Datasets:

WINGNUS
/

ACL-OCL

	{
	"paper_id": "2019",
	"header": {
	"generated_with": "S2ORC 1.0.0",
	"date_generated": "2023-01-19T07:30:22.911608Z"
	},
	"title": "Introducing Aspects of Creativity in Automatic Poetry Generation",
	"authors": [
	{
	"first": "Brendan",
	"middle": [],
	"last": "Bena",
	"suffix": "",
	"affiliation": {
	"laboratory": "",
	"institution": "Drury University",
	"location": {
	"addrLine": "900 N. Benton Ave. Springfield",
	"postCode": "65109",
	"region": "Missouri"
	}
	},
	"email": ""
	},
	{
	"first": "Jugal",
	"middle": [],
	"last": "Kalita",
	"suffix": "",
	"affiliation": {},
	"email": ""
	}
	],
	"year": "",
	"venue": null,
	"identifiers": {},
	"abstract": "Poetry Generation involves teaching systems to automatically generate text that resembles poetic work. A deep learning system can learn to generate poetry on its own by training on a corpus of poems and modeling the particular style of language. In this paper, we propose taking an approach that fine-tunes GPT-2, a pre-trained language model, to our downstream task of poetry generation. We extend prior work on poetry generation by introducing creative elements. Specifically, we generate poems that express emotion and elicit the same in readers, and poems that use the language of dreams-called dream poetry. We are able to produce poems that correctly elicit the emotions of sadness and joy 87.5 and 85 percent, respectively, of the time. We produce dreamlike poetry by training on a corpus of texts that describe dreams. Poems from this model are shown to capture elements of dream poetry with scores of no less than 3.2 on the Likert scale. We perform crowdsourced human-evaluation for all our poems. We also make use of the Coh-Metrix tool, outlining metrics we use to gauge the quality of text generated.",
	"pdf_parse": {
	"paper_id": "2019",
	"_pdf_hash": "",
	"abstract": [
	{
	"text": "Poetry Generation involves teaching systems to automatically generate text that resembles poetic work. A deep learning system can learn to generate poetry on its own by training on a corpus of poems and modeling the particular style of language. In this paper, we propose taking an approach that fine-tunes GPT-2, a pre-trained language model, to our downstream task of poetry generation. We extend prior work on poetry generation by introducing creative elements. Specifically, we generate poems that express emotion and elicit the same in readers, and poems that use the language of dreams-called dream poetry. We are able to produce poems that correctly elicit the emotions of sadness and joy 87.5 and 85 percent, respectively, of the time. We produce dreamlike poetry by training on a corpus of texts that describe dreams. Poems from this model are shown to capture elements of dream poetry with scores of no less than 3.2 on the Likert scale. We perform crowdsourced human-evaluation for all our poems. We also make use of the Coh-Metrix tool, outlining metrics we use to gauge the quality of text generated.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Abstract",
	"sec_num": null
	}
	],
	"body_text": [
	{
	"text": "Many natural language processing tasks require the generation of human-like language. Some tasks, such as image and video captioning and automatic weather and sports reporting, convert nontextual data to text. Some others, such as summarization and machine translation, convert one text to another. There are additional tasks that aim to produce text, given a topic or a few keywords such as story generation, joke generation, and poetry generation, among others.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "Poetry generation produces creative content, and delivers the content in an aesthetically pleasing manner, usually following a specific structure.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "Thus, in addition to generating text as if in a story, the lines produced usually have a certain length, quite frequently there is a rhyming scheme as well as rhythm, and organization into structures such as couplets, quatrains, quintets, and stanzas. Among other tools, creativity comes from unusual usage of words through effects such as alliteration, assonance, and elision; use of metaphors, symbolism, and other linguistic devices; licensing of underlying imagery with expressed feelings, sentiments, and emotions.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "Work in natural language generation can be traced to pioneering rule-based simulations of chatbots such as the \"psychotherapist\" Eliza (Weizenbaum et al., 1966) and paranoid schizophrenia-suffering PARRY (Colby, 1981) . Surveys such as (Hovy, 1990; Reiter and Dale, 2000; Gatt and Krahmer, 2018; Santhanam and Shaikh, 2019) have described the progress in natural language generation over 50 years. Of late, the use of deep learning has produced enviable progress in natural language generation, especially in topics such as machine translation (Bahdanau et al., 2014; Wu et al., 2016) , image captioning (Mao et al., 2014) and dialogue generation (Li et al., 2016) .",
	"cite_spans": [
	{
	"start": 135,
	"end": 160,
	"text": "(Weizenbaum et al., 1966)",
	"ref_id": "BIBREF35"
	},
	{
	"start": 204,
	"end": 217,
	"text": "(Colby, 1981)",
	"ref_id": "BIBREF3"
	},
	{
	"start": 236,
	"end": 248,
	"text": "(Hovy, 1990;",
	"ref_id": "BIBREF10"
	},
	{
	"start": 249,
	"end": 271,
	"text": "Reiter and Dale, 2000;",
	"ref_id": "BIBREF26"
	},
	{
	"start": 272,
	"end": 295,
	"text": "Gatt and Krahmer, 2018;",
	"ref_id": "BIBREF5"
	},
	{
	"start": 296,
	"end": 323,
	"text": "Santhanam and Shaikh, 2019)",
	"ref_id": "BIBREF28"
	},
	{
	"start": 544,
	"end": 567,
	"text": "(Bahdanau et al., 2014;",
	"ref_id": "BIBREF0"
	},
	{
	"start": 568,
	"end": 584,
	"text": "Wu et al., 2016)",
	"ref_id": "BIBREF37"
	},
	{
	"start": 604,
	"end": 622,
	"text": "(Mao et al., 2014)",
	"ref_id": "BIBREF16"
	},
	{
	"start": 647,
	"end": 664,
	"text": "(Li et al., 2016)",
	"ref_id": "BIBREF13"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "This paper discusses the automatic generation of natural-sounding poems that are creative. Creativity comes in many hues, and we experiment with a few established ways of creative expression in poetry generation. First, we generate poetry that can potentially evoke a response from the readers or hearers in terms of emotions and feelings they generate. Additionally, we choose the idea of mimicking the language of dreams as another form of creative expression due to its longstanding history in poetry. Dream poetry dates back to medieval times where famous fourteenth century authors, like Chaucer, experimented using dreams as the structure for an image or picture they wished to paint with a poem (Spearing, 1976a) . A dream poem is said to be characterized by the 'I' of the poem and its substance of a dream or a vision included (Lynch, 1998) . To the best of our knowledge, prior work on poetry generation, whether using deep learning or not, has not explored the incorporation of emotion-eliciting phraseology or elements of creativity such as dream poetry.",
	"cite_spans": [
	{
	"start": 702,
	"end": 719,
	"text": "(Spearing, 1976a)",
	"ref_id": "BIBREF30"
	},
	{
	"start": 836,
	"end": 849,
	"text": "(Lynch, 1998)",
	"ref_id": "BIBREF15"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "Our research provides the following contributions:",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "\u2022 generating grammatical, coherent, and flowing poetry using the powerful and versatile GPT-2 architecture, \u2022 successfully generating poetry that elicits certain emotions in readers, and \u2022 generating poems that follow time-honored tradition of dream-like language usage and imagery.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "This paper is organized as follows. Section 2 presents related work. Section 3 discusses our approach to creative text generation including pre-processing steps, architecture used, and approaches to training. Section 4 discusses our experiments and results. Finally, we present evaluation of our research in Section 5, followed by conclusions and future work in Section 6.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "Early methods for poetry generation made use of template-oriented and rule-based techniques. These approaches often required a large amount of feature picking and knowledge of syntactic and semantic rules in a language (Oliveira, 2009 (Oliveira, , 2012 . Other methods treated poetry generation as special cases of machine translation or summarization tasks (Yan et al., 2013; He et al., 2012) . We believe that forcing a model to adhere to specific rules or templates, or summarizing or translating a given text to generate new poetry is unlikely to lead to the artistically expressive quality we seek to generate.",
