuracy metrics improve by an average of 11.75% compared with those of the baseline models.