Conversation Analysis

Once conversations are collected, ChatbotLab data can be exported and analyzed with external libraries for linguistic and behavioral research.

The DLATK, ConvoKit, and text (R) tutorials all run on the same synthetic corpus, in conversation_data/. This corpus contains 19 synthetic person-to-AI conversations, generated with the Anthropic API. Each “person” speaker is tagged with a PHQ-9 depression score, along with age, gender, and a persona. No real participants are included; see conversation_data/README.md for details. The corpus ships in three formats: a ConvoKit export (conversation_data/convokit/), a flattened, DLATK-style export (conversation_data/text/), and a SQLite database built from that export (conversation_data/dlatk/).

The three tutorials use three different toolkits and three different methods: DLATK’s frequency correlation, ConvoKit’s Fighting Words, and the R text package’s Supervised Dimension Projection. All three find the same pattern in this corpus. Higher-PHQ-9 participants use more affect and hedging language, such as feel and just. Lower-PHQ-9 participants use more upbeat, closing language, such as alright and ok.

• DLATK
• ConvoKit
• text (R)