{
    "paper_id": "2022",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T07:33:00.602375Z"
    },
    "title": "ChipSong: A Controllable Lyric Generation System for Chinese Popular Song",
    "authors": [
        {
            "first": "Nayu",
            "middle": [],
            "last": "Liu",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Kuaishou Technology Co",
                "location": {
                    "settlement": "Ltd, Beijing",
                    "country": "China"
                }
            },
            "email": "liunayu@kuaishou.com"
        },
        {
            "first": "Wenjing",
            "middle": [],
            "last": "Han",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Kuaishou Technology Co",
                "location": {
                    "settlement": "Ltd, Beijing",
                    "country": "China"
                }
            },
            "email": "hanwenjing@kuaishou.com"
        },
        {
            "first": "Guangcan",
            "middle": [],
            "last": "Liu",
            "suffix": "",
            "affiliation": {},
            "email": "liuguangcan@kuaishou.com"
        },
        {
            "first": "Peng",
            "middle": [],
            "last": "Zhou",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Kuaishou Technology Co",
                "location": {
                    "settlement": "Ltd, Beijing",
                    "country": "China"
                }
            },
            "email": "zhoupeng@kuaishou.com"
        },
        {
            "first": "Ran",
            "middle": [],
            "last": "Zhang",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Kuaishou Technology Co",
                "location": {
                    "settlement": "Ltd, Beijing",
                    "country": "China"
                }
            },
            "email": "zhangran@kuaishou.com"
        },
        {
            "first": "Xiaorui",
            "middle": [],
            "last": "Wang",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Kuaishou Technology Co",
                "location": {
                    "settlement": "Ltd, Beijing",
                    "country": "China"
                }
            },
            "email": "wangxiaorui@kuaishou.com"
        },
        {
            "first": "Huabin",
            "middle": [],
            "last": "Ruan",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Kuaishou Technology Co",
                "location": {
                    "settlement": "Ltd, Beijing",
                    "country": "China"
                }
            },
            "email": "ruanhuabin@tsinghua.edu.cn"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "In this work, we take a further step towards satisfying practical demands in Chinese lyric generation from musical short-video creators, in respect of the challenges on songs' format constraints, creating specific lyrics from open-ended inspiration inputs, and language rhyme grace. One representative detail in these demands is to control lyric format at word level, that is, for Chinese songs, creators even expect fix-length words on certain positions in a lyric to match a special melody, while previous methods lack such ability. Although recent lyric generation community has made gratifying progress, most methods are not comprehensive enough to simultaneously meet these demands. As a result, we propose ChipSong, which is an assisted lyric generation system built based on a Transformerbased autoregressive language model architecture, and generates controlled lyric paragraphs fit for musical short-video display purpose, by designing 1) a novel Begin-Internal-End (BIE) word-granularity embedding sequence with its guided attention mechanism for word-level length format control, and an explicit symbol set for sentence-level length format control; 2) an open-ended trigger word mechanism to guide specific lyric contents generation; 3) a paradigm of reverse order training and shielding decoding for rhyme control. Extensive experiments show that our ChipSong generates fluent lyrics, with assuring the high consistency to predetermined control conditions.",
    "pdf_parse": {
        "paper_id": "2022",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "In this work, we take a further step towards satisfying practical demands in Chinese lyric generation from musical short-video creators, in respect of the challenges on songs' format constraints, creating specific lyrics from open-ended inspiration inputs, and language rhyme grace. One representative detail in these demands is to control lyric format at word level, that is, for Chinese songs, creators even expect fix-length words on certain positions in a lyric to match a special melody, while previous methods lack such ability. Although recent lyric generation community has made gratifying progress, most methods are not comprehensive enough to simultaneously meet these demands. As a result, we propose ChipSong, which is an assisted lyric generation system built based on a Transformerbased autoregressive language model architecture, and generates controlled lyric paragraphs fit for musical short-video display purpose, by designing 1) a novel Begin-Internal-End (BIE) word-granularity embedding sequence with its guided attention mechanism for word-level length format control, and an explicit symbol set for sentence-level length format control; 2) an open-ended trigger word mechanism to guide specific lyric contents generation; 3) a paradigm of reverse order training and shielding decoding for rhyme control. Extensive experiments show that our ChipSong generates fluent lyrics, with assuring the high consistency to predetermined control conditions.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Lyric generation is a recent emerging topic in intelligent music research community, which has attracted increasing attention and gained progress in the past few years (Watanabe et al., 2018; Manjavacas et al., 2019; Fan et al., 2019; Li et al., 2020; Zhang et al., 2020a; Nikolov et al., 2020; Sheng et al., 2021) . Meanwhile, observing a large amount of music lovers, amateurs, and professional musicians are gathering on today's fast growing Chinese short-video platforms (e.g., Kwai, TikTok, Wesee, etc.) , where they create and post musical short-videos actively, with purpose to obtain more Follows and Likes from general population; we believe it is worth to customize a lyric generation system for their short-video display purpose.",
                "cite_spans": [
                    {
                        "start": 168,
                        "end": 191,
                        "text": "(Watanabe et al., 2018;",
                        "ref_id": "BIBREF19"
                    },
                    {
                        "start": 192,
                        "end": 216,
                        "text": "Manjavacas et al., 2019;",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 217,
                        "end": 234,
                        "text": "Fan et al., 2019;",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 235,
                        "end": 251,
                        "text": "Li et al., 2020;",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 252,
                        "end": 272,
                        "text": "Zhang et al., 2020a;",
                        "ref_id": "BIBREF22"
                    },
                    {
                        "start": 273,
                        "end": 294,
                        "text": "Nikolov et al., 2020;",
                        "ref_id": "BIBREF12"
                    },
                    {
                        "start": 295,
                        "end": 314,
                        "text": "Sheng et al., 2021)",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 482,
                        "end": 508,
                        "text": "Kwai, TikTok, Wesee, etc.)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Hence, in this paper, we aim to put more emphasis on assisting creators from practical shortvideo scenario with realistic demands. In oder to collect their real demands, a qualitative investigation with 85 potential users (ages: 18\u223c40 years; 42 female, 43 males; 19 full-time musicians, 66 part-time musicians) is conducted at the very first stage. Here, we briefly release 4 representative demands as follow: 1) short lyric paragraphs are required to fit in short-video durations, mostly under 60 sec. (Zhang et al., 2020b) ; 2) open-ended inspiration inputs are desired to guide specific content generation from various creators; 3) length format controlling at sentence and even word level is expected to strictly match melody length format for flexible creation intents, where a Chinese word is generally composed of multiple characters (e.g., \"\u7231\" means love, \"\u7231 \u597d\" means hobby, \"\u7231 \u5c14 \u5170\" means Ireland) and one character sounds one syllable; 4) rich rhyme patterns are needed for smooth song singing. Although recent progress has been made on lyric generation, previous works are not comprehensive enough to simultaneously meet these customized demands; What's more, as far as we know, none of the existing work supports word-level length format control.",
                "cite_spans": [
                    {
                        "start": 503,
                        "end": 524,
                        "text": "(Zhang et al., 2020b)",
                        "ref_id": "BIBREF23"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Taking the above challenges in mind, we develop ChipSong, a lyric generation system, to assist musical short-video creators for Chinese popular song creation. As shown in Figure 1 , with ChipSong, a creator is encouraged to input a group of openended words (which are referred to as trigger words in the following) to represent his/her inspiration, Actual Lyric \u5bc2 \u5bde \u591c \u91cc \u72ec \u4f24 \u60b2 (bei) \u6211 \u7684 \u5fc3 \u65e0 \u4eba \u4f53 \u4f1a (hui) \u4f60 \u7ed9 \u7684 \u7231 \u75db \u5f7b \u6211 \u5fc3 \u6249 (fei) \u7f20 \u7ef5 \u53ea \u662f \u4e00 \u77ac \u95f4 \u7684 \u7f8e (mei) ",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 171,
                        "end": 179,
                        "text": "Figure 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Sentence-level Length Format {7; 7; 9; 9} Word-level Length Format {2, 2, 1, 2; 1, 1, 1, 2, 2; 1, 1, 1, 1, 2, 1, 2; 2, 2, 3, 1, 1} Example Input Template \u6bcf \u4e2a \u5931 \u7720 \u7684 \u591c \u665a (wan) \u8c01 \u5728 \u6211 \u5fc3 \u5934 \u8f7b \u53f9 (tan) \u662f \u7231 \u628a \u4eba \u53d8 \u5f97 \u66f4 \u654f \u611f (gan) \u601d \u5ff5 \u672c \u662f \u6298 \u4e0d \u65ad \u7684 \u7ebf ( ",
                "cite_spans": [
                    {
                        "start": 67,
                        "end": 130,
                        "text": "{2, 2, 1, 2; 1, 1, 1, 2, 2; 1, 1, 1, 1, 2, 1, 2; 2, 2, 3, 1, 1}",
                        "ref_id": null
                    },
                    {
                        "start": 236,
                        "end": 237,
                        "text": "(",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Generated Lyric of ChipSong",
                "sec_num": null
            },
            {
                "text": "\u6709 \u79cd \u604b \u7231 \u7684 \u9999 \u5473 (wei) \u4f1a \u8ba9 \u6211 \u65e0 \u6cd5 \u5165 \u7761 (shui) \u4e0e \u4f60 \u770b \u8fd9 \u6708 \u8272 \u591a \u59a9 \u5a9a (mei) \u5982 \u540c \u5973 \u5b69 \u68a6 \u5e7b \u822c \u7684 \u7f8e (mei)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Generated Lyric of ChipSong",
                "sec_num": null
            },
            {
                "text": "Figure 1: Our ChipSong system generates lyrics based on the preset template including length formats, trigger words, and rhymes. The blue box shows an example template, and the green box shows the generated lyrics.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Generated Lyric of ChipSong",
                "sec_num": null
            },
            {