	"cite_spans": [
	{
	"start": 219,
	"end": 234,
	"text": "(Oliveira, 2009",
	"ref_id": "BIBREF20"
	},
	{
	"start": 235,
	"end": 252,
	"text": "(Oliveira, , 2012",
	"ref_id": "BIBREF21"
	},
	{
	"start": 358,
	"end": 376,
	"text": "(Yan et al., 2013;",
	"ref_id": null
	},
	{
	"start": 377,
	"end": 393,
	"text": "He et al., 2012)",
	"ref_id": "BIBREF8"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Related Work",
	"sec_num": "2"
	},
	{
	"text": "More recently, deep learning methods have become prevalent in natural language generation, including poetry generation. Zhang and Lapata (2014) for instance, used Convolutional (CNN) and Recurrent Neural Networks (RNN) to generate Chinese Poetry. RNNs allow for shortterm memory of the language to be maintained by inputting the generated output of a network cell back into itself, essentially building context. Ghazvininejad et al. (2017) used Long Short-Term Memory (LSTM) units, which are advanced gated versions of RNNs, to the task of poetry generation. Wei et al. (2018) attempted to address the style issue by training the networks using particular poets and controlling for style in Chinese poetry. They found that with enough training data, adequate results could be achieved. Problems related to poetic structure were addressed by Hopkins and Kiela (2017) . They generated rhythmic poetry by training the network on only a single type of poetry to ensure produced poems adhered to a single rhythmic structure. It was found in human evaluations that while the poems produced were rated to be of lower quality than human produced poems, they were indistinguishable from human produced poems. Lau et al. (2018) took the LSTM approach one step further with the Deepspeare model by employing an attention mechanism to model interactions among generated words. They also use three neural networks, one for rhythm, one for rhyming and another for word choice in their quest to generate Shakespeare-like sonnets. Vaswani et al. (2017) developed a deep neural architecture called the Transformer that did away with any sort of need for recurrence. The Transformer also employed an elaborate attention mechanism that has been shown to be useful in natural language tasks. Radford et al. (2019) used this architecture in their Generative Pretrained Transformer 2 (GPT-2) model. GPT-2 is capable of many downstream tasks like text generation but to our knowledge, research has not been published using the GPT-2 model specifically for poetry generation.",
	"cite_spans": [
	{
	"start": 120,
	"end": 143,
	"text": "Zhang and Lapata (2014)",
	"ref_id": "BIBREF40"
	},
	{
	"start": 412,
	"end": 439,
	"text": "Ghazvininejad et al. (2017)",
	"ref_id": "BIBREF6"
	},
	{
	"start": 559,
	"end": 576,
	"text": "Wei et al. (2018)",
	"ref_id": "BIBREF34"
	},
	{
	"start": 841,
	"end": 865,
	"text": "Hopkins and Kiela (2017)",
	"ref_id": "BIBREF9"
	},
	{
	"start": 1200,
	"end": 1217,
	"text": "Lau et al. (2018)",
	"ref_id": "BIBREF12"
	},
	{
	"start": 1515,
	"end": 1536,
	"text": "Vaswani et al. (2017)",
	"ref_id": "BIBREF32"
	},
	{
	"start": 1772,
	"end": 1793,
	"text": "Radford et al. (2019)",
	"ref_id": "BIBREF25"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Related Work",
	"sec_num": "2"
	},
	{
	"text": "On a slightly different but related note, natural language generation influenced by multi-modal input was attempted by Vechtomova et al. (2018) to generate song lyrics in the style of specific artists by fusing outputs coming from lyrical inputs processed by an RNN and audio clips processed by a CNN. Text generation has also been influenced, in a cross domain manner, through images. The works of Liu et al. (2018) have shown that coupled visual-poetic embeddings can be used to pick out poetic clues in images, which in turn can be used to inspire the generated text. Though influenced natural language generation in and of itself is not a novel idea, we feel our attempt to style text with the intent of eliciting particular emotions provides a creative way to explore this subtask.",
	"cite_spans": [
	{
	"start": 119,
	"end": 143,
	"text": "Vechtomova et al. (2018)",
	"ref_id": "BIBREF33"
	},
	{
	"start": 399,
	"end": 416,
	"text": "Liu et al. (2018)",
	"ref_id": "BIBREF14"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Related Work",
	"sec_num": "2"
	},
	{
	"text": "Our goal is to successfully demonstrate the introduction of creative flair in automatic poetry generation in two exemplar ways: explicit show of emotion and the use of language that is predominantly first person with dream-like imagery. To enable the expression of emotion in generated poems, our work involves a preliminary step of scoring a corpus of downloaded poems for emotion to produce subsets of poems that express one of eight different identified emotions. This step is followed by the actual generation of poems by fine-tuning the pre-trained GPT-2 natural language model. We train eight separate models for eight different emotions, each on a sub-corpus predominantly demonstrating a particular emotion. To generate poems that use dream-like language, we create a text corpus composed of a large number of dream transcriptions created in first person by actual viewers of dreams. In this case, we apply transfer learning by fine-tuning the pre-trained GPT-2 on the dream corpus, followed by training again on poetry. We evaluate the generated poems using automated techniques as well as humans. A high-level overview of the emotion elicitation portion of our project is shown in Figure 1 . To create a corpus of poems based on the emotions they elicit, we make use of the EmoLex dic-tionary (Mohammad and Turney, 2013). EmoLex is a word-level emotion lexicon that associates English words with the eight different emotion categories we wish to explore. Each poem (or book of poems) in our dataset is given a score that is the total of the associated emotion scores in EmoLex for each word. The maximum emotion word score is taken and the poem is labeled under that emotion category. We create eight such datasets, one corresponding to each emotion category supported by EmoLex. This approach allows us to to train multiple models on our split dataset.",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 1191,
	"end": 1199,
	"text": "Figure 1",
	"ref_id": "FIGREF0"
	}
	],
	"eq_spans": [],
	"section": "Approach",
	"sec_num": "3"
	},
	{
	"text": "Currently, the emotions of joy, anticipation, trust, anger, and sadness represent a large portion of our data while the emotions of surprise, disgust, and fear are severely underrepresented. Table 1 shows key differences in models including the number of tokens in the text and the final average loss during training.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Poem Emotion Scoring",
	"sec_num": "3.1"
	},
	{
	"text": "To create a model for poetic language, we propose finetuning OpenAI's GPT-2 architecture. GPT-2 is a Transformer-based model that was trained simply to predict the next word in a 40GB text corpus (Radford et al., 2019) . This 40GB dataset, WebText, was scraped from the internet with certain heuristics that aimed to gather only quality text (i.e. only outbound Reddit links from posts with a karma rating of 3 stars or better). By training on such a large, all-encompassing corpus of text, the architecture has proven to model the English language well and has obtained state-of-theart results on downstream text-based tasks such as machine translation, question answering, and summarization. We leverage GPT-2's pre-trained knowledge of language for our downstream task of peotry generation.",
	"cite_spans": [
	{
	"start": 196,
	"end": 218,
	"text": "(Radford et al., 2019)",
	"ref_id": "BIBREF25"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "GPT Architecture",
	"sec_num": "3.2"
	},
	{
