Why Did the Chicken Cross the Road? Rephrasing and Analyzing Ambiguous Questions in VQA

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 983–1001 July 9-14, 2023 ©2023 Association for Computational Linguistics UniLG: A Unified Structure-aware Framework for Lyrics Generation Tao Qian1,2, Fan Lou2, Jiatong Shi3, Yuning Wu1, Shuai Guo1, Xiang Yin2, Qin Jin1∗ 1Renmin University of China, P.R.China 2ByteDance AI Lab 3Carnegie Mellon University, U.S.A {qiantao, yuningwu, shuaiguo, qjin}@ruc.edu.cn, tianzhong.t, yinxiang.stephen}@bytedance.com, jiatongs@cs.cmu.edu Abstract As a special task of natural language gener- ation, conditional lyrics generation needs to consider the structure of generated lyrics1 and the relationship between lyrics and music. Due to various forms of conditions, a lyrics gen- eration system is expected to generate lyrics conditioned on different signals, such as music scores, music audio, or partially-finished lyrics, etc. However, most of the previous works have ignored the musical attributes hidden behind the lyrics and the structure of the lyrics. Addi- tionally, most works only handle limited lyrics generation conditions, such as lyrics genera- tion based on music score or partial lyrics, they can not be easily extended to other generation conditions with the same framework. In this pa- per, we propose a unified structure-aware lyrics generation framework named UniLG. Specif- ically, we design compound templates that in- corporate textual and musical information to improve structure modeling and unify the dif- ferent lyrics generation conditions. Extensive experiments demonstrate the effectiveness of our framework. Both objective and subjective evaluations show significant improvements in generating structural lyrics. 1 Introduction Great progress has been made in natural language generation (NLG) with pre-trained language mod- els in recent years (Lewis et al., 2020; Radford et al., 2019; Brown et al., 2020). Lyrics gener- ation is also a special task of NLG (Chen and Lerch, 2020; Gill et al., 2020). Different from gen- eral natural language, lyrics eventually need to be presented with music after the composition. This requires the lyrics to follow song-writing rules (i.e., the structure of lyrics), such as clear paragraphs with chorus and verse concepts. However, most ∗*Corresponding Author. 1The structure of lyrics in our work means that lyrics have the chorus and verse parts among the sentences. Figure 1: Example chorus parts of a song. We use different colors for different beats2 within the bar3 and rhythm patterns are shown in 4/4 time signatures4. The same melody and rhythm pattern may repeat several times when meeting the chorus parts of the song. The melody and rhythm patterns can hint the correspondences between lyric sentences, e.g. the same or similar sentences. previous works ignore the musical concepts be- hind the lyrics and do not consider the structure of lyrics (Sheng et al., 2021; Qian et al., 2022). To explicitly model the structure of lyrics, some works introduce additional structural labels (e.g., sentence-level chorus and verse label), which in- evitably require much effort for additional human annotation (Potash et al., 2015; Lu et al., 2019). To avoid the huge annotation cost, other works either adopt predefined formats (e.g., the number of sylla- bles in each sentence) …

Content truncated to 500 words