                "text": "A minimalist generation mode of ChipSong is shown in the orange box, which extracts the format and rhyme of the actual lyric to imitatively generate a lyric. English translations of Chinese are provided in parentheses.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Generated Lyric of ChipSong",
                "sec_num": null
            },
            {
                "text": "and a sequence of numbers to tell the length of each lyric line or even each word (a combination of Chinese characters) in a line for matching melody length. The creator can also choose rhyme for the last character in each line from Chinese 14-rhyme 1 groups. Moreover, a minimalist generation mode is provided, where the creator only has to input trigger words and an actual lyric he/she is interested in, then ChipSong will extract the lyric's format and rhyme pattern, and generate a new lyric according to the input trigger words and the extracted format and rhyme, thus fully imitating the original lyric for making a cover song version.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Generated Lyric of ChipSong",
                "sec_num": null
            },
            {
                "text": "To ensure the relevance of generated lyrics with the above controlling attributes, following efforts are made in this paper: 1) A large corpus of 848K Chinese lyrics are gathered, and tailored according to proper lengths for short-video display. 2) A twostage sampling strategy is designed to produce a large number of potential trigger words from lyrics themselves without human annotation, and an autoregressive language model is self-supervisedly trained to complete the whole lyric sequence according to partially-observed trigger words, thus stimulating users' open-ended inspiration inputs.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Generated Lyric of ChipSong",
                "sec_num": null
            },
            {
                "text": "3) Both explicit and implicit control methods are proposed to arrange the format of sentenceand word-level length respectively, where sen-1 About Chinese 14-rhyme tence length is controlled via explicit character sets, and word length is controlled via a welldesigned implicit Begin-Internal-End (BIE) wordgranularity embedding sequence with its guided attention mechanism. 4) A strategy of reverse order training & shielding decoding is designed to learn a reverse language model, guaranteeing fluent text generation following rhyme control, inspired by the observation that, during lyric creation, humans usually first determine which word to use in the rhyming position of a sentence and then create the rest of that sentence based on the rhyming word. Experimentally, both automatic and human evaluations demonstrate that our ChipSong system generates fluent lyrics with high consistency to predetermined control conditions.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Generated Lyric of ChipSong",
                "sec_num": null
            },
            {
                "text": "In summary, oriented to the actual demands of musical short-video creators, we develop Chip-Song, a controllable lyric generation system, which can achieve fine-grained control over lyric generation by the proposed control methods for trigger words, format and rhyme. Especially, to the best of our knowledge, ChipSong is the first lyric generation system that can precisely control the wordlevel length format.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Generated Lyric of ChipSong",
                "sec_num": null
            },
            {
                "text": "Recent lyric generation works can be broadly categoried into three groups according to their cared artistic genres: 1) hip-pop generation, creating hippop lyrics with distinctive rhymes and rhythms constrains (Manjavacas et al., 2019; Nikolov et al., 2020; Xue et al., 2021) ; 2) poetry generation, creating some special text paradigms, such as Shakespeare's Sonnet (Oliveira et al., 2017; Li et al., 2020) , Chinese Classical Poetry (Guo et al., 2019; Hu and Sun, 2020; Li et al., 2020) , and Chinese Couplet (Yan et al., 2016) , etc; 3) popular song generation, creating full-text lyrics (Watanabe et al., 2018; Zhu et al., 2018; Lee et al., 2019; Fan et al., 2019; Zhang et al., 2020a; Sheng et al., 2021) or polishing draft lyrics (Zhang et al., 2020a) for popular songs. Our ChipSong is actually a lyric generation system within the third group, and especially for Chinese popular songs. Moreover, different from previous lyric generation works, which were mostly model-oriented for natural paragraphs generation and excluded explicit user profiles from practical application scenarios, Chip-Song customizes functions to generate lyrics for users from practical short-video scenario with re-Lyric: \u5b64\u5355\u7684\u4e0b\u96e8\u5929\uff0c\u800c\u4f60\u5df2\u8d70\u8fdc\u3002 ( Lonely rainy day, and you have gone far away. )",
                "cite_spans": [
                    {
                        "start": 209,
                        "end": 234,
                        "text": "(Manjavacas et al., 2019;",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 235,
                        "end": 256,
                        "text": "Nikolov et al., 2020;",
                        "ref_id": "BIBREF12"
                    },
                    {
                        "start": 257,
                        "end": 274,
                        "text": "Xue et al., 2021)",
                        "ref_id": "BIBREF20"
                    },
                    {
                        "start": 366,
                        "end": 389,
                        "text": "(Oliveira et al., 2017;",
                        "ref_id": "BIBREF13"
                    },
                    {
                        "start": 390,
                        "end": 406,
                        "text": "Li et al., 2020)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 434,
                        "end": 452,
                        "text": "(Guo et al., 2019;",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 453,
                        "end": 470,
                        "text": "Hu and Sun, 2020;",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 471,
                        "end": 487,
                        "text": "Li et al., 2020)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 510,
                        "end": 528,
                        "text": "(Yan et al., 2016)",
                        "ref_id": "BIBREF21"
                    },
                    {
                        "start": 590,
                        "end": 613,
                        "text": "(Watanabe et al., 2018;",
                        "ref_id": "BIBREF19"
                    },
                    {
                        "start": 614,
                        "end": 631,
                        "text": "Zhu et al., 2018;",
                        "ref_id": "BIBREF24"
                    },
                    {
                        "start": 632,
                        "end": 649,
                        "text": "Lee et al., 2019;",
                        "ref_id": "BIBREF7"
                    },
                    {
                        "start": 650,
                        "end": 667,
                        "text": "Fan et al., 2019;",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 668,
                        "end": 688,
                        "text": "Zhang et al., 2020a;",
                        "ref_id": "BIBREF22"
                    },
                    {
                        "start": 689,
                        "end": 708,
                        "text": "Sheng et al., 2021)",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 735,
                        "end": 756,
                        "text": "(Zhang et al., 2020a)",
                        "ref_id": "BIBREF22"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "E0 \u5929 E[O] E\u5929 E1 \u96e8 E[O] E\u96e8 E2 [START] E[C] E[START] E3 [6] E[B-3] E[6] E4 \u5929 E[I-3] E\u5929 E5 \u4e0b E[E-3] E\u4e0b E6 \u7684 E[B-1] E\u7684 E7 \u5355 E[B-2] E\u5355 E8 \u5b64 E[E-2] E\u5b64 E9 [SEP] E[S] E[SEP] E10 [5] E[C] E[5] E11 \u8fdc E[B-2] E\u8fdc E12 \u8d70 E[E-2] E\u8d70 E13 \u5df2 E[B-1] E\u5df2 E14 \u4f60 E[B-1] E\u4f60 E15 \u800c E[B-1] E\u800c E16 [END] E[O] E[END] \u96e8 [START] [6] \u5929 \u96e8 \u4e0b \u7684 \u5355 \u5b64 [SEP]",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "[5]",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "\u8fdc \u8d70 \u5df2 \u4f60 \u800c \u96e8 \u5929 [START]",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "[6]",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "\u5b64 \u7684 \u4e0b \u96e8 \u5929 [SEP]",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "[5] alistic demands regarding length format, trigger word, and rhyme, simultaneously.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "\u800c \u4f60 \u5df2 \u8d70 \u8fdc [END] E17 \u96e8 E[O] E\u96e8 \u5355 [END]",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "Furthormore, detailed comparison between previous lyric generation works and ChipSong are conducted as follows, from implementation view. First, when it comes to length format control of lyrics, we only notice works (Shen et al., 2019; Li et al., 2020) with sentence-level length control, and no work currently with word-level length control. Second, most of previous works lacked sufficient abilities to deal with open-ended inputs to guide lyric content (Potash et al., 2015 (Potash et al., , 2018 , as a result of the shortage of annotated training data. Fan et al. (2019) and Lu et al. (2019) regarded the user input as the first sentence and generate a continuation of lyric, but it tends to deviate from the initial input as the continuation progresses. Although, Zhang et al. (2020a) designed an interactive lyric creation system to handle the open-ended inputs, as a demo description work, it did not release sufficient implementation details and experimental evaluations. Third, in consideration of the creator's demand for rhyme control, previous works employed various rhyme modeling methods: Nikolov et al. (2020) selected output words from the a list of candidate rhyming words at the rhyming position, while forcibly adding a rhyming word could result in incoherent text in the rhyming position; SongNet (Li et al., 2020) proposed a rigid format control method to realize the rhyme modeling for Chinese lyrics; The recent DeepRapper (Xue et al., 2021) focused on continuous N-gram rhyme and rhythm modeling for rap generation, while we work on unigram rhyme control for popular songs.",
                "cite_spans": [
                    {
                        "start": 456,
                        "end": 476,
                        "text": "(Potash et al., 2015",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 477,
                        "end": 499,
                        "text": "(Potash et al., , 2018",
                        "ref_id": "BIBREF16"
                    },
                    {
                        "start": 558,
                        "end": 575,
                        "text": "Fan et al. (2019)",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 580,
                        "end": 596,
                        "text": "Lu et al. (2019)",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 770,
                        "end": 790,
                        "text": "Zhang et al. (2020a)",
                        "ref_id": "BIBREF22"
                    },
                    {
                        "start": 1318,
                        "end": 1335,
                        "text": "(Li et al., 2020)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 1447,
                        "end": 1465,
                        "text": "(Xue et al., 2021)",
                        "ref_id": "BIBREF20"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "2"
            },
            {
                "text": "As shown in Figure 2 , a Transformer-based autoregressive language model architecture is adopted as the backbone of ChipSong for lyric generation. And by modifying the internal model structure and utilizing processed external feature inputs, we apply the modeling of control conditions for length format, trigger word, and rhyme. In the subsequent section arrangement, we first describe the control condition inputs for ChipSong, and then describe the proposed condition control methods in detail.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 12,