	"text": "GPT-2 is the successor of OpenAI's first Transformer-based architecture, GPT (Radford et al., 2018) , with a few changes to the structure. The medium version of GPT-2 we use contains 345M parameters and is a 24 layer, decoder-only Transformer architecture. GPT-2 moves layer normalization to the input of each sub-block, adds another layer normalization after the final selfattention block and increases context size from 512 to 1024 tokens. This architecture allows for long term dependencies to be captured better in language modeling. GPT-2's attention mechanism is referred to as a masked multi self-attention head. This technique allows for a relationship to be modeled for all words in an input sequence. Words that have multiple meanings can then be represented based on the context they appear in. Higher attention scores from surrounding words relate to a larger contribution to the representation of a word. GPT-2 makes use of byte-pair encoding (BPE) like its predecessor GPT but on UTF-8 byte sequences (Sennrich et al., 2015) . GPT-2's encoding is somewhere in between character level and word level. The model also prevents different versions of common words from being duplicated (i.e. fate!, fate?, and fate would not be joined). This technique improves the quality of the final byte segmentation. GPT-2's encoding rids the need for pre-processing or tokenization of data and is able to assign a probability to any Unicode string.",
	"cite_spans": [
	{
	"start": 77,
	"end": 99,
	"text": "(Radford et al., 2018)",
	"ref_id": "BIBREF24"
	},
	{
	"start": 1015,
	"end": 1038,
	"text": "(Sennrich et al., 2015)",
	"ref_id": "BIBREF29"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "GPT Architecture",
	"sec_num": "3.2"
	},
	{
	"text": "The task-agnostic nature of GPT-2 allows us to take a fine-tuning approach to our downstream task of poetry generation. Our approach to generating poems that exhibit emotion as well as dream-like imagery involves training the pretrained GPT-2 model. Our training protocol for the two cases are stated briefly below.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Training for Creative Poem Generation",
	"sec_num": "3.3"
	},
	{
	"text": "Poetry is a personal form of writing that expresses human feelings, and Mill (1860) famously said \"What is poetry, but thoughts and words in which emotion spontaneously embodies itself?\" Mill (1833) also said \"The object of poetry is confessedly to act upon the emotions\". Expressing emotions, with possible motive of eliciting the same emotions in readers, is a basic characteristic of poems. Our goal in this paper is to use artificial neural networks to generate poems that explicitly evoke certain specific emotions. To generate poems with emotional content, we have split our poetry data into sub-corpora, one sub-corpus for each emotion. We train the already pre-trained GPT-2 on a sub-corpus of poems that demonstrate a certain emotion. Pre-trained GPT-2 has a very strong foundational knowledge of English. We find that training it again on emotionbearing poetry seems to enable it to generate high quality poetry, which is even able to use emotionladen words for the correct form of elicitation. We also find that the poems we generate seem to exhibit proper punctuation as well as lines that have poem-appropriate length and sentences that are grammatically correct. In addition, the poems we generate seem to be quite readable and demonstrate high coherence. Detailed analyses are reported in the next section.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Generating Emotion Poems",
	"sec_num": "3.3.1"
	},
	{
	"text": "Dream poems represent a style of poetry that was \"astonishingly\" popular in the 14th through the 16th centuries (Spearing, 1976b; Windeatt, 2003) and are still popular (Russo, 2003) . Such poems tell a story based on a dream or a number of dreams, dreamt by the narrator or by a character that the poet introduces. Spearing (1976b) claimed that dream poems are based on objective experience, but at the same time they are free of constraints of everyday possibilities. Such poems represent the outcome of a poetic process with many different influences, models, and analogues (Windeatt, 2003) , but without going into such details, our goal is to see if an ANN can produce poems which share characteristics with dream po-ems.",
	"cite_spans": [
	{
	"start": 112,
	"end": 129,
	"text": "(Spearing, 1976b;",
	"ref_id": "BIBREF31"
	},
	{
	"start": 130,
	"end": 145,
	"text": "Windeatt, 2003)",
	"ref_id": "BIBREF36"
	},
	{
	"start": 168,
	"end": 181,
	"text": "(Russo, 2003)",
	"ref_id": "BIBREF27"
	},
	{
	"start": 315,
	"end": 331,
	"text": "Spearing (1976b)",
	"ref_id": "BIBREF31"
	},
	{
	"start": 576,
	"end": 592,
	"text": "(Windeatt, 2003)",
	"ref_id": "BIBREF36"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Generating Dream Poems",
	"sec_num": "3.3.2"
	},
	{
	"text": "To generate poems that demonstrate first-person language with dream-like imagery, we take a similar approach. However, in this case, GPT-2 undergoes three separate training cycles. The first cycle is the pre-training that GPT-2 goes through before release to the public by OpenAI. Second, we train the pre-trained model on a corpus of firstperson dream descriptions. Third, we train again on poems. Our hypothesis is that pre-training by OpenAI results in good basic knowledge of English; that training on the dream corpus endows the network with the knowledge of first-person imagery-based language; and that the last training cycle teaches the network language of poems. We demonstrate in the next section that we are not far off from our being successful in our hypothesis.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Generating Dream Poems",
	"sec_num": "3.3.2"
	},
	{
	"text": "As stated by Radford et al. (2019) , the core approach of GPT-2 is language modeling. A language model can be thought of as a probability distribution over a sequence of words in the form:",
	"cite_spans": [
	{
	"start": 13,
	"end": 34,
	"text": "Radford et al. (2019)",
	"ref_id": "BIBREF25"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Text Generation and Sampling",
	"sec_num": "3.4"
	},
	{
	"text": "EQUATION",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [
	{
	"start": 0,
	"end": 8,
	"text": "EQUATION",
	"ref_id": "EQREF",
	"raw_str": "p(w 1 , ..., w n ).",
	"eq_num": "(1)"
	}
	],
	"section": "Text Generation and Sampling",
	"sec_num": "3.4"
	},
	{
	"text": "Likewise, natural language tends to have a sequential order so it can be modeled in terms of the conditional probability of a word given the words preceding it (Bengio et al., 2003) :",
	"cite_spans": [
	{
	"start": 160,
	"end": 181,
	"text": "(Bengio et al., 2003)",
	"ref_id": "BIBREF1"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Text Generation and Sampling",
	"sec_num": "3.4"
	},
	{
	"text": "EQUATION",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [
	{
	"start": 0,
	"end": 8,
	"text": "EQUATION",
	"ref_id": "EQREF",
	"raw_str": "p(w n \|w 1 , ..., w n\u22121 ).",
	"eq_num": "(2)"
	}
	],
	"section": "Text Generation and Sampling",
	"sec_num": "3.4"
	},
	{
	"text": "We make use of this probabilistic style of language modeling by sampling from the distribution in a semi-random fashion. Just as the GPT-2 paper does for its text generation, we make use of Top K sampling, limiting the possible guesses of words to 40. In addition to Top K, we make use of a temperature constant of 0.75 which controls randomness in the distribution. A temperature closer to 0 correlates to less randomness while a temperature closer to 1 relates to more randomness. Finally, at the end of the generation process, we employ a simple text cleaning algorithm that allows poems to end more naturally and rather than trail off as they do sometimes.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Text Generation and Sampling",
	"sec_num": "3.4"
	},
	{
	"text": "In order to classify emotion-eliciting poems or books, we use the NRC Word-Emotion Asso-ciation Lexicon (EmoLex) resource. EmoLex 1 was created by the National Research Council of Canada and includes 14,182 English words that are associated with different emotions and positive or negative sentiment (Mohammad and Turney, 2013). Words in EmoLex have been manually annotated via crowd-sourcing and emotions fall into one or more categories of eight basic emotions: joy, trust, fear, surprise, sadness, anticipation, anger, and disgust (Plutchik, 2014) . We elect to use this simplified version of the Wheel of Emotions due to its parallels with the available EmoLex dataset. This resource provides us with a way to fabricate a ground truth in the types of emotion-infused texts we wish to use for training data. To handle the training and generation portions of the project, we draw data from the Project Gutenberg website 2 . Project Gutenberg is a massive online database containing over 59,000 eBooks. We limit this corpus to a smaller subcorpus using an adaptation of the GutenTag tool (Brooke et al., 2015) . This tool allows us to place constraints on the amount of literature we choose to use in our work. Our final dataset includes approximately three million lines of poetic text from the Gutenberg database and is further divided by poem/book into our eight emotion categories.",
