Conversation Chronicles: Towards Diverse Temporal and Relational Dynamics in Multi-Session Conversations

1 Introduction

• Prior open-domain chatbots focus on ongoing single-session dialogue, while earlier multi-session work uses shorter intervals and does not incorporate speaker relationships.The introduction describes time elapsed and speaker relationships as dimensions that shape conversational interactions.Source: §1 · pp. 1–2
• Conversation Chronicles contains 1M multi-session dialogues in 200K five-session episodes, with intervals from hours to years and fine-grained speaker relationships.The introduction says these varied relationships induce different events and interaction flows.Source: §1 · p. 2
• The dataset is generated with LLM prompts specifying relationships, event descriptions, and time intervals, and human evaluators prefer it to other multi-session datasets.The introduction says prompts incorporate those components into generated dialogues.Source: §1 · p. 2
• ReBot uses about 630M parameters to generate dialogue reflecting chronological and relational dynamics in long-term conversations.The model's responses are reported as preferred over those of other chatbot models in long-term settings.Source: §1 · p. 2
• Extensive human evaluation finds that ReBot trained on Conversation Chronicles shows user engagement across varied temporal and relational contexts.The introduction also reports that ReBot responses are preferred over other chatbot models in long-term conversational situations.Source: §1 · p. 2

Open-domain Chatbot

• Earlier open-domain dialogue datasets consist of short, single sessions, limiting their representation of conversations separated by time intervals.The cited examples include DailyDialog, PersonaChat, Wizard of Wikipedia, Empathetic Dialogues, Blended-SkillTalk, Twitter, and Pushshift.io Reddit.Source: §2 · p. 2

Long-term Conversation

• Prior long-term approaches include information-management methods and multi-session datasets, but existing datasets have limited or fixed intervals and lack speaker relationships.The related-work discussion names DuLeMon, PLATO-LTM, MSC, and CareCallmem.Source: §2 · pp. 2–3
• Conversation Chronicles is presented as the first open-domain dialogue dataset defining fine-grained speaker relationships alongside diverse time intervals.The authors contrast this with earlier multi-session datasets' limited or fixed time ranges.Source: §2 · p. 3

Data Distillation

• Prior LLM-based dataset work spans emotional, prosocial, narrative-derived, and single-session dialogue data; the paper also uses LLMs to build a large-scale multi-session dataset.The related-work passage says these studies suggest LLMs can save time and cost while producing high-quality datasets comparable to human-written ones.Source: §2 · p. 3

3 CONVERSATION CHRONICLES

• The dataset contains 1M dialogues in 200K five-session episodes and is larger in scale than other compared multi-session datasets.Its average turns per session is smaller, but its average words per turn is higher than the other datasets.Source: §3 · p. 3, Table 3 · p. 5

3.1 Event Collection

• SODA narratives serve as events; entailment-linked events form directed graphs, from which five-event sequences are retained after removing those sharing more than three events.Event pairs are linked using an NLI classifier, with event order specified by directed edges.Source: §3.1 · p. 3

3.2 Chronological Dynamics

• Time gaps are randomly drawn from approximate intervals spanning a few hours to a couple of years, and episode-level speaker relationships are selected from ten predefined options using ChatGPT.The relationship is chosen to fit the events in the episode rather than assigned randomly.Source: §3.2 · p. 4

3.3 Conversation Episode Generation

• ChatGPT prompts condition each session on its event, time interval, and speaker relationship, while including the events and intervals from previous sessions.Automatic filtering removes generated samples with quality or format problems, and moderation removes harmful data.Source: §3.3 · p. 4, §3.4 · p. 5

3.4 Quality

• 4.34 out of 5 was the average human quality score, and Conversation Chronicles scored higher than MSC across compared metrics except Relationship.Evaluators rated 5K sampled episodes for coherence, consistency, time interval, and relationship.Source: §3.4 · p. 5

4.1 Chronological Summary

• Summarization compresses earlier sessions to preserve their events while avoiding the computational inefficiency of providing the full prior conversation history.The module takes a session dialogue as input and produces a chronological summary.Source: §4.1 · pp. 5–6
• The summarization training data uses ChatGPT-generated summaries of 100K sessions, with T5-base trained on 80K and 20K held for validation and test.The sessions are sampled from 20K Conversation Chronicles episodes.Source: §4.1 · pp. 5–6