                        "end": 20,
                        "text": "Figure 2",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Overview",
                "sec_num": "3.1"
            },
            {
                "text": "As shown in Figure 1 , the user specifies the conditional templates to formulate the lyric generation, and ChipSong generates the lyrics that meet the corresponding control conditions. Trigger word: enter a few words that are separated by \";\" to render the lyric content. Format: enter each line length and each word (i.e., a combination of Chinese characters) length of each line in the lyric, where sentences' lengths are separated by \";\", and words' lengths are separated by \",\". For example, enter \"7; 7; 9; 2, 2, 3, 1, 1\" means to generate four lines of the lyric with lengths of 7, 7, 9, and 9, respectively, and in the last sentence, the word length arrangement is specified as 2,2,3,1,1. Users can also not specify the full template, and the system retrieves similar templates to complement the length format; or directly input a lyric, and the system extracts the length format for imitative writing. Rhyme: enter the rhyme of the last word in each sentence. Rhymes are separated by \";\". For example, input \"ui;ui;ui;ui\", the generated lyrics keep the rhyme of the last word of each sentence match with \"ui\". Users can also not specify rhymes, and the system freely generates sentences.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 12,
                        "end": 20,
                        "text": "Figure 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "User Inputs",
                "sec_num": "3.2"
            },
            {
                "text": "An explicit character set C S is designed to control the length of each line of the lyrics, just like \"[CLS]\" and \"[SEP ]\" in BERT (Devlin et al., 2018) , which is constructed as follows:",
                "cite_spans": [
                    {
                        "start": 131,
                        "end": 152,
                        "text": "(Devlin et al., 2018)",
                        "ref_id": "BIBREF0"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Sentence-level Format Control",
                "sec_num": "3.3"
            },
            {
                "text": "..., [ST ART ], [4], a 1 , a 2 , a 3 , a 4 , [SEP ], [5], a 1 , a 2 , a 3 , a 4 , a 5 , [EN D]",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Sentence-level Format Control",
                "sec_num": "3.3"
            },
            {
                "text": "where [SEP ] is the interline delimiter, a i is the ith character of a sentence, [ST ART ] and [EN D] are the beginning and end of a lyric, [4] and [5] represent that the next sentence length is 4 and 5, respectively. We assign 50 learnable character embeddings",
                "cite_spans": [
                    {
                        "start": 148,
                        "end": 151,
                        "text": "[5]",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Sentence-level Format Control",
                "sec_num": "3.3"
            },
            {
                "text": "{[1], [2], [3], ..., [50]}",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Sentence-level Format Control",
                "sec_num": "3.3"
            },
            {
                "text": "to C S to represent the line length from 1 to 50, which are embedded in the lyric sequence as explicit supervisory information for training. The control character is placed after the sentence separator [SEP ] and before the beginning of the sentence to learn the correspondence between the control symbol and the sentence length. During prediction, the format control character entered by the user is inserted after the initial [ST ART ] token and the generated [SEP ] token to achieve the length control of lines.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Sentence-level Format Control",
                "sec_num": "3.3"
            },
            {
                "text": "Beyond the sentence-level format control, wordlevel format control arranges lyrics in a more refined way, benefiting fine-grained lyrics' adjustment or imitative writing lyrics. Unlike sentencegranularity format control, the explicit character control strategy makes input too verbose, and the unidirectional masked self-attention of autoregressive language model cannot model the uninput control symbols, which is difficult to reconcile and arrange the fixed-length words in fixed-length sentences. Therefore, we propose an implicit control method, Begin-Internal-End (BIE) wordgranularity embedding with its guided attention mechanism to adjust the word-level length format.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Word-level Format Control",
                "sec_num": "3.4"
            },
            {
                "text": "As shown in Figure 2 (left), each lyric token is added with a learnable embedding to record word length information 2 , just like position embeddings, that is \"Begin-Internal-End (BIE) word-granularity embedding\". The design of BIE embedding symbols is inspired by the sequence tagging task (Huang et al., 2015) . We use Note that the lyric embedding sequence is not aligned with the BIE embedding sequence; it corresponds to the BIE embedding sequence shifted to the left. This setting aims to help the model learn to predict the word length of the next token for lyric sequence, and to learn when to stop sentence generation and feed new control characters.",
                "cite_spans": [
                    {
                        "start": 291,
                        "end": 311,
                        "text": "(Huang et al., 2015)",
                        "ref_id": "BIBREF5"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 12,
                        "end": 20,
                        "text": "Figure 2",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "BIE Word-granularity Embedding",
                "sec_num": "3.4.1"
            },
            {
                "text": "The BIE word-granularity embeddings can only perceive the word length of the next lyric token in advance, but cannot predict the farther distance due to the unidirectional masked attention in autoregressive language model. When sentence length is fixed, BIE embeddings are difficult to reconcile and arrange the length of each word reasonably. Therefore, we design a word-granularity attention mechanism, which is guided by BIE embeddings, to perceive the word length information of all positions for the current token.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Word-granularity Attention",
                "sec_num": "3.4.2"
            },
            {
                "text": "Concretely, the special decoder block with the word-granularity attention is placed at the bottom of the ChipSong model, on top of which the standard Transformer decoder block is stacked. The detailed structure is shown on the right of Figure 2 (right). The calculation process is as follows:",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 236,
                        "end": 244,
                        "text": "Figure 2",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Word-granularity Attention",
                "sec_num": "3.4.2"
            },
            {
                "text": "\u00ca Fw = E Fw + E P (1) C w = Sof tmax(X W 1\u00ca Fw )\u00ca Fw (2) X out = [X ; C w ]W 2 + X (3)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Word-granularity Attention",
                "sec_num": "3.4.2"
            },
            {
                "text": "First, the BIE embedding E Fw and position embedding E P are added to obtain\u00ca Fw , so that the BIE embedding sequence carries global position information. Then, after passing through the masked self-attention layer, a word-granularity attention layer is designed to compute the attention weights of the contextual lyric embeddings X to the BIE embeddings\u00ca Fw , where a bilinear attention is applied, so as to obtain the contextual embedding C w recording global words length for each lyric token. Finally, the contextual lyric embedding X and global word-length embedding C w are concatenated and pass through a linear layer to obtain fusion representation X out , and a residual connection (He et al., 2016) is added to enhance the memory of the original input lyric embedding features X in decoder block. X out is further modeled in subsequent Transformer decoder blocks.",
                "cite_spans": [
                    {
                        "start": 691,
                        "end": 708,
                        "text": "(He et al., 2016)",
                        "ref_id": "BIBREF3"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Word-granularity Attention",
                "sec_num": "3.4.2"
            },
            {
                "text": "To produce enough trigger words during training to cover creators' input needs as much as possible, we adopt a two-stage strategy, establishing a candidate word list for each lyric in the first stage, and resampling the candidate list as trigger words during each training epoch in the second stage. Concretely, considering that general keyword extraction methods could result in a low coverage range of trigger words, all nouns, adjectives, and verbs 3 of the lyrics are reserved as the candidate word list after removing the stop words, and the candidate word list also preserves word frequency so that frequent words have a higher probability of being sampled. The number of trigger words sampled is determined according to the number of lyric sentences, and the rules are designed as follows:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Trigger Word Control",
                "sec_num": "3.5"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "k = N sent /2 \u2212 1, N sent <= 12 k = 5, else",
                        "eq_num": "(4)"
                    }
                ],
                "section": "Trigger Word Control",
                "sec_num": "3.5"
            },
            {
                "text": "where k is the number of trigger words and N sent is the number of sentences. As shown in Figure 2 , after building trigger words-lyric pair data, the trigger words sequence 3 POS-tagging information is obtained by the Jieba tool. and lyric sequence are simply spliced and fed into the language model for training, guiding model self-supervisedly complements the lyric sequence according to partially-observed trigger word sequence, where trigger words are also separated by token \"[SEP ]\". When prediction, feed trigger words, and the model complements the subsequent lyric part.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 90,
                        "end": 98,
                        "text": "Figure 2",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Trigger Word Control",
                "sec_num": "3.5"
            },
            {
                "text": "A paradigm of reverse order training and shielding decoding is designed to control rhymes. During training, we process the training data as intersentence normal order and intra-sentence reverse order, as shown in Figure 2 . For example, when the original lyric sequence is \"...,",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 213,
                        "end": 221,
                        "text": "Figure 2",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Rhyme Control",
                "sec_num": "3.6"
            },
            {
                "text": "[3], x 1 , x 2 , x 3 , [SEP ], [4], x 4 , x 5 , x 6 , x 7 , [SEP ], ...\", it is transformed into \"..., [3], x 3 , x 2 , x 1 , [SEP ], [4], x 7 , x 6 , x 5 , x 4 , [SEP ], ...\".",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Rhyme Control",
                "sec_num": "3.6"
            },
            {
