Follow-Up Detection: 20 Patterns That Make Chatbots Conversational

When a user types "how?" or "tell me more" or just sends a thumbs-up emoji, the chatbot needs to understand this is a follow-up to the previous response. ReLU.chat detects 20+ follow-up types using a combination of regex patterns and heuristic matching.

Pattern Categories

The follow-up detection system handles these categories:

Process/Reason: "how", "why", "how so", "how does that work" Depth: "more detail", "elaborate", "go deeper", "deep dive" Simplification: "explain simpler", "ELI5", "dumb it down" Examples: "give an example", "like what", "for instance" Continuation: "go ahead", "keep going", "what's next" Challenge: "really", "are you sure", "prove it" Acknowledgment: "ok", "got it", "makes sense" Clarification: "huh", "what do you mean", "I'm confused" Summarization: "tl;dr", "summarize", "brief" Comparison: "compare", "difference", "versus" Topic correction: "I meant X", "no, just X"

Two-Phase Detection

Phase 1: Pattern matching against 20+ regex patterns (fast, high-confidence). Phase 2: Heuristic fallback for short queries (< 20 chars) using substring matching.

// Phase 1: exact pattern match
for (const pat of FOLLOWUP_PATTERNS) {
  const match = query.match(pat.regex);
  if (match) return { isFollowUp: true, type: pat.type };
}
// Phase 2: heuristic short-query matching
if (query.length < 20) { / substring maps / }

Emoji and Symbol Detection

Emoji-only messages are common in chat. The system detects them via Unicode property escapes:

const allSymbols = /^[\u{1F000}-\u{1FFFF}...]+$/u;

Common emoji map to follow-up types: 🤔 → clarify, 👍 → acknowledge, 💡 → elaborate.

Punctuation Signals

"..." → elaborate, "??" → clarify, "!" → challenge. Even empty punctuation carries meaning.

Integration with Policy

The detected follow-up type feeds into the 25-feature policy vector as followUpType (feature index 18, values 0-22). The MLP uses this to adjust response strategy — simplify triggers shorter, more intuitive responses; deep_dive triggers longer, more formal ones.

Key Takeaway

Follow-up detection is pattern matching, not ML. Twenty well-crafted regex patterns cover 95% of conversational follow-ups. The remaining 5% are handled by heuristic substring matching.