	"cite_spans": [
	{
	"start": 534,
	"end": 550,
	"text": "(Plutchik, 2014)",
	"ref_id": null
	},
	{
	"start": 1089,
	"end": 1110,
	"text": "(Brooke et al., 2015)",
	"ref_id": "BIBREF2"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Datasets and Resources",
	"sec_num": "4.1"
	},
	{
	"text": "We attempt to create dream poetry by making use of the DreamBank dataset. The Dream-Bank was created by Schneider & Domhoff at UC-Santa Cruz 3 . The dataset contains a collection of over 20,000 dreams from users age 7 to 74. We scraped this dataset from the website assuring that dreams collected were recorded only in English. The DreamBank allows us to attempt trans-Heard I a song of joy, A song of happy sound, Fills all the air I breathe, To him I sing, to him I sing the happy song. All night long on the steep green grass I ride and sing Figure 4 : A hand-picked, automatically generated poem from the joy model",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 545,
	"end": 553,
	"text": "Figure 4",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Datasets and Resources",
	"sec_num": "4.1"
	},
	{
	"text": "The other, who with one accord Wrote my essay, in that he was dear And good, and knew well, how we ought to treat A man of such renown, and such love? He's a good honest man, no doubt Figure 5 : A hand-picked, automatically generated poem from the trust model fer learning by finetuning on the dream dataset first, then further finetuning on our poetry dataset. Initially, we retrained 6 GPT-2 based models. Default training parameters were used each of the 5 different emotion datasets and our dream dataset. All were trained for 12,000 steps (except for our dream model which was trained for 12k steps on dreams and on poetry) with a learning rate of 0.0001. When generating text, we do not input context: we allow the model to write the poem entirely through the sampling of conditional probability from the language it has modeled.",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 184,
	"end": 192,
	"text": "Figure 5",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Datasets and Resources",
	"sec_num": "4.1"
	},
	{
	"text": "Figures 4 through 8 give examples of 5 poems that we have hand-picked to illustrate the quality of poems generated. A cursory glance at the poems reveals the high quality of the text in terms of lexical choice, grammatical integrity, and semantic cohesion. We discuss how we quantitatively assess the poems below.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Datasets and Resources",
	"sec_num": "4.1"
	},
	{
	"text": "In the first crowd-sourced analysis of our emotioneliciting poetry we presented four poems from each category (of the five data-represented emotion categories) to ten human reviewers with undergraduate level educational backgrounds. All reviewers are native speakers of English. Poems presented were randomly selected from the top 20 EmoLex scored poems out of a pool of 1,000 generated poems. These reviewers were asked to rate each poem based on the emotions elicited within",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Evaluation",
	"sec_num": "5"
	},
	{
	"text": "We have reached the peak of the highest mountain in the world The mountain of dreams. This is the view Across the valley, One hour's journey back, We crossed it on the way between A band of beautiful young women.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Evaluation",
	"sec_num": "5"
	},
	{
	"text": "There was Figure 6 : A hand-picked, automatically generated poem from the anticipation model A long trail of falling mist Had made its way here, and now Aerily it seemed, as if to drown The discordant thunder clang. It seemed to drown the music of the rain;",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 10,
	"end": 18,
	"text": "Figure 6",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Evaluation",
	"sec_num": "5"
	},
	{
	"text": "In this lost place of sorrow Far off Figure 7 : A hand-picked, automatically generated poem from the sadness model Amidst the chaos throng'd, with angry voices each His rival's mockery; loud their scorn was fill'd; So fierce their rage, and in their eager power Met on the walls of Troy, were fill'd with dismay. Figure 8 : A hand-picked, automatically generated poem from the anger model A thousand stars at once, An hundred thousand stars!",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 37,
	"end": 45,
	"text": "Figure 7",
	"ref_id": null
	},
	{
	"start": 313,
	"end": 321,
	"text": "Figure 8",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Evaluation",
	"sec_num": "5"
	},
	{
	"text": "The sun was low, And the stars were bright, My heart would do the same.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Evaluation",
	"sec_num": "5"
	},
	{
	"text": "A thousand stars at once, A hundred thousand stars! The night had begun, And the stars were all the same. When I came back from the dead, I saw the stars Figure 9 : A hand-picked, automatically generated poem from the dream model For she was mine. I was the only one She had, And a thousand other friends, And a hundred more She held me dear. Her eyes were clear, her cheeks were bright, Her heart was like a rose, Her mouth was full of music, Her lips were white As snow, And the music she sang Figure 10 : A hand-picked poem, automatically generated from the dream model Emotion Anger Antic. Joy Sad. Trust % 65 40 85 87.5 32.5 Table 2 : Average percentage of correctly elicited emotion across four poems in each category them after reading. An emotion was deemed correctly elicited if the associated Likert score was 4 or greater from the reviewer. Table 2 illustrates the results from our evaluation. When taking the average percentage of correct emotion-eliciting poems, the models of joy, sadness, and anger produced the most promising results while the trust and anticipation models were less than satisfactory. We believe this is because joy, sadness and anger are basic or fundamental emotions compared to trust and anticipation, which are more complex and difficult to explain. Although there are many opinions among psychologists about what constitute basic emotions, joy, sadness and anger, (especially the last two) seem to occur the most often in proposals that demarcate a set of basic emotions (Ortony and Turner, 1990) . To preserve consistency in our experiments, we evaluate our dream model poetry in a manner similar to our evaluation of the emotion poems. Four poems from the model were presented to the same ten judges and they were asked to assess the poems based on qualities of dream poetry. These poems were cherry picked from a pool of 1,000 generated poems. A dream poem is said to have the following qualities (Windeatt, 2003; Spearing, 1976b; Russo, 2003) among many other qualities. We believe these three are the least ambiguous and easiest to decipher for human evaluation.",
	"cite_spans": [
	{
	"start": 1510,
	"end": 1535,
	"text": "(Ortony and Turner, 1990)",
	"ref_id": "BIBREF22"
	},
	{
	"start": 1939,
	"end": 1955,
	"text": "(Windeatt, 2003;",
	"ref_id": "BIBREF36"
	},
	{
	"start": 1956,
	"end": 1972,
	"text": "Spearing, 1976b;",
	"ref_id": "BIBREF31"
	},
	{
	"start": 1973,
	"end": 1985,
	"text": "Russo, 2003)",
	"ref_id": "BIBREF27"
	}
	],
	"ref_spans": [
	{
	"start": 154,
	"end": 162,
	"text": "Figure 9",
	"ref_id": null
	},
	{
	"start": 496,
	"end": 505,
	"text": "Figure 10",
	"ref_id": "FIGREF0"
	},
	{
	"start": 630,
	"end": 637,
	"text": "Table 2",
	"ref_id": null
	},
	{
	"start": 852,
	"end": 859,
	"text": "Table 2",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Evaluation",
	"sec_num": "5"
	},
	{
	"text": "Poem 1 2 3 4 Qual 1 5 4.9 4.8 4.5 Qual 2 3.5 4.1 3.2 3.3 Qual 3 3.9 4.2 3.7 3.7 Analysis of results show that machine generated poems are able to capture the first person perspective well, achieving between 4.5 and 5 average Likert scores. The poems often appear to retell a story or an event, scoring between 3.7 and 4.2 average Likert scores. The nature of poetry and dream recounts that make up our data is often narrative, so this result stands to reason. However, Quality 2 scores of the poem substance containing a dream or vision are questionable. We suspect the Quality 2 score is lower due to the ambiguity in ascertaining dream text from regular text. Table 3 highlights our results for the dream model.",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 662,
	"end": 669,
	"text": "Table 3",
	"ref_id": "TABREF2"
	}
	],
	"eq_spans": [],
	"section": "Evaluation",
	"sec_num": "5"
	},
	{
