MaiChat: A Text-based Dialogue Corpus Rich In Conversational Features

MaiChat is an English corpus of typed instant-messaging dialogues designed to capture spontaneous, interaction-rich written conversation. The dataset contains 50 dyadic conversations between 100 native English speakers (4,975 messages in total), collected via a custom-built real-time chat platform that records not only final message content but also detailed behavioural and temporal metadata. Each message is accompanied by a keystroke-level typing log — over 189,000 timestamped text-state snapshots with millisecond precision — from which inter-keystroke intervals, deletions, insertions, self-corrections, typing pauses, send delays, and turn timing can be derived, alongside per-message device information (desktop, mobile, tablet). The corpus is particularly rich in conversational phenomena commonly observed in spontaneous interaction, including onomatopoeic and lexical laughter, emojis and emoticons, filled pauses, expressive respellings, intentional capitalisation, and self-corrections. Conversations were collected between participants who know each other well, encouraging natural and informal exchanges. Designed with a transparent data collection protocol, MaiChat enables scalable collection while ensuring privacy and consent. The dataset supports research in conversational AI, cognitive modelling, human–computer interaction, behavioural biometrics, and the study of spontaneous written communication.