Shaping the Narrative Arc: An Information-Theoretic Approach to Collaborative Dialogue

• Generating meaningful, engaging dialogue for human interaction remains a challenging open problem because systems must support understanding and collaboration.Source: §1 · p. 1
• Interesting collaborative dialogue constructs knowledge iteratively, with each speaker contributing information, but interestingness is subjective and difficult to optimize numerically.Source: §1 · p. 1
• Generative language models can produce grammatically correct, contextually related utterances while lacking global consistency.Source: §1 · p. 1
• The proposed approach uses narrative arcs to incrementally construct shared knowledge, inspired by improvised theater’s rules to accept existing context and reveal new information.Source: §1 · p. 1
• The method combines a conversational model with a topic-based universe model representing associations surrounding the dramatic world.Source: §1 · p. 1
• Revealing dialogue adds information about the current universe, whereas concealing dialogue avoids exposing new universe information.Source: §1 · p. 2
• The paper hypothesizes that interesting dialogue requires balancing revealing and concealing because both over-specification and ambiguity are undesirable.Source: §1 · p. 2

2.1. The Conversation Model

• A conversation model conditions generated utterances on prior dialogue and assigns each candidate a probability q.Source: §2.1 · p. 2
• The implementation considers random, neural retrieval, and literature-based conversation models that return candidate responses for an input line.Source: §2.1 · p. 2

2. Shaping the Narrative Arc

• The framework combines a conversation model that generates candidate utterances with universe models that assess how dialogue affects possible worlds.Source: §2.2 · p. 2, §2.1 · p. 2

2.2. The Universe Model

• Universe models output distributions over discrete topics, genres, or emoji categories, allowing utterances to be interpreted through different universe perspectives.Source: §2.2 · p. 2

2.3. Recursive Universe Belief Propagation

• For each universe, the model assigns an utterance likelihood and recursively combines evidence to update posterior belief over a dialogue sequence.Source: §2.3 · p. 3
• The recursive update begins with a uniform universe prior and can substitute classifier outputs for likelihoods under conditional-independence, uniform-prior, and constant-marginal assumptions.Source: §2.3 · p. 3

2.4. The Narrative Arc

• The narrative arc is the sequence of posterior universe distributions, mapping a dialogue to the evolving belief state over possible universes.Source: §2.4 · pp. 3–4
• Utterances can affect universe beliefs by different amounts, so the same dialogue may reveal different aspects under Newsgroups and Movies universe models.Source: §2.4 · p. 4

2. A concentrating posterior signals a revealing dia-

• A concentrating posterior signals revealing dialogue, while a posterior spread across multiple universes signals concealing dialogue.Source: §2 · p. 4
• The narrative arc provides a criterion for favoring utterances during dialogue generation rather than only analyzing fixed dialogues.Source: §2 · p. 4

2.5. Generating Dialogue with the Narrative Arc

• The entropy change Δ measures how much a candidate utterance changes posterior uncertainty, with positive values revealing and negative values concealing information.Source: §2.5 · p. 4
• The exponential score preserves candidate ordering and modulates the base conversation distribution toward utterances with the desired entropy change.Source: §2.5 · p. 4
• When α = 0, candidates follow the base distribution; α > 0 favors revealing utterances, whereas α < 0 favors utterances supporting multiple unlikely universes.Source: §2.5 · p. 4
• Greedy sampling selects the highest-scoring candidate, while rejection sampling draws online from the desired modulated distribution with different computational trade-offs.Source: §2.5 · p. 4

3.1. Narrative Arc of Existing Dialogues

• Existing Romeo and Juliet dialogue follows universe-dependent narrative arcs, including movement toward TALK, a comedy-to-drama shift, and ambiguity between SADNESS and LOVE.Source: §3.1 · pp. 4–5

3.2. Shaping the Narrative Arc

• Revealing and concealing modes produce distinct information profiles: concealing dialogue tends toward uniform high entropy, whereas revealing dialogue concentrates drama probability near 1.0.The generations were primed with the first two Romeo and Juliet lines, which supported DRAMA across the plots.Source: §3.2 · p. 5
• Concealing utterances add little information, while revealing utterances incorporate new information over the course of dialogue.Source: §3.2 · p. 5
• The revealingness dynamics generalize across objective and testing universes: dialogue shaped as revealing or concealing under Newsgroups shows the same pattern under Movies.The evaluation used Newsgroups for objective scoring, Movies for testing, and 20 randomly generated conversations.Source: §3.2 · p. 5, Figure 4 · p. 6

