Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

CHIP – Child/Parent Interaction Profile

Status: Current Last updated: 2026-05-11 17:18 EDT

Purpose

Analyzes interaction patterns between a child speaker and their conversational partners. Categorizes successive utterance pairs to measure imitation, repetition, and overlap. CHIP is commonly used in child language research to quantify how much a child imitates or echoes their interlocutor.

Usage

chatter clan chip file.cha
chatter clan chip --speaker CHI file.cha
chatter clan chip --format json file.cha

Options

OptionCLAN FlagDescription
--speaker <CODE>+t*CHIInclude speaker
--exclude-speaker <CODE>-t*CHIExclude speaker
--format <FMT>–Output format: text, json, csv

Display Modes (+dN / --display-mode N) — DRAFT, awaiting PI review

Status: drafted from CLAN manual; not yet implemented. The rewriter at crates/clan/clan-core/src/clan_args.rs:101 translates +dN → --display-mode N, but no clap field consumes that token today. Drafted from CLAN manual §7.4.5 (Unique Options, CHIP) for PI review.

NCLAN behavior (verbatim from manual)
+d (no number)“Using +d with no further number outputs only coding tiers, which are useful for iterative analyses.”
+d1“Using +d1 outputs only summary statistics, which can then be sent to a statistical program.”

Open questions for PI review

  1. CHIP’s +d shape (coding-tier-only vs summary-only) is orthogonal to the FREQ/KWAL/MLU shape (output format selectors). Should the --display-mode enum’s variants be CHIP-specific (coding-tiers / summary) or share a name space with FREQ’s variants?
  2. “Useful for iterative analyses” implies the coding-tier output is intended to be piped to another chatter clan command. Should chatter prefer making this the default JSON output shape, with --display-mode summary collapsing to just the matrix?

Interaction Categories

For each adjacent utterance pair (speaker A followed by speaker B):

CategoryCondition
Exact repetitionB’s words are identical to A’s (order-independent)
OverlapB shares >= 50% of words with A (smaller unique-word set as denominator)
No overlapB shares < 50% of words with A

Only cross-speaker adjacency is considered; consecutive utterances by the same speaker do not produce interaction records. Adjacency state is reset at file boundaries.

Output

36-measure matrix format structurally matching CLAN:

  • ADU (adult) / CHI (child) / ASR (adult-speech-related) / CSR (child-speech-related) columns
  • Per directed speaker pair (MOT->CHI is distinct from CHI->MOT)
  • Counts and percentages for each interaction category
  • Grand totals across all pairs

Current implementation status (2026-05-12): the matrix header, row labels, and ADU/CHI/ASR/CSR column layout are implemented and written by chip.rs:246-260. The actual measure values are written as zeros — chip.rs:253 carries a comment “currently all zeros — full computation not yet implemented.” So the structural parity is in place but the analytical numbers are not yet computed. See §Golden test parity below for the consequence.

Echo behavior

When displaying matched utterances, CHIP echoes:

  • Main tier text
  • %mor tier (if present)

It does not echo %gra tiers, matching CLAN’s behavior.

Differences from CLAN

Matrix format

Uses the exact 36-measure matrix format with ADU/CHI/ASR/CSR columns. Header, row labels, and column structure match CLAN character-for-character; per §Output, the measure-value cells are currently emitted as zeros pending full implementation.

Echo content

Echoes main tier + %mor only (not %gra tiers), matching CLAN.

Word identification

Uses AST-based is_countable_word() instead of CLAN’s string-prefix matching. Overlap comparison operates on parsed word content, not raw text.

Output formats

Supports text, JSON, and CSV. CLAN produces text only.

Golden test parity

No CHIP-specific golden tests are currently checked in (rg 'chip_golden|chip.*golden' returns nothing in tests/ and crates/clan/clan-core/). The previous “100% parity” claim was incorrect given the stub measure values described above. Once the 36 measures are computed, golden tests should be added against a real CLAN CHIP run.