                "text": "In the same way, the BIE word-granularity embedding sequence is accordingly processed to keep consistency. The reverse order sentences input enables to learn a reverse language model, so that rhyming position is predicted first in a sentence, and the subsequent predictions coordinate the rhyming word for protecting the text fluency from being affected.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Rhyme Control",
                "sec_num": "3.6"
            },
            {
                "text": "During prediction, the input pattern of \"intersentence normal order, intra-sentence reverse order\" is maintained; that is, the last word to rhyme in each sentence is predicted first, and then the rest of the sentence is predicted in reverse order. Then, a decoding shielding strategy is adopted to control the prediction of rhyming words. According to the Chinese 14-rhyme scheme, we build a rhyme dictionary whose key is the rhyme and value is the words that the rhyme matches. At the position of the last word in the sentence, the rhyme dictionary is queried according to the input rhyme, and the softmax output values corresponding to all non-rhymed words are reduced, so that the model selects outputs from rhyming words based on predicted probability distributions.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Rhyme Control",
                "sec_num": "3.6"
            },
            {
                "text": "We prepare a large lyric corpus to train the Chip-Song model. The lyric data is constructed with reference to ChineseLyrics 4 , and we gather 848K Chinese popular lyrics, where the number of lyrics sentences is 27,181K, the average lyric length is 253 (excluding punctuation), the average number of sentences is 32, and the average length of each sentence is 8. To fit short-video durations, we tailor the lyrics into small segments of 8, 10, 12, 14, or 16 sentences in length, which considers that 8 to 12 lines of lyrics are generally required for a 60 sec. short-video song. Lyric tailoring also increases first line diversity. In addition, unsegmented lyrics are also incorporated as training data to preserve semantic integrity and learn long-range dependencies.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Data Preparation and Processing",
                "sec_num": "4.1"
            },
            {
                "text": "To comprehensively evaluate the proposed system, we formulate multiple conditional templates for generation, which are provided in https://github.com/korokes/chipsong. Concretely, we build 15 groups of trigger words from users and construct 500 format templates from the format library, where the formats are extracted based on actual lyrics. Each format template is randomly assigned a rhyme pattern and a group of trigger words as a complete evaluation template for lyric generation. In addition, we sample the original lyrics corresponding to the format template to generate another group of trigger words as described in Section 3.5, which are only used to evaluate the effect of trigger words guiding content. The lyrics corresponding to these evaluation templates are eliminated from the training data.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation Templates",
                "sec_num": "4.2"
            },
            {
                "text": "For automatic evaluations, each template generates 20 samples, and a total of 500\u00d720=10,000 samples are finally generated for evaluation. For human evaluations, 200 samples of 10 conditional templates are randomly reserved for evaluation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation Templates",
                "sec_num": "4.2"
            },
            {
                "text": "We use a 8-head, 8-layer, 512-dimensional Transformer to build the ChipSong model (39.5M). Actually, larger hidden layer dimensions can make the model perform better. For training, the Adam optimizer (Kingma and Ba, 2014) is used with an initial learning rate of 1.5e-4. The model is trained for about 3.5 days on two GTX 2080ti GPUs with a batch size of 8. The training data combines the segmented lyrics data and the raw lyrics data, eliminating lyrics corresponding to evaluation templates. Due to the large corpus and the duplication of segmented lyrics and original lyrics, we do not set too many training epochs, and set the training epochs to 3. Owing to sufficient lyrics gathering, we do not use the pretraining strategy. For prediction, we use TopK decoding with a sampling value of 8.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training settings",
                "sec_num": "4.3"
            },
            {
                "text": "Automatic Evaluation We use the trained lyric generation models on our corpus to evaluate the perplexity (PPL) and use the Distinct (MA-D1,D2, MI-D1,D2) metrics to evaluate the diversity of generated lyric texts. Moreover, we design the following metrics to evaluate the proposed conditional control ability: 1) sentencelevel format accuracy (SA), the percentage of generated sentences with correct length. 2) wordlevel format accuracy (WA), the percentage of generated sentences whose words length arrangement is exactly the same as the label. 3) rhyme accuracy (RA), the percentage of generated sentences with correct rhymes. 4) word length accuracy (WA-N ), the percentage of generated words containing N Chinese characters that are correct in position and length; as the Chinese word lengths are basically within 4, we evaluate the control accuracy of 1 to 5 word length. 5) trigger word effect, we first use trigger words extracted from the original lyrics to generate samples, and then use BLEU (Papineni et al., 2002) to compare content similarity between the original lyric and the generated lyric to evaluate the relevance of trigger words and contents indirectly. Human Evaluation We recruit three postgraduates engaged in audio and music fields to score the generated lyrics on fluency, relevance, and listenability: (1) fluency (F), the quality of the generated lyrics, whether they are smooth, grammatical, and whether there are ill-formed sentences; (1=Bad to 3=Good). (2) relevance (R), the degree of relevance of the trigger words and the lyric content; (1=Bad to 3=Good). (3) listenability (L) (Watanabe et al., 2018) , as lyrics, are the positions of words, lines, and segments natural? (1=Bad to 3=Good).",
                "cite_spans": [
                    {
                        "start": 1001,
                        "end": 1024,
                        "text": "(Papineni et al., 2002)",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 1611,
                        "end": 1634,
                        "text": "(Watanabe et al., 2018)",
                        "ref_id": "BIBREF19"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation Metrics",
                "sec_num": "4.4"
            },
            {
                "text": "For reference, a standard Transformer-based autoregressive language model (ALM) is trained as a baseline, with the same configuration as ChipSong. We also respectively train ALM-F, ALM-T, and ALM-R for observing a single conditional control effect, where the proposed control strategies of format (F), trigger word (T), and rhyme (R) are individually modeled into ALM for training (modeling all the three conditions into ALM is equal to ChipSong), with the same configuration as ChipSong. In addition, we compare ChipSong with SongNet (Li et al., 2020) that proposes a rigid format and rhyme control method. Due to differences in model settings and usage data, we use our data to retrain the SongNet model. For the models that lack the trigger word mechanism, the input is used as the first sentence and let the model continue to write like previous methods. Table 1 and Table 2 show the experimental results of different models under each evaluation metrics. As can be seen from Table 1 , ChipSong demonstrates good conditional control ability on format and rhyme, where the sentence-granularity format accuracy (SA) is 98.54%, the word-level format accuracy (WA) is 86.64%, and the rhyme accuracy (RA) is 98.56%. Since SongNet's rhyming modeling method cannot actively select rhymes and requires specific rhyming corpus for training, it isn't easy to exert its role in rhyming modeling to achieve good rhyming accuracy. ChipSong also demonstrates better PPL and generative diversity. Interestingly, the reverse order training of rhyme control (No.5) has little impact on the model PPL, indicating that the reverse language model still learns language rules. As shown in Table 2 , Chip-Song embodies better content control capabilities via the trigger word mechanism. It can also be observed that a single conditional control model (No.3, 4, 5) generally performs better on its corresponding control condition because there are no constraints of the other two conditional controls. For example, in Table 1 , without the constraints of trigger words and rhyme decoding shielding, ALM-F can focus more on controlling length format and obtain higher WA and WA-N scores, even achieving 92.72 on WA; in Table 2 , without format and rhyme constraints, ALM-T has more opportunities to generate content related to trigger words for better BLEU scores.",
                "cite_spans": [
                    {
                        "start": 535,
                        "end": 552,
                        "text": "(Li et al., 2020)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 1833,
                        "end": 1839,
                        "text": "(No.3,",
                        "ref_id": null
                    },
                    {
                        "start": 1840,
                        "end": 1842,
                        "text": "4,",
                        "ref_id": null
                    },
                    {
                        "start": 1843,
                        "end": 1845,
                        "text": "5)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 859,
                        "end": 878,
                        "text": "Table 1 and Table 2",
                        "ref_id": "TABREF4"
                    },
                    {
                        "start": 980,
                        "end": 987,
                        "text": "Table 1",
                        "ref_id": "TABREF4"
                    },
                    {
                        "start": 1672,
                        "end": 1679,
                        "text": "Table 2",
                        "ref_id": "TABREF5"
                    },
                    {
                        "start": 1999,
                        "end": 2006,
                        "text": "Table 1",
                        "ref_id": "TABREF4"
                    },
                    {
                        "start": 2199,
                        "end": 2206,
                        "text": "Table 2",
                        "ref_id": "TABREF5"
                    }
                ],
                "eq_spans": [],
                "section": "Baselines",
                "sec_num": "4.5"
            },
            {
                "text": "To further analyze the effect of each proposed conditional control, we respectively remove the conditional control of 1) word-level format control (WC); 2) sentence-level format control (SC); 3) trigger word control (TC); 4) rhyme control (RC) to train the ablation models. Two internal structures of WC, 6) BIE embedding in WC (WC-Emb) and 7) word-granularity attention in WC (WC-Att), are also ablated for evaluation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Ablation",
                "sec_num": "5.2"
            },
            {