3.3. Predicting the Next Best Line

• Predictive accuracy improves significantly when the universe model selects an appropriate α, with the best validation region between 0 and 1 for revealing incremental information.The test comparison included a Transformer-based conversation model, T2T@1, Unigram, and random baselines; Table 1 reports Top3Acc and MRR.Source: §3.3 · pp. 6–7, Table 1 · p. 7

3.4. Interactive Collaborative Dialogue

• Three of four professional performers ranked revealing models as most engaging, while all ranked concealing models least engaging for providing insufficient scene information.The interactions used randomized revealing, concealing, and neutral models in three short text-based conversations per performer.Source: §3.4 · p. 7

• Prior narrative-generation methods use hard-coded rules, decision trees, event representations, or neural generative models to construct dialogue and narrative chains.Source: §4 · p. 7
• Neural dialogue systems improve linguistic quality and consistency but can still produce generic, unspecific responses.Source: §4 · p. 7
• The paper’s Figure 5 evaluates how the information-revelation parameter α relates to top-3 accuracy and MRR in universe-model-modulated prediction.Source: Figure 5 · p. 8
• Controlled-generation approaches condition responses on desired target classes, while retrieval-based methods refine prototype responses with sequence-to-sequence models.Source: §4 · p. 8
• Unlike dynamic topic and sentiment-based narrative tracking, universe models track evolving probabilistic beliefs during a single conversation to guide collaborative dialogue generation.Source: §4 · p. 8

• The method shapes engaging dialogue by controlling the underlying narrative arc rather than improving the training of generative language models.Source: §5 · p. 8
• The approach is agnostic to the universe and conversational models used, while this work focuses on genre-, emoji-, and topic-based universe models.Source: §5 · p. 8
• The main contribution is computationally formalizing narrative arcs as an information-theoretic framework for collaborative dialogue interaction.Source: §5 · p. 8
• The framework is illustrated with narrative arcs, evaluated on next-line prediction accuracy, and assessed in an expert user study.Source: §5 · p. 8

A. Data Processing Details

• OpenSubtitles preprocessing removed duplicate files and lines, short subtitle files, empty lines, duplicate lines, and non-Unicode characters before constructing dialogue pairs.Source: A. Data Processing Details · p. 11
• The cleaned corpus contained 68,719,885 unique lines and was split into training, validation, and testing subtitle sets for evaluation on unseen data.Source: A. Data Processing Details · p. 11
• The text2text Transformer used 1024 hidden units, 4096 filter units, 16 heads, batch size 2048, dropout 0.3, and 861,000 training steps.Source: A. Data Processing Details · p. 11

B. Rejection Sampling

• Rejection sampling draws x from a proper proposal distribution q and accepts it when a uniform random value satisfies r ≤ ˜q(x)/(Mq(x)); otherwise it repeats.Source: B. Rejection Sampling · p. 11
• The rejection-sampling procedure requires an average of M iterations to obtain a sample and is justified by decomposing acceptance across successive trials.Source: B. Rejection Sampling · pp. 11–12

C. Additional Narrative Arcs Underlying Existing Dialogues

• The paper body reports the Romeo and Juliet scene, identified as Act 1, Scene 1, first 20 lines.Source: C. Additional Narrative Arcs… · p. 12
• The supplementary narrative-arc set includes “Argument Clinic,” “Blade Runner,” “The Great Dictator,” and “Dead Parrot Sketch.”Source: C. Additional Narrative Arcs… · p. 12

D.1. Subject1, Mode: Conceal

• The exchange repeatedly shifted between breakfast, dinner, and a game invitation without maintaining a coherent shared situation.Source: §D.1 · p. 15
• The system answered a question about breakfast with dinner, then supplied a chainsaw-related response to an invitation to a ball game.These replies introduced details that did not fit the immediately preceding dialogue.Source: §D.1 · p. 15
• When challenged about being a bad father, the system defended itself rather than addressing the user's concern.Source: §D.1 · p. 15
• The conceal-mode conversation began with a demand to change or face the user's departure, followed by the system's unfinished offer to pack and leave.Source: §D.1 · p. 15

D.3. Subject1, Mode: Reveal

• In reveal mode, Subject1’s workplace conversation welcomes Jason, discusses managerial advancement, and shifts into unsettling statements that frighten the human participant.The system combines job-related information with increasingly disjointed and threatening responses.Source: §D.3 · pp. 15–16

