Controlled Cue Generation for Play Scripts

• Cues enrich dialogue by conveying actions, stage information, moods, thoughts, reactions, and silent gestures.Source: §1 · p. 1
• The paper proposes generating theatrical cues from dialogue, a task it identifies as novel and previously unexplored.Source: §1 · p. 2
• The dataset contains over 1500 play scripts with 775,000 dialogue lines and more than 277,000 cues across varied topics.Source: §1 · p. 2
• The approach uses transformer language models with a cue/dialogue discriminator and also explores emotion- and topic-controlled generation.Source: §1 · p. 2

2.1 Text generation

• Transformer-based language models support controlled generation, but many prior approaches require separate fine-tuning for each attribute; this work instead uses plug-and-play steering.Source: §2.1 · p. 2

2.2 Story Generation

• Story-generation systems decompose events and sentences or condition recurrent generators on extracted controls, but new intents generally require separate models.Source: §2.2 · p. 2

2.3 Dialogue Systems

• Dialogue systems commonly extract user intent before generating intent-dependent responses using conditional generation models.Source: §2.3 · p. 3

2.4 Play Script Generation

• Theater play generation has received less attention than other creative-generation tasks, and fine-grained controllable play-script generation remains unexplored.Source: §2.4 · p. 3
• THEaiTRE fine-tunes GPT-2 on a small bilingual script dataset, but its dataset, experiments, and generated scripts were not released for evaluation.Source: §2.4 · p. 3

• The corpus contains 1511 English-language scripts with over 775,000 dialogue lines and over 277,000 cues across themes including Comedy, Romance, Satire, and Greek.Source: §3 · p. 3
• Play scripts are organized into acts and scenes, with scenes representing actions in a shared place and time among a small set of characters.Source: §3 · p. 3
• Dialogue lines begin with capitalized character names, while cues appear in parentheses and are retained from pages containing both dialogue and cues.Source: §3 · p. 3
• Cues express stage directions, descriptions, character states, moods, feelings, thoughts, and silent expressions, and can reinterpret plays without changing dialogue.Source: §3 · p. 4
• 11K of 227K cues contain the word “stage”; cues range from zero to 10 character names and from zero to 20 actions.Source: §3 · p. 4, Figure 1 · p. 4

4 Methodology

• PPLM combines GPT-2 with attribute models and shifts hidden representations toward the desired attribute while preserving fluency through KL-divergence minimization.The framework creates conditional generation without retraining the underlying language model.Source: §4 · pp. 4–5

4.1 Controlled Generation using Cue/Dialogue Discriminator

• A cue/dialogue discriminator steers generation toward the sentence type selected by the user.The binary classifier is trained on 10% of the dataset and its likelihood guides PPLM activation shifts.Source: §4.1 · p. 5, §4 · p. 5

4.2 Controlled Generation using LDA

• LDA automatically extracts keywords from ten cue topics, enabling user-selected topics to steer generation toward topic-specific cues.The target topic maximizes the likelihood of its extracted keywords during generation.Source: §4.2 · p. 5, §4 · p. 5

4.3 Controlled Generation with Emotions

• DeepMoji-derived emotion labels train a multi-label classifier that steers generated text toward a user-specified emotion.The emotion classifier predicts labels from dialogue lines mapped to a subset of Plutchik’s emotions and does not necessarily generate cues.Source: §4.3 · p. 6, §4.4 · p. 7

4.4 Experimental Setup

• The experiments compare GPT-2+FT, ILM, PPLM+LDA, PPLM+CueDisc, and PPLM+Emotion using dataset splits reserved for generation, steering, and attribute modeling.PPLM uses a GPT-2 345M model fine-tuned on 80% of the dataset, while ILM infills masked dialogue-cue-dialogue triples.Source: §4.4 · pp. 6–7

4.5 Quantitative Results

• PPLM+CueDisc achieves the highest LCSR and BI-SIM scores, followed by PPLM+Emotion and PPLM+LDA, while GPT-2+FT and ILM score low.The results indicate that the discriminator best generates cues resembling the reference cue corpus.Source: Table 2 · p. 7, §4.5 · p. 7
• PPLM+CueDisc and PPLM+Emotion are comparable to or better than GPT-2+FT in diversity, whereas ILM and PPLM+LDA perform worst.LDA sometimes produces repetitive non-character stage-direction text, and ILM also tends to generate repetitive stage directions.Source: §4.5 · p. 7, Table 3 · p. 8

4.6 Qualitative Results

• Human evaluation finds that PPLM+CueDisc achieves high cue accuracy while preserving overall conversational coherence.GPT-2+FT generates diverse text but fails to generate cues, motivating future improvements through preprocessing and persona-based discriminators.Source: §4.6 · p. 8, Table 4 · p. 8

• The authors propose generating theatrical cues from dialogue and use plug-and-play language models with cue/dialogue, topic, and emotion controls.Source: §5 · p. 8