4.2 Dialogue Generation

• The BART-large generator conditions the next utterance on speaker relationship, time interval, prior-session summary, and current dialogue context.These inputs are represented as P(c|r, t, s, h), where c is the next utterance and r, t, s, and h denote the relationship, interval, summary, and current context.Source: §4.2 · p. 6

4 REBOT

• REBOT combines chronological summarization and dialogue generation, using about 630M parameters to generate multi-session dialogue reflecting chronological and relational dynamics.The summarization module provides prior-session context to the dialogue generator.Source: §4 · p. 5, §4.2 · p. 6

5.1 Implementation and Training Details

• Training uses 160K episodes, with 20K each for validation and test, and AdamW with cross-entropy loss for generation tasks.The splits are drawn from 200K episodes comprising 1M dialogue sessions.Source: §5.1 · p. 6

5.2 Human Evaluation

• Because reference-based metrics may not suit open-domain dialogue, the study relies on human evaluation of dataset and model dialogue quality.Evaluation considers dimensions including coherence, contradiction, and engagement.Source: §5.2 · p. 6

5.3 Dataset Quality Evaluation

• Dataset quality is assessed by having evaluators rate 5K sampled episodes for coherence, consistency, time interval, and relationship.The criteria are intended to check whether the dialogues reflect events, intervals, and relationships.Source: §5.3 · p. 6

5.4 Comparison to Other Datasets

• For comparison with MSC, the study samples 0.5K episodes from each dataset and averages ratings from three evaluators per episode.The comparison omits relationship because MSC has no speaker relationships, and uses the other shared evaluation metrics.Source: §5.4 · pp. 6–7

5.5 Model Performance Evaluation

• Generated summaries are checked for whether they fully describe speaker interactions, while generated episodes are rated for engagingness, humanness, and memorability.The evaluation samples 3K summaries and 1K generated episodes.Source: §5.5 · p. 7

5.6 Comparison to Other Models

• Interactive comparison uses 50 live chats per model, with in-house evaluators conversing for at least six turns and rating the same criteria.REBOT is compared with MSC 2.7B; evaluators consider the prior-session event summary for REBOT or persona for MSC.Source: §5.6 · p. 7, Figure 5 · p. 7

Dialogue Generation

• Generated dialogue episodes receive high scores across evaluation metrics and are described as natural, engaging, and retaining earlier-session information with few contradictions.The paper relates this consistency to the dataset’s quality evaluation.Source: §6 · p. 7

Interactive Dialogue Generation

• REBOT performs better than MSC 2.7B in interactive evaluation, with the authors associating its advantage with coherence, consistency, time awareness, and speaker relationships.Evaluators also report that predefined relationships make conversations more focused than conversations without a relationship.Source: §6 · p. 7

Speaker Relationship and Time Interval

• Examples show that REBOT varies dialogue flows by speaker relationship, recognizes relationship shifts, and recalls past events across sessions while reflecting elapsed time.A coach and athlete’s relationship shifts to former coach and peer; past-event recall can span beyond the immediately preceding session.Source: §6 · p. 8, Table 8 · p. 9

Chronological Summarization

• Human evaluation gives generated chronological summaries an average score of 4.3 out of 5, which the authors consider sufficient to provide important context for generation.The evaluation covers 3K samples.Source: §6 · p. 9

Ablation Study

• Without time-interval data, the model trends toward generic time references; without relationship data, it fails to maintain a consistent relationship context with the user.Source: §6 · p. 9

7 Conclusion

• Conversation Chronicles is a multi-session dialogue dataset that implements chronological dynamics by integrating time intervals and speaker relationships; its dialogues are distilled from ChatGPT using prompts based on an event graph.Source: §7 · p. 9
• The dataset is built by creating an event graph and distilling dialogues from ChatGPT using prompts based on events, time intervals, and speaker relationships.Source: §7 · p. 9
• Human evaluations report that REBOT generates diverse, coherent responses reflecting different time intervals and speaker relationships, with high engagement and no contradiction in long-term conversations.Source: §7 · p. 9

Limitations

• The study uses a limited number of specific time intervals and speaker relationships, potentially limiting the generalizability of its findings.The authors plan to expand coverage of both.Source: Limitations · p. 9
• Different LLMs can produce different dialogue types and results even under the same framework.The authors plan to consider configurations mixing different LLMs.Source: Limitations · p. 9

