对话状态跟踪

CUISINE_SYNONYMS = {
    "italian": ["italian", "pasta", "pizza", "italy"],
    "chinese": ["chinese", "chow mein", "noodles"],
}


def extract_cuisine(utterance):
    for canonical, synonyms in CUISINE_SYNONYMS.items():
        if any(syn in utterance.lower() for syn in synonyms):
            return canonical
    return None

def update_state(state, utterance):
    new_state = dict(state)
    for slot, extractor in SLOT_EXTRACTORS.items():
        value = extractor(utterance)
        if value is not None:
            new_state[slot] = value
    for slot in NEGATION_CLEARS:
        if is_negated(utterance, slot):
            new_state[slot] = None
    return new_state

from pydantic import BaseModel
from typing import Literal, Optional
import instructor

class RestaurantState(BaseModel):
    cuisine: Optional[Literal["italian", "chinese", "indian", "thai", "any"]] = None
    area: Optional[Literal["north", "south", "east", "west", "center"]] = None
    price: Optional[Literal["cheap", "moderate", "expensive"]] = None
    people: Optional[int] = None
    day: Optional[str] = None


def llm_dst(history, llm):
    prompt = f"""You track the slot values of a restaurant booking across turns.
Dialogue so far:
{render(history)}

Update the state based on the latest user turn. Output only the JSON state."""
    return llm(prompt, response_model=RestaurantState)

def joint_goal_accuracy(predicted_states, gold_states):
    correct = sum(1 for p, g in zip(predicted_states, gold_states) if p == g)
    return correct / len(predicted_states)

CORRECTION_CUES = {"actually", "no wait", "on second thought", "change that to"}


def is_correction(utterance):
    return any(cue in utterance.lower() for cue in CORRECTION_CUES)

---
name: dst-designer
description: Design a dialogue state tracker — schema, extractor, update policy, evaluation.
version: 1.0.0
phase: 5
lesson: 29
tags: [nlp, dialogue, task-oriented]
---

Given a use case (domain, languages, vocab openness, compliance needs), output:

1. Schema. Domain list, slots per domain, open vs closed vocabulary per slot.
2. Extractor. Rule-based / seq2seq / LLM-with-Pydantic. Reason.
3. Update policy. Regenerate-whole-state / incremental; correction handling; negation handling.
4. Evaluation. Joint Goal Accuracy on a held-out dialogue set, slot-level precision/recall, confusion on the hardest slot.
5. Confirmation flow. When to explicitly ask the user to confirm (destructive actions, low-confidence extractions).

Refuse LLM-only DST for compliance-sensitive slots without a rule-based secondary check. Refuse any DST that cannot roll back a slot on user correction. Flag schemas without version tags.