	"text": "Currently, there exists no widely available standard for evaluating poetry generation. Scores like BLEU, ROUGE, METEOR, etc. are more suited for Machine Translation (MT) tasks (Zhang et al., 2019) . For example, they compare how similar sentence P is to translated-sentenceP. Instead, we outline some metrics from the Coh-Metrix web tool that helps us further quantitatively evaluate the quality of text generated. With the goal of eliciting emotions, we claim that subjective analysis of generated poetry is superior to any available objective metrics.",
	"cite_spans": [
	{
	"start": 176,
	"end": 196,
	"text": "(Zhang et al., 2019)",
	"ref_id": "BIBREF39"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Evaluation",
	"sec_num": "5"
	},
	{
	"text": "To provide a quantitative calculation of the caliber of text our models produce, we outline in this section relevant metrics from the University of Memphis Coh-Metrix tool (Graesser et al., 2004) . Coh-Metrix is a text evaluation software kit and from it, we have chosen 8 forms of assessment. The first two, Flesch-Kincaid Grade Level (FKGL) and Flesch Reading Ease (FRE), are two standard measures that deal with text readability and ease (Klare, 1974) . The FKGL scores a text from grade level 0 to 18, while the FRE score is a 0-100 index with 100 being an easily readable text. We aim to produce text that is readable by all, so a low FKGL score and high FRE score would be ideal. The next metrics we employ evaluate at the word level. The word imageability (IMGc) and word concreteness (CNCc) scores measure content words on their ability to create an image in the reader's mind and their ability to appeal to a reader's senses, respectively (Coltheart, 1981) . We aim for our art to create a connection between the reader and poem, so we believe imageability and concreteness of content words are two good measures with this in mind. We also make use of three text easibility principal component scores: narrativity (PCNARp), referential cohesion (PCREFp), and syntactic simplicity (PC-SYNp) (Graesser et al., 2004) . The text easibility PC scores are percentile scales, and thus we aim for higher numbers for these scores. Finally, we make use of the Lexical Diversity Type:Token Ratio score (LDTTRa) for all words. LDTTRa measures the ratio of type (unique) words to all tokens in the text. Because our text is relatively short, we aim for a middle ground in the LDTTRa ratio, meaning there is uniqueness in the word choice of the text, but cohesion is still upheld.",
	"cite_spans": [
	{
	"start": 172,
	"end": 195,
	"text": "(Graesser et al., 2004)",
	"ref_id": "BIBREF7"
	},
	{
	"start": 441,
	"end": 454,
	"text": "(Klare, 1974)",
	"ref_id": "BIBREF11"
	},
	{
	"start": 948,
	"end": 965,
	"text": "(Coltheart, 1981)",
	"ref_id": null
	},
	{
	"start": 1299,
	"end": 1322,
	"text": "(Graesser et al., 2004)",
	"ref_id": "BIBREF7"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Coh-Metrix",
	"sec_num": "5.1"
	},
	{
	"text": "Inspection of our Coh-Metrix results show that randomly selected poems from all models fall at or below the 2nd-grade reading level (in FKGL scores) and are greater than 93 on the FRE scale. This suggests generated poems are easily readable by the majority of viewers. Looking at the IMGc and CNCc scores, we see that our poems, except for the dream model concreteness, fall in the 400s. Words with higher imageability and concreteness fall around the low 600s while words that are lower fall around the upper 200s on this scale. These scores reveal that our models are generating text that is concrete in word choice and that paint a picture. Our dream model scoring lower in the concreteness is reasonable as the word choice of dreams tends to be more abstract. Lastly, percentile scores of PCSYNp and PCNARp show that the majority of models are producing poems that are both syntactically simplistic and narrative. Most PCREFp scores are on the lower end of the scale. We suspect the reason these scores are lower is because the poems are not necessarily related and were all input at once. Table 4 highlights these scores for each poetry model.",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 1094,
	"end": 1101,
	"text": "Table 4",
	"ref_id": "TABREF4"
	}
	],
	"eq_spans": [],
	"section": "Coh-Metrix",
	"sec_num": "5.1"
	},
	{
	"text": "In this paper we influenced automatic natural language generation to create poetry through the use of classified emotion poems and dream text. To do so, we first leveraged a word-level emotion lexicon to construct a meaning for emotion-eliciting text and used that text to train separate language models. Next, we gathered data of dream records and employed transfer learning in attempts to generate dream-like poetry. The work reported in this paper seeks to create art in the form of auto-generated poetry while opening the door to more projects involving emotion-eliciting text-based tasks and influenced creative neural generation.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Conclusion & Future Work",
	"sec_num": "6"
	},
	{
	"text": "We would like to thank the reviewers for their feedback on this project. Comments and suggestions from reviewers -both those that the were incorporated into this article and those on which we will report in future work -provide invaluable insight as to improving our results. Importantly, our continuing research involves gathering a more comprehensive human-evaluation with a larger number of reviewers and poems to be judged. We also wish to gather data for the underrepresented emotion categories, leading, ideally, to a more robust language model for each emotion. Our work thus far provides a baseline for introducing emotions into generated text via a word-level lexicon, but we wish to employ other tools -segment-level lexicons, for example -in an attempt to better capture the contextual dependencies of emotion. Additionally, the word-level baseline we have produced focuses on generating single-emotion text. We are interested in examining poems of multiple emotions and different levels of intensity to expand on this study. Finally, we wish to seek out additional forms of replicating creativity that artists incorporate in their work.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Conclusion & Future Work",
	"sec_num": "6"
	},
	{
	"text": "https://saifmohammad.com/WebPages/ NRC-Emotion-Lexicon.htm 2 https://www.gutenberg.org/ 3 https://www.dreambank.net/",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	}
	],
	"back_matter": [],
	"bib_entries": {
	"BIBREF0": {
	"ref_id": "b0",
	"title": "Neural machine translation by jointly learning to align and translate",
	"authors": [
	{
	"first": "Dzmitry",
	"middle": [],
	"last": "Bahdanau",
	"suffix": ""
	},
	{
	"first": "Kyunghyun",
	"middle": [],
	"last": "Cho",
	"suffix": ""
	},
	{
	"first": "Yoshua",
	"middle": [],
	"last": "Bengio",
	"suffix": ""
	}
	],
	"year": 2014,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {
	"arXiv": [
	"arXiv:1409.0473"
	]
	},
	"num": null,
	"urls": [],
	"raw_text": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Ben- gio. 2014. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473.",
	"links": null
	},
	"BIBREF1": {
	"ref_id": "b1",
	"title": "A neural probabilistic language model",
	"authors": [
	{
	"first": "Yoshua",
	"middle": [],
	"last": "Bengio",
	"suffix": ""
	},
	{
	"first": "R\u00e9jean",
	"middle": [],
	"last": "Ducharme",
	"suffix": ""
	},
	{
	"first": "Pascal",
	"middle": [],
	"last": "Vincent",
	"suffix": ""
	},
	{
	"first": "Christian",
	"middle": [],
	"last": "Jauvin",
	"suffix": ""
	}
	],
	"year": 2003,
	"venue": "Journal of machine learning research",
	"volume": "3",
	"issue": "",
	"pages": "1137--1155",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Yoshua Bengio, R\u00e9jean Ducharme, Pascal Vincent, and Christian Jauvin. 2003. A neural probabilistic lan- guage model. Journal of machine learning research, 3(Feb):1137-1155.",
	"links": null
	},
	"BIBREF2": {
	"ref_id": "b2",
	"title": "Gutentag: an nlp-driven tool for digital humanities research in the project gutenberg corpus",
	"authors": [
	{
	"first": "Julian",
	"middle": [],
	"last": "Brooke",
	"suffix": ""
	},
	{
	"first": "Adam",
	"middle": [],
	"last": "Hammond",
	"suffix": ""
	},
	{
	"first": "Graeme",
	"middle": [],
	"last": "Hirst",
	"suffix": ""
	}
	],
	"year": 2015,
	"venue": "Proceedings of the Fourth Workshop on Computational Linguistics for Literature",
	"volume": "",
	"issue": "",