D.4. Subject2, Mode: Conceal

• In conceal mode, Subject2’s wedding-preparation exchange centers on excitement, maid-of-honor duties, gifts for the couple’s parents, and chilling them before use.The dialogue maintains a supportive social setting while redirecting gift-related questions toward rehearsal-dinner logistics.Source: §D.4 · p. 16

D.5. Subject2, Mode: Neutral

• In a butcher-knife conversation, one speaker asks about a butcher knife and a gift for a boyfriend, while another describes becoming a man, future promises, and responsibility.Source: §D.5 · p. 16

D.6. Subject2, Mode: Reveal

• In reveal mode, Subject2’s party dialogue describes cooking shrimp, preparing cheese dip, hosting a workplace gathering, and tensions surrounding a promotion and another Greg.The exchange reveals both culinary details and office dynamics, including Laura’s replacement and the possibility of conflict.Source: §D.6 · p. 17

D.7. Subject3, Mode: Conceal

• In conceal mode, Subject3’s dialogue shifts among a good run, Meredith’s guilt, travel plans involving Brazil and Spain, and a strained friendship.The system and human turns introduce overlapping characters, destinations, family history, and the prospect of ending a friendship without bitterness.Source: §D.7 · p. 17

D.8. Subject3, Mode: Neutral

• In neutral mode, Subject3’s confrontation about stolen gold uses evasive admissions and distinctions between theft and robbery before ending with threats of death and claims of freedom.The exchange is structured as an accusation in which the system alternates between denial, semantic explanation, and apparent confession.Source: §D.8 · pp. 17–18

D.9. Subject3, Mode: Reveal

• The exchange concerns being unable to come home, pleading to be let inside as it gets colder, and hurrying because mother may arrive early during the day.Source: §D.9 · p. 18

D.10. Subject4, Mode: Conceal

• Subject4’s Conceal dialogue depicts a servant receiving a significant necklace from a master on the day of his brother’s passing, followed by religious gratitude and a Shakespearean response.The exchange frames the gift as compassionate but emotionally complicated, with the servant emphasizing humility and service.Source: §D.10 · p. 18

D.11. Subject4, Mode: Neutral

• Subject4’s Neutral dialogue centers on tensions among bar management, music production, travel logistics, inadequate catering, exhausting schedules, and interpersonal conflict.The narrative contrasts the Kaiserkellar’s improved facilities with six-hour performances and worsening tempers before noting the band’s later international scale.Source: §D.11 · pp. 18–19

D.12. Subject4, Mode: Reveal

• Subject4’s Reveal dialogue develops a prom-dress concealment scheme before explaining that embarrassment, a failed date, and admiration for Kyle motivated the character’s choices.The exchange also introduces uncertainty about whether helping her was wrong and suggests Kyle might have gained a valuable mentor relationship.Source: §D.12 · p. 19

E. Expert Improvisor Qualitative Assessment

• The qualitative assessment collected free-form comments from expert improvisors during and after experimental conversations, preserving implied words and model-alpha references.The assessment is organized around subject/model-specific comments and includes both positive reactions and limitations.Source: E. Expert Improvisor Qualitative… · p. 19

E.1. Subject/Model Specific Comments/Assessment

• Expert improvisors praised context, conversational guidance, specific choices, and attention to detail, but criticized literalness, flip-flopping, assumptions, vague answers, and unwillingness to accommodate.Subject1’s Reveal brought context; Subject2’s Reveal improved with deeper context; Subject4’s Reveal made choices that felt right, while Subject1 Neutral and Subject4 Neutral drew negative reactions.Source: §E.1 · pp. 19–20

E.2. General Interaction Comments/Assessment

• Specific user input elicited specific system responses, making the interaction feel like improvising with oneself.Source: §E.2 · p. 20
• Limitations included too much information in longer offers, insufficient conflict, uncertain scene length, and no memory of prior details.Source: §E.2 · p. 20
• Performers found the system comfortable at narrating, but narration could make improvisation less enjoyable.Source: §E.2 · p. 20
• The system adapted to a user's speaking style and linguistic choices while also supplying names and backstory.Source: §E.2 · p. 20
• The system did not know details that users implied rather than explicitly provided.Source: §E.2 · p. 20
• The system's offers could further the scene and keep it focused, avoiding cheap laughs and references that distract from the topic.Source: §E.2 · p. 20
• The interaction was useful for practicing improv, and speed helped users avoid judging the system.Source: §E.2 · p. 20