                "text": "The experimental results are shown in Table  3 and Table 4 . As can be observed from the tables, format (No.2, 3, 4, 5) and rhyme control (No.7) increase the diversity of generation while trigger word control (No.6) decreases the diversity. The modeling of word-level format control, WC, WC-Att, and WC-Emb, plays an important role in reducing the PPL. When the modeling of WC, SC, RC, or TC is removed separately, the accuracy of the corresponding evaluations, SA, WA, RA, or BLEU, is obviously reduced, indicating the effectiveness of the proposed control methods. Although WC-Att and WC-Emb both play a positive role in wordlevel format control, trigger word control (TC) and rhyme control (RC) have a negative effect on word- level format control, where the scores of WA and WA-N rise when TC or RC is ablated. Table 5 shows the experimental results of human evaluation in fluency (F), relevance (R), and Listenability (L). On the whole, the fluency scores of the models are not much different (2.46-2.56 points), which is attributed to sufficient corpus for training. ChipSong scores far higher than baselines (No.1,2) in relevance evaluation due to the modeling of trigger word mechanism. It can also be seen that when the control of rhyme or format (No.4,5,7) is lifted, the relevance is improved; we conjecture that the model has more opportunities to generate related content when the format or rhyme is not restricted. ChipSong w/o RC (No.7) gain the best average score; this is because our human evaluation does not consider the evaluation of lyric rhymes. The listenability scores of the models without format control (No.1,5) drop from nearly full marks, because the free generation is prone to generate too short or too long sentences, or two consecutive sentences with large length differences, which is not conducive to fit songs.",
                "cite_spans": [
                    {
                        "start": 104,
                        "end": 110,
                        "text": "(No.2,",
                        "ref_id": null
                    },
                    {
                        "start": 111,
                        "end": 113,
                        "text": "3,",
                        "ref_id": null
                    },
                    {
                        "start": 114,
                        "end": 116,
                        "text": "4,",
                        "ref_id": null
                    },
                    {
                        "start": 117,
                        "end": 119,
                        "text": "5)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 38,
                        "end": 58,
                        "text": "Table  3 and Table 4",
                        "ref_id": "TABREF7"
                    },
                    {
                        "start": 815,
                        "end": 822,
                        "text": "Table 5",
                        "ref_id": "TABREF9"
                    }
                ],
                "eq_spans": [],
                "section": "Ablation",
                "sec_num": "5.2"
            },
            {
                "text": "We count the sampled trigger words and the sampled trigger words without repetition in training, which aims to observe the trigger word coverage. The results are shown in Figure 3 . Due to the two-stage strategies, a large number of trigger words are produced for training. The number of sampling words is 3.98e7, and is only 3.16e5 after deduplication. As shown in Figure 3 , as the extracted trigger words increase, new trigger words increase very little, which shows that our method covers a relatively comprehensive range of trigger words to handle out-of-distribution and cover the general input needs for users.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 171,
                        "end": 179,
                        "text": "Figure 3",
                        "ref_id": null
                    },
                    {
                        "start": 366,
                        "end": 374,
                        "text": "Figure 3",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Trigger word coverage",
                "sec_num": "5.4"
            },
            {
                "text": "As shown in Figure 4 , we enumerate some generated lyrics of the ChipSong system in several scenarios: 1) a Chinese Hanmai song with a specific format; 2) customizing format according to a song; 3) imitative writing a lyric for a cover song version, where the sentence-and word-level format are extracted from the original lyric, Han Hong's \"Qingchun\". For the first case in each template in the figure, we also provide the English translation for understanding the generated lyrics. For the first template in a given word-granularity length format, we provide a human-annotated word segmentation boundary with the green vertical line |. As can be seen from the generated results, ChipSong can fine-grainly adjust the generated lyrics' format to adapt to any song, render the content guided",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 12,
                        "end": 20,
                        "text": "Figure 4",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Case Analysis",
                "sec_num": "5.5"
            },
            {
                "text": "Words length is obtained by the Chinese text segmentation tool, Jieba.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "https://github.com/dengxiuqi/ChineseLyrics",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [
            {
                "text": " Figure 3 : The count of sampled trigger words. Three figures use different scales. The total sampling number is 3.98e7, and is only 3.16e5 after deduplication. New trigger words increase very little as the increase of sampling trigger words, indicating that sampled trigger words cover a relatively comprehensive range to handle input needs.Sentence granularity: 7;7;7;7;7;7;7;7 Word granularity: T: In love R: an, uan, ian Stars, river, sun, and moon lingering together The breeze blows, took away the years Do not need to ask depth of love in the world At the end of the road, we will meet Look at the high mountains and far water Time travels between clouds and water Carve an ageless appearance of beauty To the ends of the earth with you Figure 4 : Cases of lyrics generated by ChipSong. The generated results are marked in blue font. The rhymes are marked in grey font. T: Trigger words; R: Rhyme. For the fist template, the green vertical line | is used to manually annotate word segmentation boundaries. For the first case in each template, the English translation is given.by open-ended trigger words, and maintain the rhyme. More generated cases are provided in https://github.com/korokes/chipsong.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 1,
                        "end": 9,
                        "text": "Figure 3",
                        "ref_id": null
                    },
                    {
                        "start": 744,
                        "end": 752,
                        "text": "Figure 4",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "annex",
                "sec_num": null
            },
            {
                "text": "In this work, we develop ChipSong, a lyric generation system, to assist musical short-video creators for Chinese popular song creation. ChipSong fine-grainly adjusts lyric generation to meet the creator's needs in various scenarios via the proposed strategies of sentence-and word-level format control, trigger word control, and rhyme control.In the future, we would like to consider melody generation and other attributions control of lyrics.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "6"
            }
        ],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Bert: Pre-training of deep bidirectional transformers for language understanding",
                "authors": [
                    {
                        "first": "Jacob",
                        "middle": [],
                        "last": "Devlin",
                        "suffix": ""
                    },
                    {
                        "first": "Ming-Wei",
                        "middle": [],
                        "last": "Chang",
                        "suffix": ""
                    },
                    {
                        "first": "Kenton",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    },
                    {
                        "first": "Kristina",
                        "middle": [],
                        "last": "Toutanova",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1810.04805"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understand- ing. arXiv preprint arXiv:1810.04805.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "A hierarchical attention based seq2seq model for Chinese lyrics generation",
                "authors": [
                    {
                        "first": "Haoshen",
                        "middle": [],
                        "last": "Fan",
                        "suffix": ""
                    },
                    {
                        "first": "Jie",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Bojin",
                        "middle": [],
                        "last": "Zhuang",
                        "suffix": ""
                    },
                    {
                        "first": "Shaojun",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Jing",
                        "middle": [],
                        "last": "Xiao",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the Pacific Rim International Conference on Artificial Intelligence",
                "volume": "",
                "issue": "",
                "pages": "279--288",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Haoshen Fan, Jie Wang, Bojin Zhuang, Shaojun Wang, and Jing Xiao. 2019. A hierarchical attention based seq2seq model for Chinese lyrics generation. In Pro- ceedings of the Pacific Rim International Conference on Artificial Intelligence, pages 279-288.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Jiuge: A humanmachine collaborative Chinese classical poetry generation system",
                "authors": [
                    {
                        "first": "Zhipeng",
                        "middle": [],
                        "last": "Guo",
                        "suffix": ""
                    },
                    {
                        "first": "Xiaoyuan",
                        "middle": [],
                        "last": "Yi",
                        "suffix": ""
                    },
                    {
                        "first": "Maosong",
                        "middle": [],
                        "last": "Sun",
                        "suffix": ""
                    },
                    {
                        "first": "Wenhao",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Cheng",
                        "middle": [],
                        "last": "Yang",
                        "suffix": ""
                    },
                    {
                        "first": "Jiannan",
                        "middle": [],
                        "last": "Liang",
                        "suffix": ""
                    },
                    {
                        "first": "Huimin",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Yuhui",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Ruoyu",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: System Demonstrations",
                "volume": "",
                "issue": "",
                "pages": "25--30",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/P19-3005"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Zhipeng Guo, Xiaoyuan Yi, Maosong Sun, Wenhao Li, Cheng Yang, Jiannan Liang, Huimin Chen, Yuhui Zhang, and Ruoyu Li. 2019. Jiuge: A human- machine collaborative Chinese classical poetry gen- eration system. In Proceedings of the 57th An- nual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 25-30, Florence, Italy. Association for Computational Lin- guistics.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Deep residual learning for image recognition",
                "authors": [
                    {
                        "first": "Kaiming",
                        "middle": [],
                        "last": "He",
                        "suffix": ""
                    },
                    {
                        "first": "Xiangyu",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Shaoqing",
                        "middle": [],
                        "last": "Ren",
                        "suffix": ""
                    },
                    {
                        "first": "Jian",
                        "middle": [],
                        "last": "Sun",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the IEEE conference on computer vision and pattern recognition",
