IM-TQA: A Chinese Table Question Answering Dataset with Implicit and Multi-type Table Structures

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 447–467 July 9-14, 2023 ©2023 Association for Computational Linguistics Songs Across Borders: Singable and Controllable Neural Lyric Translation Longshen Ou and Xichu Ma and Min-Yen Kan and Ye Wang National University of Singapore {longshen, ma_xichu, kanmy, wangye}@comp.nus.edu.sg Abstract The development of general-domain neural ma- chine translation (NMT) methods has advanced significantly in recent years, but the lack of naturalness and musical constraints in the out- puts makes them unable to produce singable lyric translations. This paper bridges the singa- bility quality gap by formalizing lyric transla- tion into a constrained translation problem, con- verting theoretical guidance and practical tech- niques from translatology literature to prompt- driven NMT approaches, exploring better adap- tation methods, and instantiating them to an English-Chinese lyric translation system. Our model achieves 99.85%, 99.00%, and 95.52% on length accuracy, rhyme accuracy, and word boundary recall. In our subjective evaluation, our model shows a 75% relative enhancement on overall quality, compared against naive fine- tuning1. 1 Introduction With the globalization of entertainment, it is becom- ing increasingly common for people to appreciate songs in foreign languages. Meanwhile, artists are internationalizing their work and building territo- ries worldwide. Nevertheless, an unfriendly barrier exists between the artists and the audience: most commercial songs are not written in multiple lan- guages. Worse still, most existing song translations entirely ignore the music constraints, rendering 1Code available at https://github.com/Sonata165/ControllableLyricTranslation them unsingable alone with the music. As a result, the language barrier complicates the interaction between artists and their audience. Obtaining singable lyric translations can facili- tate the globalization of the music publishing in- dustry to further promote the growth of its $5.9 billion USD market size (Verified Market Research, 2022). However, song translation is unusually diffi- cult for human translators, due to music constraints and style requirements. If we can construct lyric- specific machine translation (MT) systems that can produce drafts that satisfy these constraints and requirements, the difficulty and cost of lyric trans- lation will be largely reduced, as lyricists and trans- lators can start with such automatic drafts and can focus on post-processing for quality and creativity. However, obtaining singable lyrics from MT sys- tems is challenging. Figure 1 shows two sentences of lyrics from the song Let It Go, together with an MT output and a singable translation. We observe a notable quality gap between them. While the MT output correctly translates the source, it ignores all the criteria that matter to make the output singable: (1) The second sentence of the MT outputs is unnat- ural because of incoherent vocabulary selection and lack of aesthetics. (2) Overcrowded syllables in the first sentence of the MT outputs force performers to break music notes in the orange box into multi- ple pieces to align them with lyrics. The rhythm pattern consequently diverges from the composer’s intention. (3) The two-syllable word in the red box is situated across a musical pause (blue box), let�it�go�in�intro Subtitle …

Content truncated to 500 words