	"pages": "42--47",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Julian Brooke, Adam Hammond, and Graeme Hirst. 2015. Gutentag: an nlp-driven tool for digital hu- manities research in the project gutenberg corpus. In Proceedings of the Fourth Workshop on Compu- tational Linguistics for Literature, pages 42-47.",
	"links": null
	},
	"BIBREF3": {
	"ref_id": "b3",
	"title": "Modeling a paranoid mind",
	"authors": [
	{
	"first": "Kenneth",
	"middle": [],
	"last": "Mark",
	"suffix": ""
	},
	{
	"first": "Colby",
	"middle": [],
	"last": "",
	"suffix": ""
	}
	],
	"year": 1981,
	"venue": "Behavioral and Brain Sciences",
	"volume": "4",
	"issue": "4",
	"pages": "515--534",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Kenneth Mark Colby. 1981. Modeling a paranoid mind. Behavioral and Brain Sciences, 4(4):515- 534.",
	"links": null
	},
	"BIBREF4": {
	"ref_id": "b4",
	"title": "The mrc psycholinguistic database",
	"authors": [],
	"year": 1981,
	"venue": "The Quarterly Journal of Experimental Psychology Section A",
	"volume": "33",
	"issue": "4",
	"pages": "497--505",
	"other_ids": {
	"DOI": [
	"10.1080/14640748108400805"
	]
	},
	"num": null,
	"urls": [],
	"raw_text": "Max Coltheart. 1981. The mrc psycholinguistic database. The Quarterly Journal of Experimental Psychology Section A, 33(4):497-505.",
	"links": null
	},
	"BIBREF5": {
	"ref_id": "b5",
	"title": "Survey of the state of the art in natural language generation: Core tasks, applications and evaluation",
	"authors": [
	{
	"first": "Albert",
	"middle": [],
	"last": "Gatt",
	"suffix": ""
	},
	{
	"first": "Emiel",
	"middle": [],
	"last": "Krahmer",
	"suffix": ""
	}
	],
	"year": 2018,
	"venue": "Journal of Artificial Intelligence Research",
	"volume": "61",
	"issue": "",
	"pages": "65--170",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Albert Gatt and Emiel Krahmer. 2018. Survey of the state of the art in natural language generation: Core tasks, applications and evaluation. Journal of Artifi- cial Intelligence Research, 61:65-170.",
	"links": null
	},
	"BIBREF6": {
	"ref_id": "b6",
	"title": "Hafez: an interactive poetry generation system",
	"authors": [
	{
	"first": "Marjan",
	"middle": [],
	"last": "Ghazvininejad",
	"suffix": ""
	},
	{
	"first": "Xing",
	"middle": [],
	"last": "Shi",
	"suffix": ""
	},
	{
	"first": "Jay",
	"middle": [],
	"last": "Priyadarshi",
	"suffix": ""
	},
	{
	"first": "Kevin",
	"middle": [],
	"last": "Knight",
	"suffix": ""
	}
	],
	"year": 2017,
	"venue": "Proceedings of ACL 2017, System Demonstrations",
	"volume": "",
	"issue": "",
	"pages": "43--48",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Marjan Ghazvininejad, Xing Shi, Jay Priyadarshi, and Kevin Knight. 2017. Hafez: an interactive poetry generation system. In Proceedings of ACL 2017, System Demonstrations, pages 43-48, Vancouver, Canada. Association for Computational Linguistics.",
	"links": null
	},
	"BIBREF7": {
	"ref_id": "b7",
	"title": "Coh-metrix: Analysis of text on cohesion and language",
	"authors": [
	{
	"first": "Arthur",
	"middle": [
	"C"
	],
	"last": "Graesser",
	"suffix": ""
	},
	{
	"first": "Danielle",
	"middle": [
	"S"
	],
	"last": "Mcnamara",
	"suffix": ""
	},
	{
	"first": "Max",
	"middle": [
	"M"
	],
	"last": "Louwerse",
	"suffix": ""
	},
	{
	"first": "Zhiqiang",
	"middle": [],
	"last": "Cai",
	"suffix": ""
	}
	],
	"year": 2004,
	"venue": "Behavior Research Methods, Instruments, & Computers",
	"volume": "36",
	"issue": "",
	"pages": "193--202",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Arthur C. Graesser, Danielle S. McNamara, Max M. Louwerse, and Zhiqiang Cai. 2004. Coh-metrix: Analysis of text on cohesion and language. Behav- ior Research Methods, Instruments, & Computers, 36:193-202.",
	"links": null
	},
	"BIBREF8": {
	"ref_id": "b8",
	"title": "Generating chinese classical poems with statistical machine translation models",
	"authors": [
	{
	"first": "Jing",
	"middle": [],
	"last": "He",
	"suffix": ""
	},
	{
	"first": "Ming",
	"middle": [],
	"last": "Zhou",
	"suffix": ""
	},
	{
	"first": "Long",
	"middle": [],
	"last": "Jiang",
	"suffix": ""
	}
	],
	"year": 2012,
	"venue": "Twenty-Sixth AAAI Conference on Artificial Intelligence",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Jing He, Ming Zhou, and Long Jiang. 2012. Generat- ing chinese classical poems with statistical machine translation models. In Twenty-Sixth AAAI Confer- ence on Artificial Intelligence.",
	"links": null
	},
	"BIBREF9": {
	"ref_id": "b9",
	"title": "Automatically generating rhythmic verse with neural networks",
	"authors": [
	{
	"first": "Jack",
	"middle": [],
	"last": "Hopkins",
	"suffix": ""
	},
	{
	"first": "Douwe",
	"middle": [],
	"last": "Kiela",
	"suffix": ""
	}
	],
	"year": 2017,
	"venue": "Proceedings of the 55th Annual Meeting of the ACL",
	"volume": "1",
	"issue": "",
	"pages": "168--178",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Jack Hopkins and Douwe Kiela. 2017. Automatically generating rhythmic verse with neural networks. In Proceedings of the 55th Annual Meeting of the ACL (Volume 1: Long Papers), pages 168-178.",
	"links": null
	},
	"BIBREF10": {
	"ref_id": "b10",
	"title": "Pragmatics and natural language generation",
	"authors": [
	{
	"first": "H",
	"middle": [],
	"last": "Eduard",
	"suffix": ""
	},
	{
	"first": "",
	"middle": [],
	"last": "Hovy",
	"suffix": ""
	}
	],
	"year": 1990,
	"venue": "Artificial Intelligence",
	"volume": "43",
	"issue": "2",
	"pages": "153--197",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Eduard H Hovy. 1990. Pragmatics and natural lan- guage generation. Artificial Intelligence, 43(2):153- 197.",
	"links": null
	},
	"BIBREF11": {
	"ref_id": "b11",
	"title": "Assessing readability. Reading research quarterly",
	"authors": [
	{
	"first": "",
	"middle": [],
	"last": "George R Klare",
	"suffix": ""
	}
	],
	"year": 1974,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "62--102",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "George R Klare. 1974. Assessing readability. Reading research quarterly, pages 62-102.",
	"links": null
	},
	"BIBREF12": {
	"ref_id": "b12",
	"title": "Deep-speare: A joint neural model of poetic language, meter and rhyme",
	"authors": [
	{
	"first": "Trevor",
	"middle": [],
	"last": "Jey Han Lau",
	"suffix": ""
	},
	{
	"first": "Timothy",
	"middle": [],
	"last": "Cohn",
	"suffix": ""
	},
	{
	"first": "Julian",
	"middle": [],
	"last": "Baldwin",
	"suffix": ""
	},
	{
	"first": "Adam",
	"middle": [],
	"last": "Brooke",
	"suffix": ""
	},
	{
	"first": "",
	"middle": [],
	"last": "Hammond",
	"suffix": ""
	}
	],
	"year": 2018,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Jey Han Lau, Trevor Cohn, Timothy Baldwin, Julian Brooke, and Adam Hammond. 2018. Deep-speare: A joint neural model of poetic language, meter and rhyme. CoRR, abs/1807.03491.",
	"links": null
	},
	"BIBREF13": {
	"ref_id": "b13",
	"title": "Deep reinforcement learning for dialogue generation",
	"authors": [
	{
	"first": "Jiwei",
	"middle": [],
	"last": "Li",
	"suffix": ""
	},
	{
	"first": "Will",
	"middle": [],
	"last": "Monroe",
	"suffix": ""
	},
	{
	"first": "Alan",
	"middle": [],
	"last": "Ritter",
	"suffix": ""
	},
	{
	"first": "Michel",
	"middle": [],
	"last": "Galley",
	"suffix": ""
	},
	{
	"first": "Jianfeng",
	"middle": [],
	"last": "Gao",
	"suffix": ""
	},
	{
	"first": "Dan",
	"middle": [],
	"last": "Jurafsky",
	"suffix": ""
	}
	],
	"year": 2016,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {
	"arXiv": [
	"arXiv:1606.01541"
	]
	},
	"num": null,
	"urls": [],
	"raw_text": "Jiwei Li, Will Monroe, Alan Ritter, Michel Galley, Jianfeng Gao, and Dan Jurafsky. 2016. Deep rein- forcement learning for dialogue generation. arXiv preprint arXiv:1606.01541.",
	"links": null
	},
	"BIBREF14": {
	"ref_id": "b14",
	"title": "Beyond narrative description: Generating poetry from images by multi-adversarial training",
	"authors": [
	{
	"first": "Bei",
	"middle": [],
	"last": "Liu",
	"suffix": ""
	},
	{
	"first": "Jianlong",
	"middle": [],
	"last": "Fu",
	"suffix": ""
	},
	{
	"first": "P",
	"middle": [],
	"last": "Makoto",
	"suffix": ""
	},
	{
	"first": "Masatoshi",
	"middle": [],
	"last": "Kato",
	"suffix": ""
	},
	{
	"first": "",
	"middle": [],
	"last": "Yoshikawa",
	"suffix": ""
	}
	],
	"year": 2018,
	"venue": "Proceedings of the 26th ACM International Conference on Multimedia, MM '18",
	"volume": "",
	"issue": "",
	"pages": "783--791",
	"other_ids": {
	"DOI": [
	"10.1145/3240508.3240587"
	]
	},
	"num": null,
	"urls": [],
	"raw_text": "Bei Liu, Jianlong Fu, Makoto P. Kato, and Masatoshi Yoshikawa. 2018. Beyond narrative description: Generating poetry from images by multi-adversarial training. In Proceedings of the 26th ACM Interna- tional Conference on Multimedia, MM '18, pages 783-791, New York, NY, USA. ACM.",
