Conversational AI › Visualizations

Michael BrenndoerferMarch 1, 20267 min read
Language AI HandbookAll chapters
Part 63 of 67LLM Applications
Chapter 483Conversational AI › Visualizations
Reading section 10 of 12 · Visualizations · 7 minSection 10 of 12 · 7 min

Book contents

Language AI Handbook67 parts · 513 chaptersView the whole book →
Active chapter 483 of 513

Section 10 of 12 · 7 min read

Conversational AI › VisualizationsOne section · Sections in this chapter10 / 12 ↓
Sections in this chapter12
Reading tools Definitions & inline links

Choose your expertise level to adjust how many terms are explained. Beginners see more tooltips, experts see fewer to maintain reading flow. Hover over underlined terms for instant definitions.

Article links

Make inline references clickable

Let's visualize some key aspects of conversational AI systems to build intuition.

Out[12]:
Visualization
Grouped bar chart showing frequency of dialogue acts for user and assistant turns.
Distribution of dialogue acts across user and assistant turns in a typical task-oriented conversation. Users primarily request and inform; the assistant primarily informs, confirms, and clarifies. Understanding this distribution helps tune system prompts toward the right mix of responses.
Out[13]:
Visualization
Bar chart comparing three dialogue system architectures across four capability dimensions.
Comparison of rule-based, statistical pipeline, and LLM-based dialogue system architectures across four capability dimensions. Rule-based systems lead on controllability and data efficiency but lag on generalization and domain coverage. LLM-based systems reverse this tradeoff, excelling at generalization while requiring more data and offering less explicit control.
Out[14]:
Visualization
Line chart showing token count vs conversation turns for three memory management strategies.
Context window token growth across conversation turns under three memory strategies: full history, sliding window (last 5 turns), and summarization. Full history grows linearly without bound; the sliding window caps at a fixed token cost; summarization keeps costs nearly constant by replacing old turns with a compact summary once the compression trigger fires.
Out[15]:
Visualization
Stacked bar chart showing slot filling status across conversation turns.
Slot filling progress across conversation turns for a restaurant booking task. Required slots (date, time, party size, cuisine) are filled first; optional slots (dietary restrictions, price range) remain unfilled unless the user volunteers the information. By turn 4, all required slots are collected and the booking can proceed.
Out[16]:
Visualization
Scatter plot of BLEU scores vs human ratings showing near-zero correlation.
BLEU score vs. human quality rating for simulated dialogue responses. The deliberately near-zero correlation illustrates how n-gram overlap can fail to distinguish responses that humans judge differently, so BLEU should not be the primary quality metric for a conversational system.
Scatter plot of LLM evaluator scores vs human ratings showing strong positive correlation.
LLM evaluator score vs. human quality rating for the same simulated responses. The strong positive correlation illustrates how model-based evaluation can track human judgments more closely, while still requiring regular validation against fresh human ratings.