Ethics Statement

• Potentially harmful content may remain in the data despite toxic filtering, and generated responses may be unwanted or unsuitable for real-world situations.The authors recommend research use or care in real-world applications, noting medical advice as an example.Source: Ethics Statement · p. 9
• The authors say they checked with the evaluation agency that annotators were compensated fairly.Source: Ethics Statement · p. 10

A Prompts Details

• ChatGPT assigns each episode one of ten predefined fine-grained speaker relationships based on its event description.Options include relationships such as spouses, classmates, and patient and doctor.Source: A Prompts Details · p. 12
• The dialogue prompt supplies the relationship, previous and current session topics, and the interval between sessions, with speaker labels and an instruction to link to prior topics if needed.Source: A Prompts Details · p. 12
• The summary prompt asks ChatGPT to select important events from a conversation and summarize them in two sentences.Source: A Prompts Details · p. 12
• The authors use gpt-3.5-turbo-0301 to make the prompting setup reproducible because responses can vary by model version.Source: A Prompts Details · p. 12

B Dataset Filtering Process

• An episode is removed if any session has more than two speakers, unclear speaker alignment, unlisted speakers, or unnecessary information such as stage directions.Source: B Dataset Filtering Process · p. 12

C Implementation and Training Details

• The summarization and generation modules use T5-base (222M parameters) and BART-large (406M parameters), respectively.Both use linear scheduling and early stopping, with different batch sizes, input and output length limits, and training durations.Source: C Implementation and Training Details · p. 12

D Dialogue Example for Each Relationship

• The examples show ten relationship categories shaping the same homework context into different responses, such as classmate advice or parental empathy.Source: D Dialogue Example for Each Relationship · pp. 12–13
• Human evaluation indicates that speaker relationships help achieve high conversation engagement.Source: D Dialogue Example for Each Relationship · p. 13

E Chronological Summary Example

• The chronological summary examples capture event changes across sessions, including a training-time change and later recollection of an anniversary restaurant.Source: E Chronological Summary Example · p. 13

F Ablation Study Example

• Ablation examples suggest time information yields interval-specific rather than generic replies, while relationship information yields more contextually consistent responses.The authors conclude that both inputs appear to improve contextual appropriateness and specificity.Source: F Ablation Study Example · p. 13

G Generation Quality per Relationship

• REBOT's dialogue-generation scores are described as balanced across all relationship categories.The authors interpret this as effectively mirroring relational dynamics across the categories.Source: G Generation Quality per Relationship · p. 13

H Episode Example

• Figures 6–8 provide examples of a full dataset episode and live chats with REBOT and MSC 2.7B.Source: H Episode Example · p. 13

Human Evaluation

• Evaluators rate full episodes from 1 to 5 for coherence, consistency, time-interval realism, and maintenance of the designated speaker relationship.The criteria assess natural event transitions, avoiding contradictions across sessions, reflecting elapsed time, and sustaining the specified relationship.Source: I Human Evaluation · pp. 13–14
• Model performance is assessed for engagingness, humanness, and memorability, with memorability starting at 3 absent contradictions and adjusted for correct recalls.Engagingness concerns interesting, context-immersed interaction; humanness concerns emotional understanding and natural language and thought processes.Source: I Human Evaluation · p. 14
• Examples contrast relationship-conditioned responses, time-aware replies using specific intervals with generic replies without time information, and inconsistencies when relationship information is removed.Table 9 describes different responses to the same event depending on the relationship; Tables 22 and 23 describe temporal and relational ablation examples.Source: Table 22 · p. 18, Table 23 · p. 19, Table 9

• Figure 8 gives an MSC 2.7B live-chat example in which an athlete asks for feedback, receives encouragement, and asks what training to focus on.The exchange presents the coach as praising the athlete’s improvement and hard work before the athlete asks for further training guidance.Source: Figure 8 · p. 22, I Human Evaluation · p. 23
• Figure 9 depicts the page used by human evaluators.Source: Figure 9 · p. 23
• Figures 10 and 11 show live-chat pages with ReBot and MSC 2.7B, respectively.Source: Figure 10 · p. 23, Figure 11 · p. 23