                "volume": "",
                "issue": "",
                "pages": "770--778",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recog- nition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770- 778.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Generating major types of Chinese classical poetry in a uniformed framework. CoRR, abs",
                "authors": [
                    {
                        "first": "Jinyi",
                        "middle": [],
                        "last": "Hu",
                        "suffix": ""
                    },
                    {
                        "first": "Maosong",
                        "middle": [],
                        "last": "Sun",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jinyi Hu and Maosong Sun. 2020. Generating major types of Chinese classical poetry in a uniformed framework. CoRR, abs/2003.11528.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Bidirectional LSTM-CRF models for sequence tagging",
                "authors": [
                    {
                        "first": "Zhiheng",
                        "middle": [],
                        "last": "Huang",
                        "suffix": ""
                    },
                    {
                        "first": "Wei",
                        "middle": [],
                        "last": "Xu",
                        "suffix": ""
                    },
                    {
                        "first": "Kai",
                        "middle": [],
                        "last": "Yu",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1508.01991"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Zhiheng Huang, Wei Xu, and Kai Yu. 2015. Bidi- rectional LSTM-CRF models for sequence tagging. arXiv preprint arXiv:1508.01991.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Adam: A method for stochastic optimization",
                "authors": [
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Diederik",
                        "suffix": ""
                    },
                    {
                        "first": "Jimmy",
                        "middle": [],
                        "last": "Kingma",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Ba",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1412.6980"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "iComposer: An automatic songwriting system for Chinese popular music",
                "authors": [
                    {
                        "first": "Hsin-Pei",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    },
                    {
                        "first": "Jhih-Sheng",
                        "middle": [],
                        "last": "Fang",
                        "suffix": ""
                    },
                    {
                        "first": "Wei-Yun",
                        "middle": [],
                        "last": "Ma",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics (Demonstrations)",
                "volume": "",
                "issue": "",
                "pages": "84--88",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Hsin-Pei Lee, Jhih-Sheng Fang, and Wei-Yun Ma. 2019. iComposer: An automatic songwriting sys- tem for Chinese popular music. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics (Demonstrations), pages 84-88.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "A diversity-promoting objective function for neural conversation models",
                "authors": [
                    {
                        "first": "Jiwei",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Michel",
                        "middle": [],
                        "last": "Galley",
                        "suffix": ""
                    },
                    {
                        "first": "Chris",
                        "middle": [],
                        "last": "Brockett",
                        "suffix": ""
                    },
                    {
                        "first": "Jianfeng",
                        "middle": [],
                        "last": "Gao",
                        "suffix": ""
                    },
                    {
                        "first": "Bill",
                        "middle": [],
                        "last": "Dolan",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2016. A diversity-promoting ob- jective function for neural conversation models. In Proceedings of the Conference of the North Ameri- can Chapter of the Association for Computational Linguistics: Human Language Technologies.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Rigid formats controlled text generation",
                "authors": [
                    {
                        "first": "Piji",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Haisong",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Xiaojiang",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Shuming",
                        "middle": [],
                        "last": "Shi",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "742--751",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Piji Li, Haisong Zhang, Xiaojiang Liu, and Shuming Shi. 2020. Rigid formats controlled text generation. In Proceedings of the Annual Meeting of the Associa- tion for Computational Linguistics, pages 742-751.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "A syllable-structured, contextuallybased conditionally generation of Chinese lyrics",
                "authors": [
                    {
                        "first": "Xu",
                        "middle": [],
                        "last": "Lu",
                        "suffix": ""
                    },
                    {
                        "first": "Jie",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Bojin",
                        "middle": [],
                        "last": "Zhuang",
                        "suffix": ""
                    },
                    {
                        "first": "Shaojun",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Jing",
                        "middle": [],
                        "last": "Xiao",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the Pacific Rim International Conference on Artificial Intelligence",
                "volume": "",
                "issue": "",
                "pages": "257--265",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Xu Lu, Jie Wang, Bojin Zhuang, Shaojun Wang, and Jing Xiao. 2019. A syllable-structured, contextually- based conditionally generation of Chinese lyrics. In Proceedings of the Pacific Rim International Confer- ence on Artificial Intelligence, pages 257-265.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Generation of hip-hop lyrics with hierarchical modeling and conditional templates",
                "authors": [
                    {
                        "first": "Enrique",
                        "middle": [],
                        "last": "Manjavacas",
                        "suffix": ""
                    },
                    {
                        "first": "Mike",
                        "middle": [],
                        "last": "Kestemont",
                        "suffix": ""
                    },
                    {
                        "first": "Folgert",
                        "middle": [],
                        "last": "Karsdorp",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the International Conference on Natural Language Generation",
                "volume": "",
                "issue": "",
                "pages": "301--310",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Enrique Manjavacas, Mike Kestemont, and Folgert Karsdorp. 2019. Generation of hip-hop lyrics with hierarchical modeling and conditional templates. In Proceedings of the International Conference on Nat- ural Language Generation, pages 301-310.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "Rapformer: Conditional rap lyrics generation with denoising autoencoders",
                "authors": [
                    {
                        "first": "Eric",
                        "middle": [],
                        "last": "Nikola I Nikolov",
                        "suffix": ""
                    },
                    {
                        "first": "Curtis",
                        "middle": [],
                        "last": "Malmi",
                        "suffix": ""
                    },
                    {
                        "first": "Loreto",
                        "middle": [],
                        "last": "Northcutt",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Parisi",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the International Conference on Natural Language Generation",
                "volume": "",
                "issue": "",
                "pages": "360--373",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Nikola I Nikolov, Eric Malmi, Curtis Northcutt, and Loreto Parisi. 2020. Rapformer: Conditional rap lyrics generation with denoising autoencoders. In Proceedings of the International Conference on Nat- ural Language Generation, pages 360-373.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Multilingual extension and evaluation of a poetry generator",
                "authors": [
                    {
                        "first": "Raquel",
                        "middle": [],
                        "last": "Hugo Gon\u00e7alo Oliveira",
                        "suffix": ""
                    },
                    {
                        "first": "Alberto",
                        "middle": [],
                        "last": "Herv\u00e1s",
                        "suffix": ""
                    },
                    {
                        "first": "Pablo",
                        "middle": [],
                        "last": "D\u00edaz",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Gerv\u00e1s",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Natural Language Engineering",
                "volume": "23",
                "issue": "6",
                "pages": "929--967",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Hugo Gon\u00e7alo Oliveira, Raquel Herv\u00e1s, Alberto D\u00edaz, and Pablo Gerv\u00e1s. 2017. Multilingual extension and evaluation of a poetry generator. Natural Language Engineering, 23(6):929-967.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Bleu: A method for automatic evaluation of machine translation",
                "authors": [
                    {
                        "first": "Kishore",
                        "middle": [],
                        "last": "Papineni",
                        "suffix": ""
                    },
                    {
                        "first": "Salim",
                        "middle": [],
                        "last": "Roukos",
                        "suffix": ""
                    },
                    {
                        "first": "Todd",
                        "middle": [],
                        "last": "Ward",
                        "suffix": ""
                    },
                    {
                        "first": "Wei-Jing",
                        "middle": [],
                        "last": "Zhu",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "Proceedings of the annual meeting on association for computational linguistics (ACL)",
                "volume": "",
                "issue": "",
                "pages": "311--318",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei- Jing Zhu. 2002. Bleu: A method for automatic evaluation of machine translation. In Proceedings of the annual meeting on association for computational linguistics (ACL), pages 311-318. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Ghostwriter: Using an LSTM for automatic rap lyric generation",
                "authors": [
                    {
                        "first": "Peter",
                        "middle": [],
                        "last": "Potash",
                        "suffix": ""
                    },
                    {
