Wajinimi/Children_Intent_Classification

MAMA Communicative Intent Dataset (INCA-A Annotated)

Overview

The MAMA Communicative Intent Dataset is a linguistically annotated corpus of child utterances designed to support research in child-centred Natural Language Processing (NLP) and communicative intent recognition in early language development.

The dataset contains 10,800 child utterances annotated using the INCA Communicative Coding System (Ninio et al., 1994), a developmental framework that identifies the communicative functions underlying children's speech.

To make the dataset suitable for machine learning, the original INCA codes were mapped to 23 refined intent categories representing distinct communicative behaviours in early child language.

This dataset was created as part of the MAMA (Machine-Assisted Maternal Assistant) research project, which investigates how artificial intelligence systems can better understand the communicative behaviour of young children.

Unlike many NLP corpora that normalise or correct non-standard language, this dataset preserves authentic developmental linguistic features, including telegraphic speech and missing grammatical markers.

---

Dataset Summary

Property	Value
Total utterances	10,800
Total labelled instances	11,410
Intent categories	23
Annotation framework	INCA Communicative Coding System
Source corpus	CHILDES
Language	English
Task	Intent Classification

The dataset reflects naturalistic child language, resulting in class imbalance typical of real-world conversational data.

Example distribution:

Intent	Frequency
Observation / Reference	2,964
Narrative / Storytelling	1,664
Comfort	4

---

Annotation Framework

The dataset is based on the INCA Communicative Coding System, introduced in:

Ninio, A., Snow, C., Pan, B., & Rollins, P. (1994).
Classifying communicative acts in children's interactions.

The INCA system categorises communicative functions in children's speech rather than grammatical structure alone.

In this dataset, INCA codes were mapped into refined NLP intent categories suitable for supervised machine learning.

---

Mapping from INCA Codes to Refined Intent Categories

INCA Category	INCA Code	Refined Intent Category
Directing hearer’s attention	DHA / CL	Attention
Speech elicitation	EI, RT, EA	Imitation
Questions	QN, YQ, TQ	Question
Evaluation	ET	Excitement
Discussing related-to-present	DRP	Narrative or Storytelling
Discussing joint focus	DJF	Observation or Reference
Statements	WS	Desire or Action
Negotiating activity	NIA / DW	Disagreement or Correction
Marking	MRK	Gratitude
Comforting	CMO	Comfort
Directiveness	RP	Request
Directiveness	RD, CS	Refusal
Directiveness	GR	Explanation or Justification
Declaration	YD / AP	Agreement or Acknowledgment
Marking	MK	Greeting
Marking	EM	Distress or Pain
Marking	EN	Emotion
Fantasy discussion	DFW	Playtalk or Fantasy
Possession negotiation	PSS	Possession
Request / Suggest	RP	Need
Dare / Challenge	DR	Command
Disapprove / Protest	DS, ED, DW	Complaint

---

Annotation Protocol

Annotation followed a two-stage validation procedure.

Stage 1 — Initial Annotation

All utterances were initially labelled by the primary researcher using the INCA communicative coding framework.

Stage 2 — Expert Re-annotation

To strengthen validity, the dataset was independently reviewed by two domain experts:

• Developmental Psychologist
• Experienced Early-Years Teacher

This ensured both developmental theoretical grounding and practical child-language expertise.

---

Inter-Annotator Reliability

Agreement between annotators was measured using Cohen's Kappa (κ).

Observed Agreement

[ P_o = \frac{\sum C_{ii}}{N} ]

Where:

• (C_{ii}) = number of rows where annotators assigned the same category
• (N) = total number of annotated rows

Cohen's Kappa

[ \kappa = \frac{P_o - P_e}{1 - P_e} ]

Where expected agreement is defined as:

[ P_e = \sum \left(\frac{R_i}{N} \cdot \frac{C_i}{N}\right) ]

Where:

• (R_i) = rows assigned to category (i) by annotator 1
• (C_i) = rows assigned to category (i) by annotator 2

The resulting score was:

κ = 0.81

According to Landis and Koch (1977), this represents almost perfect agreement, indicating strong reliability in intent categorisation.

---

Linguistic Characteristics

Average Utterance Length

Average utterance length was computed as:

[ \text{Average utterance length} = \frac{\sum_{i=1}^{N} (\text{token length of utterance}_i)}{N} ]

Analysis revealed that:

• Explanation or Justification
• Narrative or Storytelling
• Desire or Action

tend to produce longer utterances, indicating more verbose communicative behaviour.

In contrast:

• Observation or Reference

typically contains shorter utterances, reflecting concise descriptions of objects or events in the shared environment.

---

Lexical Diversity

Lexical diversity analysis showed variation across communicative intents.

Categories such as:

• Observation
• Narrative or Storytelling

exhibited higher vocabulary diversity, reflecting descriptive language use.

Conversely:

• Agreement or Acknowledgement

showed low lexical diversity, as these responses often rely on short, formulaic expressions such as: