Dongxu Li


Automatic Gloss Dictionary for Sign Language Learners
Chenchen Xu | Dongxu Li | Hongdong Li | Hanna Suominen | Ben Swift
Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: System Demonstrations

A multi-language dictionary is a fundamental tool for language learning, allowing the learner to look up unfamiliar words. Searching an unrecognized word in the dictionary does not usually require deep knowledge of the target language. However, this is not true for sign language, where gestural elements preclude this type of easy lookup. This paper introduces GlossFinder, an online tool supporting 2, 000 signs to assist language learners in determining the meaning of given signs. Unlike alternative systems of complex inputs, our system requires only that learners imitate the sign in front of a standard webcam. A user study conducted among sign language speakers of varying ability compared our system against existing alternatives and the interviews indicated a clear preference for our new system. This implies that GlossFinder can lower the barrier in sign language learning by addressing the common problem of sign finding and make it accessible to the wider community.

The Devil in Linear Transformer
Zhen Qin | Xiaodong Han | Weixuan Sun | Dongxu Li | Lingpeng Kong | Nick Barnes | Yiran Zhong
Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing

Linear transformers aim to reduce the quadratic space-time complexity of vanilla transformers. However, they usually suffer from degraded performances on various tasks and corpus. In this paper, we examine existing kernel-based linear transformers and identify two key issues that lead to such performance gaps: 1) unbounded gradients in the attention computation adversely impact the convergence of linear transformer models; 2) attention dilution which trivially distributes attention scores over long sequences while neglecting neighbouring structures. To address these issues, we first identify that the scaling of attention matrices is the devil in unbounded gradients, which turns out unnecessary in linear attention as we show theoretically and empirically. To this end, we propose a new linear attention that replaces the scaling operation with a normalization to stabilize gradients. For the issue of attention dilution, we leverage a diagonal attention to confine attention to only neighbouring tokens in early layers. Benefiting from the stable gradients and improved attention, our new linear transformer model, transNormer, demonstrates superior performance on text classification and language modeling tasks, as well as on the challenging Long-Range Arena benchmark, surpassing vanilla transformer and existing linear variants by a clear margin while being significantly more space-time efficient. The code is available at .