                        "first": "Alexey",
                        "middle": [],
                        "last": "Romanov",
                        "suffix": ""
                    },
                    {
                        "first": "Anna",
                        "middle": [],
                        "last": "Rumshisky",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Proceedings of the Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "1919--1924",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Peter Potash, Alexey Romanov, and Anna Rumshisky. 2015. Ghostwriter: Using an LSTM for automatic rap lyric generation. In Proceedings of the Con- ference on Empirical Methods in Natural Language Processing, pages 1919-1924.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Evaluating creative language generation: The case of rap lyric ghostwriting",
                "authors": [
                    {
                        "first": "Peter",
                        "middle": [],
                        "last": "Potash",
                        "suffix": ""
                    },
                    {
                        "first": "Alexey",
                        "middle": [],
                        "last": "Romanov",
                        "suffix": ""
                    },
                    {
                        "first": "Anna",
                        "middle": [],
                        "last": "Rumshisky",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the Second Workshop on Stylistic Variation",
                "volume": "",
                "issue": "",
                "pages": "29--38",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Peter Potash, Alexey Romanov, and Anna Rumshisky. 2018. Evaluating creative language generation: The case of rap lyric ghostwriting. In Proceedings of the Second Workshop on Stylistic Variation, pages 29- 38.",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Controlling sequence-tosequence models-a demonstration on neural-based acrostic generator",
                "authors": [
                    {
                        "first": "Pei-Lun",
                        "middle": [],
                        "last": "Liang-Hsin Shen",
                        "suffix": ""
                    },
                    {
                        "first": "Chao-Chung",
                        "middle": [],
                        "last": "Tai",
                        "suffix": ""
                    },
                    {
                        "first": "Shou-De",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Lin",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the Conference on Empirical Methods in Natural Language Processing and the International Joint Conference on Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "43--48",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Liang-Hsin Shen, Pei-Lun Tai, Chao-Chung Wu, and Shou-De Lin. 2019. Controlling sequence-to- sequence models-a demonstration on neural-based acrostic generator. In Proceedings of the Confer- ence on Empirical Methods in Natural Language Processing and the International Joint Conference on Natural Language Processing, pages 43-48.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "SongMASS: Automatic song writing with pre-training and alignment constraint",
                "authors": [
                    {
                        "first": "Zhonghao",
                        "middle": [],
                        "last": "Sheng",
                        "suffix": ""
                    },
                    {
                        "first": "Kaitao",
                        "middle": [],
                        "last": "Song",
                        "suffix": ""
                    },
                    {
                        "first": "Xu",
                        "middle": [],
                        "last": "Tan",
                        "suffix": ""
                    },
                    {
                        "first": "Yi",
                        "middle": [],
                        "last": "Ren",
                        "suffix": ""
                    },
                    {
                        "first": "Wei",
                        "middle": [],
                        "last": "Ye",
                        "suffix": ""
                    },
                    {
                        "first": "Shikun",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Tao",
                        "middle": [],
                        "last": "Qin",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "Proceedings of the AAAI Conference on Artificial Intelligence",
                "volume": "35",
                "issue": "",
                "pages": "13798--13805",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Zhonghao Sheng, Kaitao Song, Xu Tan, Yi Ren, Wei Ye, Shikun Zhang, and Tao Qin. 2021. SongMASS: Automatic song writing with pre-training and align- ment constraint. In Proceedings of the AAAI Con- ference on Artificial Intelligence, volume 35, pages 13798-13805.",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "A melody-conditioned lyrics language model",
                "authors": [
                    {
                        "first": "Kento",
                        "middle": [],
                        "last": "Watanabe",
                        "suffix": ""
                    },
                    {
                        "first": "Yuichiroh",
                        "middle": [],
                        "last": "Matsubayashi",
                        "suffix": ""
                    },
                    {
                        "first": "Satoru",
                        "middle": [],
                        "last": "Fukayama",
                        "suffix": ""
                    },
                    {
                        "first": "Masataka",
                        "middle": [],
                        "last": "Goto",
                        "suffix": ""
                    },
                    {
                        "first": "Kentaro",
                        "middle": [],
                        "last": "Inui",
                        "suffix": ""
                    },
                    {
                        "first": "Tomoyasu",
                        "middle": [],
                        "last": "Nakano",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
                "volume": "1",
                "issue": "",
                "pages": "163--172",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kento Watanabe, Yuichiroh Matsubayashi, Satoru Fukayama, Masataka Goto, Kentaro Inui, and To- moyasu Nakano. 2018. A melody-conditioned lyrics language model. In Proceedings of the Con- ference of the North American Chapter of the Asso- ciation for Computational Linguistics: Human Lan- guage Technologies, Volume 1 (Long Papers), pages 163-172.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "DeepRapper: Neural rap generation with rhyme and rhythm modeling",
                "authors": [
                    {
                        "first": "Lanqing",
                        "middle": [],
                        "last": "Xue",
                        "suffix": ""
                    },
                    {
                        "first": "Kaitao",
                        "middle": [],
                        "last": "Song",
                        "suffix": ""
                    },
                    {
                        "first": "Duocai",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    },
                    {
                        "first": "Xu",
                        "middle": [],
                        "last": "Tan",
                        "suffix": ""
                    },
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Nevin",
                        "suffix": ""
                    },
                    {
                        "first": "Tao",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Wei-Qiang",
                        "middle": [],
                        "last": "Qin",
                        "suffix": ""
                    },
                    {
                        "first": "Tie-Yan",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2107.01875"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Lanqing Xue, Kaitao Song, Duocai Wu, Xu Tan, Nevin L Zhang, Tao Qin, Wei-Qiang Zhang, and Tie- Yan Liu. 2021. DeepRapper: Neural rap generation with rhyme and rhythm modeling. arXiv preprint arXiv:2107.01875.",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "Chinese couplet generation with neural network structures",
                "authors": [
                    {
                        "first": "Rui",
                        "middle": [],
                        "last": "Yan",
                        "suffix": ""
                    },
                    {
                        "first": "Cheng-Te",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Xiaohua",
                        "middle": [],
                        "last": "Hu",
                        "suffix": ""
                    },
                    {
                        "first": "Ming",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics",
                "volume": "1",
                "issue": "",
                "pages": "2347--2357",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/P16-1222"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Rui Yan, Cheng-Te Li, Xiaohua Hu, and Ming Zhang. 2016. Chinese couplet generation with neural net- work structures. In Proceedings of the 54th An- nual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2347- 2357, Berlin, Germany. Association for Computa- tional Linguistics.",
                "links": null
            },
            "BIBREF22": {
                "ref_id": "b22",
                "title": "Youling: An ai-assisted lyrics creation system",
                "authors": [
                    {
                        "first": "Rongsheng",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Xiaoxi",
                        "middle": [],
                        "last": "Mao",
                        "suffix": ""
                    },
                    {
                        "first": "Le",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Lin",
                        "middle": [],
                        "last": "Jiang",
                        "suffix": ""
                    },
                    {
                        "first": "Lin",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Zhiwei",
                        "middle": [],
                        "last": "Hu",
                        "suffix": ""
                    },
                    {
                        "first": "Yadong",
                        "middle": [],
                        "last": "Xi",
                        "suffix": ""
                    },
                    {
                        "first": "Changjie",
                        "middle": [],
                        "last": "Fan",
                        "suffix": ""
                    },
                    {
                        "first": "Minlie",
                        "middle": [],
                        "last": "Huang",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "85--91",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Rongsheng Zhang, Xiaoxi Mao, Le Li, Lin Jiang, Lin Chen, Zhiwei Hu, Yadong Xi, Changjie Fan, and Minlie Huang. 2020a. Youling: An ai-assisted lyrics creation system. In Proceedings of the Conference on Empirical Methods in Natural Language Process- ing, pages 85-91.",
                "links": null
            },
            "BIBREF23": {
                "ref_id": "b23",
                "title": "Autosight: Distributed edge caching in short video network",
                "authors": [
                    {
                        "first": "Yuchao",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Pengmiao",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Zhili",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Bo",
                        "middle": [],
                        "last": "Bai",
                        "suffix": ""
                    },
