Improving Chinese Punctuation Restoration via External POS Tagger

Automatic Speech Recognition (ASR) systems typically generate transcriptions without punctuation. To enhance readability and meet the expected input requirements for downstream language models, it's crucial to add punctuation marks in these transcripts. Most state-of-the-art improve the performance of punctuation restoration models by incorporating external information, such as part-of-speech(POS) tags. Although these models take POS tags into account, they predominantly focus on the English language, with limited research examining how POS tags enhance punctuation restoration performance in Chinese. In this paper, we validate the effectiveness of POS tags on the Chinese punctuation restoration task, and develop an innovative method to fusing POS tags with contextual embeddings. For English, we use the IWSLT dataset to verify the effectiveness of the fusion approach, while for Chinese we develop a new Chinese language dataset to evaluate the proposed methods. Experimental results show that our proposed method can consistently obtain performance gains, indicating its effectiveness for punctuation restoration tasks.