Dynamic Hard Classification – Adaptive Mapping (DHC-AM) Dialogue Model

Core Idea:
AI-human dialogue systems are not static classifiers, but dynamic person perception systems.
Misjudging or misclassifying a user is normal and inevitable.
The key is: can the system, like a human, correct itself in time and reclassify?

Deficiencies of traditional AI (including large models):

· Either one-time classification, never updated
· Or black-box fuzziness, never admitting mistakes
· No mechanism for "I misjudged you, let me re-evaluate you"

The proposed mechanism:
Classification → Mapping → Error detection → Reclassification → New mapping

This is human-level social intelligence, not machine Q&A.

---

DHC-AM Formal Description

I. Basic Symbols

· X: User feature space
· C = \{c_1, c_2, \dots, c_n\}: Set of user hard classifications
· Q: User request
· Y: AI response
· \tau: Confidence threshold
· D: Mismatch detection function; D=1 requires reclassification, D=0 keep current class
· t: Time step

---

II. Initial Hard Classification (t=0)

\hat{c}_0 =
\begin{cases}
c_k, & \text{if priority rules match} \\
\arg\max\limits_{c_i \in C} P(c_i \mid x), & \text{if } \max P \geq \tau \\
\varnothing, & \text{otherwise}
\end{cases}

---

III. Mapping

y_0 = g(\hat{c}_0, q)

---

IV. Mismatch Detection

D(\hat{c}_{t-1}, q_t, y_{t-1}, \text{user\_feedback}) = 1 \text{ or } 0

---

V. Dynamic Reclassification (t \geq 1)

\hat{c}_t =
\begin{cases}
c_k, & \text{new rule match} \\
\arg\max\limits_{c_i \in C} P(c_i \mid x_{\text{hist}} + x_t), & \text{if } \max P \geq \tau \\
\varnothing, & \text{otherwise}
\end{cases}

---

VI. Adaptive Mapping

y_t = g(\hat{c}_t, q_t)

---

VII. Overall Closed Loop

y_t = g\big( C(X_t), q_t \big)

---