Cues are unspoken lines that can drastically effect how the dialogues are perceived by the audience. A cue can be a trigger for an action, an informative description of the stage, or thoughts or body language of the characters.

Given dialogue lines as input, we use a fine-tuned GPT-2 model, ILM and PPLM frameworks to generate cues. In addition to cue generation, we extend the PPLM framework with an automatic topic keyword extraction module and an emotion-based attribute model to generate text with the target topics or emotions.

Cue/Dialogue Discriminator

• We train a single-layer cue/dialogue classifier using 10% of our data
• We use the trained classifier to steer the generation towards the higher log-likelihood of dialogues or cues as specified by the user

Topic Modeling with LDA

• We use 10% of our dataset for topic modeling and automatic topic keyword extraction
• WA target topic selected by the user is then used to steer the language generation process to maximize the log-likelihood of the extracted keywords of the target topic

Emotion Classifier

• We use DeepMoji to create an emoji-labeled subset with 10% of our data
• We map the predicted emojis to a subset of emotions from Plutchik’s Wheel of Emotions, such as "happy", "angry", "disappointed", and train a multi-label emotion classifier
• We use the trained classifier to steer the generation towards generating text with the target emotion

• GPT-2+ FT: Given a line of dialogue as input, we use a fine-tuned GPT-2 medium model to generate text.
• ILM: ILM enables LMs to infill variable-length spans using both preceding and subsequent text. We fine-tune the GPT-2 small model on dialogue - cue - dialogue line triples. The cue is masked during training and sampling since our goal is to generate cues. Infilling is performed by using the preceding and succeeding dialogues.
• PPLM+LDA: We detect the topic of the input dialogue line using LDA and steer the generation towards generating text with the same topic.
• PPLM+CueDisc: We train a cue/dialogue sentence type discriminator and control the generation process using this classifier.
• PPLM+Emotion: We train a multi-label emotion classifier and steer the generation process to generate text that reflects the target emotion specified by the user

We use n-gram similarity (LCSR and BI-SIM) and distance metrics to measure the similarity of the generated text to a reference cue corpus. We measure the diversity of the text generated by each model by the number of distinct n-grams (normalized by the length of text) and report the Dist-1, Dist-2, and Dist-3 scores.

In this paper, we use a large-scale play script dataset and propose the novel task of generating theatrical cues from dialogues. We approach the cue generation problem as a controlled text generation task and use a plug-and-play language model with a cue/dialogue discriminator, LDA-based topic keyword lists, and a multi-label emotion classifier to steer the language model to the desired attributes without re-training the model. Our experiments show that language models can be successfully used to generate plausible and attribute-controlled text in highly specialized domains such as plays. In the future, we plan to explore character and person-based cue and dialogue generation tasks with plug-and-play models.