                    {
                        "first": "Gong",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Wendong",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Bo",
                        "middle": [],
                        "last": "Lian",
                        "suffix": ""
                    },
                    {
                        "first": "Ke",
                        "middle": [],
                        "last": "Xu",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "IEEE Network",
                "volume": "34",
                "issue": "3",
                "pages": "194--199",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yuchao Zhang, Pengmiao Li, Zhili Zhang, Bo Bai, Gong Zhang, Wendong Wang, Bo Lian, and Ke Xu. 2020b. Autosight: Distributed edge caching in short video network. IEEE Network, 34(3):194-199.",
                "links": null
            },
            "BIBREF24": {
                "ref_id": "b24",
                "title": "Xiaoice band: A melody and arrangement generation framework for pop music",
                "authors": [
                    {
                        "first": "Hongyuan",
                        "middle": [],
                        "last": "Zhu",
                        "suffix": ""
                    },
                    {
                        "first": "Qi",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Nicholas",
                        "middle": [
                            "Jing"
                        ],
                        "last": "Yuan",
                        "suffix": ""
                    },
                    {
                        "first": "Chuan",
                        "middle": [],
                        "last": "Qin",
                        "suffix": ""
                    },
                    {
                        "first": "Jiawei",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Kun",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Guang",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    },
                    {
                        "first": "Furu",
                        "middle": [],
                        "last": "Wei",
                        "suffix": ""
                    },
                    {
                        "first": "Yuanchun",
                        "middle": [],
                        "last": "Xu",
                        "suffix": ""
                    },
                    {
                        "first": "Enhong",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery & Data Mining",
                "volume": "",
                "issue": "",
                "pages": "2837--2846",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Hongyuan Zhu, Qi Liu, Nicholas Jing Yuan, Chuan Qin, Jiawei Li, Kun Zhang, Guang Zhou, Furu Wei, Yuanchun Xu, and Enhong Chen. 2018. Xiaoice band: A melody and arrangement generation frame- work for pop music. In Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 2837-2846.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "num": null,
                "uris": null,
                "text": "The left figure shows the overall architecture of ChipSong. The right figure shows the internal structure of layers; for simplicity, we've omitted the drawing of residual connection and layer normalization.",
                "type_str": "figure"
            },
            "FIGREF1": {
                "num": null,
                "uris": null,
                "text": "[B \u2212 {length}],[I \u2212 {length}],[E \u2212 {length}]  to indicate the beginning, inside, and end of a word or term (i.e., a combination of Chinese characters), and specifically splice the \"BIE\" mark with a number to record word length.For example, \"[B \u2212 1]\" indicates the word length is 1, and \"[B \u2212 4], [I \u2212 4], [I \u2212 4], [E \u2212 4]\" indicates the word length is 4. This labeling strategy can avoid word boundary confusion during training. We set additional embedding symbols [S] (i.e., separator) and [C] (i.e., count) to respectively correspond to the separator [SEP ] and sentencelevel control characters, and [O] (i.e., outside) to correspond to trigger words and the ending character [EN D] in the lyric sequence.",
                "type_str": "figure"
            },
            "TABREF2": {
                "text": "\u96e8\u5929 ( Rainy day ) , Format: {6; 5}&{2,1,3; 1,1,1,2} , Rhyme: an, ian, uan",
                "type_str": "table",
                "html": null,
                "content": "<table><tr><td>Autoregressive Language Model Output Trigger word: BIE Word-granularity Embeddings</td><td>Transformer Decoder Block with Word-granularity Attention Autoregressive Transformer Decoder Block Autoregressive Transformer Decoder Block ...</td><td>Feed-Forward Word-granularity Concat&amp;Linear Feed-Forward Masked Self-Attention</td></tr><tr><td/><td/><td>Attention</td></tr><tr><td>Position</td><td/><td/></tr><tr><td>Embeddings</td><td/><td>Masked</td></tr><tr><td>Token</td><td/><td>Self-Attention</td></tr><tr><td>Embeddings</td><td/><td/></tr><tr><td/><td/><td>BIE Word-granularity</td></tr><tr><td>Reverse Order</td><td/><td>Embeddings</td></tr><tr><td/><td/><td>Position Embeddings</td></tr><tr><td>Input</td><td/><td/></tr><tr><td/><td/><td>Token Embeddings</td></tr></table>",
                "num": null
            },
            "TABREF3": {
                "text": "98.38 92.72 16.44 92.68 94.22 88.35 89.54 31.58",
                "type_str": "table",
                "html": null,
                "content": "<table><tr><td colspan=\"2\">No. Model</td><td colspan=\"5\">PPL (\u2193) MA-D1 MI-D1 MA-D2 MI-D2</td><td>SA</td><td>WA</td><td colspan=\"5\">RA WA-1 WA-2 WA-3 WA-4 WA-5</td></tr><tr><td>1</td><td>ALM</td><td>15.77</td><td>83.40</td><td>5.01</td><td>97.33</td><td>15.76</td><td>9.96</td><td colspan=\"2\">0.67 15.43 0.58</td><td>0.94</td><td>0.09</td><td>0.05</td><td>-</td></tr><tr><td>2</td><td>SongNet</td><td>12.33</td><td>88.05</td><td>4.77</td><td>98.05</td><td colspan=\"4\">18.92 97.80 5.89 13.82 3.67</td><td>6.82</td><td>0.75</td><td>0.59</td><td>-</td></tr><tr><td colspan=\"10\">3 17.39 4 ALM-F 8.49 89.24 4.47 98.48 ALM-T 14.78 86.27 3.49 97.20 14.84 10.68 0.51 12.01 0.45</td><td>0.59</td><td>0.07</td><td>-</td><td>-</td></tr><tr><td>5</td><td>ALM-R</td><td>15.55</td><td>89.49</td><td>4.76</td><td>98.28</td><td>19.63</td><td>9.57</td><td colspan=\"2\">0.36 98.38 0.35</td><td>0.46</td><td>0.08</td><td>-</td><td>-</td></tr><tr><td>6</td><td>ChipSong</td><td>7.69</td><td>89.20</td><td>5.22</td><td>98.04</td><td colspan=\"8\">21.69 98.54 86.64 98.56 86.04 89.03 78.74 76.11 18.42</td></tr></table>",
                "num": null
            },
            "TABREF4": {
                "text": "Automatic evaluation results of different models. SA: sentence-level format accuracy, WA: word-level format accuracy, RA: rhyme , WA-N: N-length word accuracy. Overall, ChipSong shows better control ability in all conditions. Note that single conditional control models perform better on corresponding conditions than ChipSong with full-conditional control applied, such as ALM-F on WA and WA-N metrics, because there are no constraints of other conditional controls, which is explained in result analysis.",
                "type_str": "table",
                "html": null,
                "content": "<table><tr><td colspan=\"2\">No. Model</td><td colspan=\"4\">BLEU-1 BLEU-2 BLEU-3 BLEU-4</td></tr><tr><td>1</td><td>ALM</td><td>23.50</td><td>7.63</td><td>2.61</td><td>0.90</td></tr><tr><td>2</td><td>SongNet</td><td>25.41</td><td>6.25</td><td>1.88</td><td>0.61</td></tr><tr><td>3</td><td>ALM-F</td><td>26.29</td><td>7.88</td><td>2.69</td><td>1.03</td></tr><tr><td>4 5</td><td>ALM-T ALM-R</td><td>29.84 23.83</td><td>11.33 6.11</td><td>4.53 1.75</td><td>1.83 0.54</td></tr><tr><td>6</td><td>ChipSong</td><td>28.55</td><td>9.56</td><td>3.38</td><td>1.35</td></tr></table>",
                "num": null
            },
            "TABREF5": {
                "text": "Effects of trigger words controlling contents.",
                "type_str": "table",
                "html": null,
                "content": "<table/>",
                "num": null
            },
            "TABREF6": {
                "text": "No. Ablation Model PPL (\u2193) MA-D1 MI-D1 MA-D2 MI-D2",
                "type_str": "table",
                "html": null,
                "content": "<table><tr><td/><td/><td/><td/><td/><td/><td/><td>SA</td><td>WA</td><td colspan=\"5\">RA WA-1 WA-2 WA-3 WA-4 WA-5</td></tr><tr><td>1</td><td>ChipSong</td><td>7.69</td><td>89.20</td><td>5.22</td><td>98.04</td><td colspan=\"8\">21.69 98.45 86.64 98.56 86.04 89.03 78.74 76.11 18.42</td></tr><tr><td>2</td><td>w/o WC-Emb</td><td>8.00</td><td>88.40</td><td>5.18</td><td>97.21</td><td colspan=\"7\">22.62 98.04 79.48 98.54 79.67 83.43 62.38 56.22</td><td>7.89</td></tr><tr><td>3</td><td>w/o WC-Att</td><td>9.75</td><td>88.89</td><td>4.94</td><td>97.90</td><td colspan=\"8\">21.53 97.98 78.95 98.53 77.94 82.75 66.88 58.01 15.79</td></tr><tr><td>4</td><td>w/o WC</td><td>12.55</td><td>85.51</td><td>4.11</td><td>96.65</td><td colspan=\"4\">15.99 98.36 5.39 98.22 3.79</td><td>6.55</td><td>0.78</td><td>0.36</td><td>-</td></tr><tr><td>5</td><td>w/o WC, SC</td><td>14.50</td><td>87.00</td><td>4.69</td><td>98.05</td><td>14.84</td><td>9.40</td><td colspan=\"2\">0.53 97.68 0.72</td><td>1.05</td><td>0.13</td><td>0.07</td><td>-</td></tr><tr><td>6</td><td>w/o TC</td><td>8.37</td><td>91.16</td><td>5.73</td><td>98.95</td><td colspan=\"8\">23.07 98.31 89.86 98.87 90.09 91.94 83.07 84.68 21.05</td></tr><tr><td>7</td><td>w/o RC</td><td>7.84</td><td>87.68</td><td>4.85</td><td>97.34</td><td colspan=\"8\">18.30 98.75 89.37 15.19 88.36 91.10 83.48 79.86 28.95</td></tr></table>",
                "num": null
            },
            "TABREF7": {
                "text": "Ablation results. WC: word-level format control, SC: sentence-level format control, TC: trigger word control, RC: rhyme control. Ablating one control of ChipSong causes the corresponding evaluation score to decrease, while evaluation scores of other controls increase due to the reduction of constraints for generation.",
                "type_str": "table",
                "html": null,
                "content": "<table><tr><td colspan=\"2\">No. Model</td><td colspan=\"4\">BLEU-1 BLEU-2 BLEU-3 BLEU-4</td></tr><tr><td>1</td><td>ChipSong</td><td>28.55</td><td>9.56</td><td>3.38</td><td>1.35</td></tr><tr><td>2</td><td>w/o WC-Emb</td><td>27.97</td><td>8.98</td><td>3.10</td><td>1.18</td></tr><tr><td>3</td><td>w/o WC-Att</td><td>26.93</td><td>8.01</td><td>2.67</td><td>1.02</td></tr><tr><td>4</td><td>w/o WC</td><td>27.49</td><td>8.39</td><td>2.81</td><td>1.00</td></tr><tr><td>5</td><td>w/o WC, SC</td><td>25.38</td><td>10.28</td><td>4.57</td><td>2.02</td></tr><tr><td>6</td><td>w/o TC</td><td>24.85</td><td>6.26</td><td>1.83</td><td>0.63</td></tr><tr><td>7</td><td>w/o RC</td><td>31.23</td><td>12.69</td><td>5.58</td><td>2.49</td></tr></table>",
                "num": null
            },
            "TABREF8": {
                "text": "Ablation results of trigger words controlling contents.",
                "type_str": "table",
                "html": null,
                "content": "<table><tr><td colspan=\"2\">No. Model</td><td>F</td><td>R</td><td>L</td><td>Avg</td></tr><tr><td>1</td><td>ALM</td><td colspan=\"4\">2.51 2.07 2.61 2.40</td></tr><tr><td>2</td><td>SongNet</td><td colspan=\"4\">2.53 1.82 2.96 2.44</td></tr><tr><td>3</td><td>ChipSong</td><td colspan=\"4\">2.56 2.44 2.96 2.65</td></tr><tr><td>4</td><td>ChipSong w/o WC</td><td colspan=\"4\">2.48 2.47 2.95 2.63</td></tr><tr><td>5</td><td colspan=\"5\">ChipSong w/o WC, SC 2.46 2.67 2.75 2.63</td></tr><tr><td>6</td><td>ChipSong w/o TC</td><td colspan=\"4\">2.51 1.75 2.97 2.41</td></tr><tr><td>7</td><td>ChipSong w/o RC</td><td colspan=\"4\">2.54 2.60 2.95 2.70</td></tr></table>",
                "num": null
            },
            "TABREF9": {
                "text": "Results of human evaluations. F: Fluency; R: Relevance; L: Listenability. Avg is the average score of F, R and L.",
                "type_str": "table",
                "html": null,
                "content": "<table/>",
                "num": null
            }
        }
    }
}