	"links": null
	},
	"BIBREF15": {
	"ref_id": "b15",
	"title": "Medieval Dream-Poetry",
	"authors": [
	{
	"first": "L",
	"middle": [],
	"last": "Kathryn",
	"suffix": ""
	},
	{
	"first": "",
	"middle": [],
	"last": "Lynch",
	"suffix": ""
	}
	],
	"year": 1998,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Kathryn L. Lynch. 1998. Medieval Dream-Poetry. Cambridge University Press.",
	"links": null
	},
	"BIBREF16": {
	"ref_id": "b16",
	"title": "Deep captioning with multimodal recurrent neural networks (m-rnn)",
	"authors": [
	{
	"first": "Junhua",
	"middle": [],
	"last": "Mao",
	"suffix": ""
	},
	{
	"first": "Wei",
	"middle": [],
	"last": "Xu",
	"suffix": ""
	},
	{
	"first": "Yi",
	"middle": [],
	"last": "Yang",
	"suffix": ""
	},
	{
	"first": "Jiang",
	"middle": [],
	"last": "Wang",
	"suffix": ""
	},
	{
	"first": "Zhiheng",
	"middle": [],
	"last": "Huang",
	"suffix": ""
	},
	{
	"first": "Alan",
	"middle": [],
	"last": "Yuille",
	"suffix": ""
	}
	],
	"year": 2014,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {
	"arXiv": [
	"arXiv:1412.6632"
	]
	},
	"num": null,
	"urls": [],
	"raw_text": "Junhua Mao, Wei Xu, Yi Yang, Jiang Wang, Zhiheng Huang, and Alan Yuille. 2014. Deep captioning with multimodal recurrent neural networks (m-rnn). arXiv preprint arXiv:1412.6632.",
	"links": null
	},
	"BIBREF17": {
	"ref_id": "b17",
	"title": "What is poetry? Monthly Repository",
	"authors": [
	{
	"first": "John",
	"middle": [],
	"last": "Stuart Mill",
	"suffix": ""
	}
	],
	"year": null,
	"venue": "",
	"volume": "7",
	"issue": "",
	"pages": "60--70",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "John Stuart Mill. 1833. What is poetry? Monthly Repository, 7(73):60-70.",
	"links": null
	},
	"BIBREF18": {
	"ref_id": "b18",
	"title": "1860. Thoughts on poetry and its varieties. The Crayon",
	"authors": [
	{
	"first": "John",
	"middle": [],
	"last": "Stuart Mill",
	"suffix": ""
	}
	],
	"year": null,
	"venue": "",
	"volume": "7",
	"issue": "",
	"pages": "123--128",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "John Stuart Mill. 1860. Thoughts on poetry and its va- rieties. The Crayon, 7(5):123-128.",
	"links": null
	},
	"BIBREF19": {
	"ref_id": "b19",
	"title": "Crowdsourcing a word-emotion association lexicon",
	"authors": [
	{
	"first": "M",
	"middle": [],
	"last": "Saif",
	"suffix": ""
	},
	{
	"first": "",
	"middle": [],
	"last": "Mohammad",
	"suffix": ""
	},
	{
	"first": "D",
	"middle": [],
	"last": "Peter",
	"suffix": ""
	},
	{
	"first": "",
	"middle": [],
	"last": "Turney",
	"suffix": ""
	}
	],
	"year": 2013,
	"venue": "",
	"volume": "29",
	"issue": "",
	"pages": "436--465",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Saif M. Mohammad and Peter D. Turney. 2013. Crowdsourcing a word-emotion association lexicon. 29(3):436-465.",
	"links": null
	},
	"BIBREF20": {
	"ref_id": "b20",
	"title": "Automatic generation of poetry: an overview",
	"authors": [
	{
	"first": "Hugo",
	"middle": [],
	"last": "Oliveira",
	"suffix": ""
	}
	],
	"year": 2009,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Hugo Oliveira. 2009. Automatic generation of poetry: an overview. Universidade de Coimbra.",
	"links": null
	},
	"BIBREF21": {
	"ref_id": "b21",
	"title": "Poetryme: a versatile platform for poetry generation",
	"authors": [
	{
	"first": "",
	"middle": [],
	"last": "Hugo Gon\u00e7alo Oliveira",
	"suffix": ""
	}
	],
	"year": 2012,
	"venue": "Concept Invention, and General Intelligence",
	"volume": "1",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Hugo Gon\u00e7alo Oliveira. 2012. Poetryme: a versatile platform for poetry generation. Computational Cre- ativity, Concept Invention, and General Intelligence, 1:21.",
	"links": null
	},
	"BIBREF22": {
	"ref_id": "b22",
	"title": "What's basic about basic emotions? Psychological review",
	"authors": [
	{
	"first": "Andrew",
	"middle": [],
	"last": "Ortony",
	"suffix": ""
	},
	{
	"first": "Terence",
	"middle": [
	"J"
	],
	"last": "Turner",
	"suffix": ""
	}
	],
	"year": 1990,
	"venue": "",
	"volume": "97",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Andrew Ortony and Terence J Turner. 1990. What's basic about basic emotions? Psychological review, 97(3):315.",
	"links": null
	},
	"BIBREF24": {
	"ref_id": "b24",
	"title": "Improving language understanding by generative pre-training",
	"authors": [
	{
	"first": "Alec",
	"middle": [],
	"last": "Radford",
	"suffix": ""
	},
	{
	"first": "Karthik",
	"middle": [],
	"last": "Narasimhan",
	"suffix": ""
	},
	{
	"first": "Tim",
	"middle": [],
	"last": "Salimans",
	"suffix": ""
	},
	{
	"first": "Ilya",
	"middle": [],
	"last": "Sutskever",
	"suffix": ""
	}
	],
	"year": 2018,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. 2018. Improving language understanding by generative pre-training. URL https://s3-us-west-2. amazonaws. com/openai- assets/researchcovers/languageunsupervised/language understanding paper. pdf.",
	"links": null
	},
	"BIBREF25": {
	"ref_id": "b25",
	"title": "Language models are unsupervised multitask learners",
	"authors": [
	{
	"first": "Alec",
	"middle": [],
	"last": "Radford",
	"suffix": ""
	},
	{
	"first": "Jeffrey",
	"middle": [],
	"last": "Wu",
	"suffix": ""
	},
	{
	"first": "Rewon",
	"middle": [],
	"last": "Child",
	"suffix": ""
	},
	{
	"first": "David",
	"middle": [],
	"last": "Luan",
	"suffix": ""
	},
	{
	"first": "Dario",
	"middle": [],
	"last": "Amodei",
	"suffix": ""
	},
	{
	"first": "Ilya",
	"middle": [],
	"last": "Sutskever",
	"suffix": ""
	}
	],
	"year": 2019,
	"venue": "OpenAI Blog",
	"volume": "",
	"issue": "8",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. OpenAI Blog, 1(8).",
	"links": null
	},
	"BIBREF26": {
	"ref_id": "b26",
	"title": "Building natural language generation systems",
	"authors": [
	{
	"first": "Ehud",
	"middle": [],
	"last": "Reiter",
	"suffix": ""
	},
	{
	"first": "Robert",
	"middle": [],
	"last": "Dale",
	"suffix": ""
	}
	],
	"year": 2000,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Ehud Reiter and Robert Dale. 2000. Building natural language generation systems. Cambridge university press.",
	"links": null
	},
	"BIBREF27": {
	"ref_id": "b27",
	"title": "Dream poetry as dream work",
	"authors": [
	{
	"first": "A",
	"middle": [],
	"last": "Richard",
	"suffix": ""
	},
	{
	"first": "",
	"middle": [],
	"last": "Russo",
	"suffix": ""
	}
	],
	"year": 2003,
	"venue": "Dreaming",
	"volume": "13",
	"issue": "1",
	"pages": "13--27",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Richard A Russo. 2003. Dream poetry as dream work. Dreaming, 13(1):13-27.",
	"links": null
	},
	"BIBREF28": {
	"ref_id": "b28",
	"title": "A survey of natural language generation techniques with a focus on dialogue systems-past, present and future directions",
	"authors": [
	{
	"first": "Sashank",
	"middle": [],
	"last": "Santhanam",
	"suffix": ""
	},
	{
	"first": "Samira",
	"middle": [],
	"last": "Shaikh",
	"suffix": ""
	}
	],
	"year": 2019,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {
	"arXiv": [
	"arXiv:1906.00500"
	]
	},
	"num": null,
	"urls": [],
	"raw_text": "Sashank Santhanam and Samira Shaikh. 2019. A sur- vey of natural language generation techniques with a focus on dialogue systems-past, present and future directions. arXiv preprint arXiv:1906.00500.",
	"links": null
	},
	"BIBREF29": {
	"ref_id": "b29",
	"title": "Neural machine translation of rare words with subword units",
	"authors": [
	{
	"first": "Rico",
	"middle": [],
	"last": "Sennrich",
	"suffix": ""
	},
	{
	"first": "Barry",
	"middle": [],
	"last": "Haddow",
	"suffix": ""
	},
	{
	"first": "Alexandra",
	"middle": [],
	"last": "Birch",
	"suffix": ""
	}
	],
	"year": 2015,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {
	"arXiv": [
	"arXiv:1508.07909"
	]
	},
	"num": null,
	"urls": [],
	"raw_text": "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2015. Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909.",
	"links": null
	},
	"BIBREF30": {
	"ref_id": "b30",
	"title": "The High Medieval Dream",
	"authors": [
	{
	"first": "A",
	"middle": [
	"C"
	],
	"last": "Spearing",
	"suffix": ""
	}
	],
	"year": 1976,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "A. C. Spearing. 1976a. The High Medieval Dream. Stanford University Press.",
	"links": null
	},
	"BIBREF31": {
	"ref_id": "b31",
	"title": "Medieval dreampoetry",
	"authors": [
	{
	"first": "Anthony Colin",
	"middle": [],
	"last": "Spearing",
	"suffix": ""
	}
	],
	"year": 1976,
	"venue": "CUP Archive",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Anthony Colin Spearing. 1976b. Medieval dream- poetry. CUP Archive.",
	"links": null
	},
	"BIBREF32": {
	"ref_id": "b32",
	"title": "Attention is all you need",
	"authors": [
