Controllable Lexical Alignment in Dialogue Systems: Data

This record contains the data of our study on weighted decoding for controllable lexical alignment in dialogue systems. It contains the 10,545 dialogues of the automatic evaluation, the alignment and perplexity scores, the rankings of the LLM judges, the dialogues of a generalisation sweep with newer models, and the materials and anonymised responses of the human evaluation (50 participants, pseudonymised). We license the data under CC BY 4.0. Three folders contain turns from DailyDialog, and they keep the DailyDialog licence (CC BY-NC-SA 4.0). README.md lists these folders. We report the study in the following paper, and in Chapter 6 of the dissertation of Sumit Srivastava (University of Twente): Sumit Srivastava, Mariët Theune, and Alejandro Catalá. Controllable Lexical Alignment in Dialogue Systems. Accepted at the International Natural Language Generation Conference (INLG 2026).