	{
	"first": "Ashish",
	"middle": [],
	"last": "Vaswani",
	"suffix": ""
	},
	{
	"first": "Noam",
	"middle": [],
	"last": "Shazeer",
	"suffix": ""
	},
	{
	"first": "Niki",
	"middle": [],
	"last": "Parmar",
	"suffix": ""
	},
	{
	"first": "Jakob",
	"middle": [],
	"last": "Uszkoreit",
	"suffix": ""
	},
	{
	"first": "Llion",
	"middle": [],
	"last": "Jones",
	"suffix": ""
	},
	{
	"first": "Aidan",
	"middle": [
	"N"
	],
	"last": "Gomez",
	"suffix": ""
	},
	{
	"first": "Illia",
	"middle": [],
	"last": "Kaiser",
	"suffix": ""
	},
	{
	"first": "",
	"middle": [],
	"last": "Polosukhin",
	"suffix": ""
	}
	],
	"year": 2017,
	"venue": "Advances in Neural Information Processing Systems",
	"volume": "30",
	"issue": "",
	"pages": "5998--6008",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, \u0141 ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Gar- nett, editors, Advances in Neural Information Pro- cessing Systems 30, pages 5998-6008. Curran As- sociates, Inc.",
	"links": null
	},
	"BIBREF33": {
	"ref_id": "b33",
	"title": "Generating lyrics with variational autoencoder and multi-modal artist embeddings",
	"authors": [
	{
	"first": "Olga",
	"middle": [],
	"last": "Vechtomova",
	"suffix": ""
	},
	{
	"first": "Hareesh",
	"middle": [],
	"last": "Bahuleyan",
	"suffix": ""
	},
	{
	"first": "Amirpasha",
	"middle": [],
	"last": "Ghabussi",
	"suffix": ""
	},
	{
	"first": "Vineet",
	"middle": [],
	"last": "John",
	"suffix": ""
	}
	],
	"year": 2018,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Olga Vechtomova, Hareesh Bahuleyan, Amirpasha Ghabussi, and Vineet John. 2018. Generating lyrics with variational autoencoder and multi-modal artist embeddings. CoRR, abs/1812.08318.",
	"links": null
	},
	"BIBREF34": {
	"ref_id": "b34",
	"title": "Poet-based poetry generation: Controlling personal style with recurrent neural networks",
	"authors": [
	{
	"first": "Jia",
	"middle": [],
	"last": "Wei",
	"suffix": ""
	},
	{
	"first": "Qiang",
	"middle": [],
	"last": "Zhou",
	"suffix": ""
	},
	{
	"first": "Yici",
	"middle": [],
	"last": "Cai",
	"suffix": ""
	}
	],
	"year": 2018,
	"venue": "2018 International Conference on Computing, Networking and Communications (ICNC)",
	"volume": "",
	"issue": "",
	"pages": "156--160",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Jia Wei, Qiang Zhou, and Yici Cai. 2018. Poet-based poetry generation: Controlling personal style with recurrent neural networks. In 2018 International Conference on Computing, Networking and Com- munications (ICNC), pages 156-160. IEEE.",
	"links": null
	},
	"BIBREF35": {
	"ref_id": "b35",
	"title": "Eliza-a computer program for the study of natural language communication between man and machine",
	"authors": [
	{
	"first": "Joseph",
	"middle": [],
	"last": "Weizenbaum",
	"suffix": ""
	}
	],
	"year": 1966,
	"venue": "Communications of the ACM",
	"volume": "9",
	"issue": "1",
	"pages": "36--45",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Joseph Weizenbaum et al. 1966. Eliza-a computer program for the study of natural language communi- cation between man and machine. Communications of the ACM, 9(1):36-45.",
	"links": null
	},
	"BIBREF36": {
	"ref_id": "b36",
	"title": "Literary structures in chaucer. The Cambridge Companion to Chaucer",
	"authors": [
	{
	"first": "Barry",
	"middle": [],
	"last": "Windeatt",
	"suffix": ""
	}
	],
	"year": 2003,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "214--232",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Barry Windeatt. 2003. Literary structures in chaucer. The Cambridge Companion to Chaucer, pages 214- 232.",
	"links": null
	},
	"BIBREF37": {
	"ref_id": "b37",
	"title": "Google's neural machine translation system",
	"authors": [
	{
	"first": "Yonghui",
	"middle": [],
	"last": "Wu",
	"suffix": ""
	},
	{
	"first": "Mike",
	"middle": [],
	"last": "Schuster",
	"suffix": ""
	},
	{
	"first": "Zhifeng",
	"middle": [],
	"last": "Chen",
	"suffix": ""
	},
	{
	"first": "V",
	"middle": [],
	"last": "Quoc",
	"suffix": ""
	},
	{
	"first": "Mohammad",
	"middle": [],
	"last": "Le",
	"suffix": ""
	},
	{
	"first": "Wolfgang",
	"middle": [],
	"last": "Norouzi",
	"suffix": ""
	},
	{
	"first": "Maxim",
	"middle": [],
	"last": "Macherey",
	"suffix": ""
	},
	{
	"first": "Yuan",
	"middle": [],
	"last": "Krikun",
	"suffix": ""
	},
	{
	"first": "Qin",
	"middle": [],
	"last": "Cao",
	"suffix": ""
	},
	{
	"first": "Klaus",
	"middle": [],
	"last": "Gao",
	"suffix": ""
	},
	{
	"first": "",
	"middle": [],
	"last": "Macherey",
	"suffix": ""
	}
	],
	"year": 2016,
	"venue": "Bridging the gap between human and machine translation",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {
	"arXiv": [
	"arXiv:1609.08144"
	]
	},
	"num": null,
	"urls": [],
	"raw_text": "Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. 2016. Google's neural machine translation system: Bridging the gap between hu- man and machine translation. arXiv preprint arXiv:1609.08144.",
	"links": null
	},
	"BIBREF38": {
	"ref_id": "b38",
	"title": "2013. I, poet: automatic chinese poetry composition through a generative summarization framework under constrained optimization",
	"authors": [
	{
	"first": "Rui",
	"middle": [],
	"last": "Yan",
	"suffix": ""
	},
	{
	"first": "Han",
	"middle": [],
	"last": "Jiang",
	"suffix": ""
	},
	{
	"first": "Mirella",
	"middle": [],
	"last": "Lapata",
	"suffix": ""
	},
	{
	"first": "Shou-De",
	"middle": [],
	"last": "Lin",
	"suffix": ""
	},
	{
	"first": "Xueqiang",
	"middle": [],
	"last": "Lv",
	"suffix": ""
	},
	{
	"first": "Xiaoming",
	"middle": [],
	"last": "Li",
	"suffix": ""
	}
	],
	"year": null,
	"venue": "Twenty-Third International Joint Conference on Artificial Intelligence",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Rui Yan, Han Jiang, Mirella Lapata, Shou-De Lin, Xueqiang Lv, and Xiaoming Li. 2013. I, poet: au- tomatic chinese poetry composition through a gen- erative summarization framework under constrained optimization. In Twenty-Third International Joint Conference on Artificial Intelligence.",
	"links": null
	},
	"BIBREF39": {
	"ref_id": "b39",
	"title": "Bertscore: Evaluating text generation with BERT. CoRR",
	"authors": [
	{
	"first": "Tianyi",
	"middle": [],
	"last": "Zhang",
	"suffix": ""
	},
	{
	"first": "Varsha",
	"middle": [],
	"last": "Kishore",
	"suffix": ""
	},
	{
	"first": "Felix",
	"middle": [],
	"last": "Wu",
	"suffix": ""
	},
	{
	"first": "Kilian",
	"middle": [
	"Q"
	],
	"last": "Weinberger",
	"suffix": ""
	},
	{
	"first": "Yoav",
	"middle": [],
	"last": "Artzi",
	"suffix": ""
	}
	],
	"year": 2019,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2019. Bertscore: Evaluating text generation with BERT. CoRR, abs/1904.09675.",
	"links": null
	},
	"BIBREF40": {
	"ref_id": "b40",
	"title": "Chinese poetry generation with recurrent neural networks",
	"authors": [
	{
	"first": "Xingxing",
	"middle": [],
	"last": "Zhang",
	"suffix": ""
	},
	{
	"first": "Mirella",
	"middle": [],
	"last": "Lapata",
	"suffix": ""
	}
	],
	"year": 2014,
	"venue": "Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
	"volume": "",
	"issue": "",
	"pages": "670--680",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Xingxing Zhang and Mirella Lapata. 2014. Chinese poetry generation with recurrent neural networks. In Proceedings of the 2014 Conference on Em- pirical Methods in Natural Language Processing (EMNLP), pages 670-680.",
	"links": null
	}
	},
	"ref_entries": {
	"FIGREF0": {
	"type_str": "figure",
	"text": "A high-level overview of our project implementation for emotion eliciting poetry",
	"uris": null,
	"num": null
	},
	"FIGREF1": {
	"type_str": "figure",
	"text": "GPT Architecture. Adapted from(Radford et al., 2018(Radford et al., , 2019",
	"uris": null,
	"num": null
	},
	"FIGREF2": {
	"type_str": "figure",
	"text": "American pyschologist Robert Plutchik's Wheel of Emotions",
	"uris": null,
	"num": null
	},
	"TABREF1": {
	"type_str": "table",
	"content": "<table/>",
	"html": null,
	"text": "Comparison of 5 emotion models trained.",
	"num": null
	},
	"TABREF2": {
	"type_str": "table",
	"content": "<table><tr><td>\u2022 Quality 1: The poem is generally a first-</td></tr><tr><td>person expression</td></tr><tr><td>\u2022 Quality 2: The poem's main substance is</td></tr><tr><td>dream or vision like</td></tr><tr><td>\u2022 Quality 3: The poem recounts or foretells an</td></tr><tr><td>experience or event</td></tr></table>",
	"html": null,
	"text": "Average Likert score of users for each poem",
	"num": null
	},
	"TABREF4": {
	"type_str": "table",
	"content": "<table/>",
	"html": null,
	"text": "Average Coh-Metrix evaluations across 25 randomly selected poems from each model.",
	"num": null
	}
	}
	}
	}