Social-Cognitive Annotation

Model-specific prompts, generated responses, derived SCET aggregates, stability-analysis variables, and other intermediate experimental files are not part of the dataset release.Data SourceThe dialogue content is derived from DuLeMon, introduced by Xu et al. (2022):Xu, X., Gou, Z., Wu, W., Niu, Z.-Y., Wu, H., Wang, H., & Wang, S. (2022). Long Time No See! Open-Domain Conversation with Long-Term Persona Memory.

Findings of the Association for Computational Linguistics: ACL 2022, 2639–2650.DuLeMon was developed for research on open-domain multi-turn dialogue with long-term persona information.The present resource does not introduce the underlying dialogue text as a newly collected corpus. Instead, it provides a processed and socially annotated subset of DuLeMon for the experiments reported in our study.Social-Cognitive AnnotationsThe released data extend the dialogue representation with multiple categories of social-cognitive information.EmotionEmotion is represented using six continuous dimensions:angry fear happy neutral sad surprise These variables provide continuous affective information rather than a single discrete emotion category.For the analyses reported in the paper, emotional activation is operationalized asEt=1−neutralt.E_t = 1-\text{neutral}_t.The original six-dimensional representation remains available in the released CSV files.PersonalityPersonality-related information is represented using 16 MBTI-oriented dimensions:INTJ INTP ENTJ ENTP INFJ INFP ENFJ ENFP ISTJ ISFJ ESTJ ESFJ ISTP ISFP ESTP ESFP These variables are continuous computational representations and should not be interpreted as clinically validated or human-verified personality assessments.The dataset additionally contains latent personality-related variables used in the original annotation and analysis pipeline.Conversational StyleConversational-style information is represented separately for the user and assistant sides of the dialogue.The annotated style dimensions include:warmth emotional intensity politeness humor supportiveness self-disclosure formality verbosity Corresponding fields are provided using the user_* and bot_* prefixes.These variables make it possible to analyze conversational-style consistency and variation across dialogue turns.Relationship InformationThe dataset contains interaction-level relationship annotations covering multiple dimensions, including:trust intimacy familiarity dependency affection engagement For the operational analyses reported in the paper, the relationship state is summarized asRt=16∑j=16rt,j.R_t=\frac{1}{6}\sum_{j=1}^{6}r_{t,j}.These variables should be interpreted as computational interaction-level annotations rather than direct measurements of human psychological relationships.Memory-Related InformationThe released data contain memory- and event-related information, including variables such as:event_type event_importance event_people event_time_reference memory_decay_score turn_embedding_id For the operational analyses reported in the paper, memory activation is defined asMt=event_importancet×memory_decay_scoret.M_t= \text{event\_importance}_t \times \text{memory\_decay\_score}_t.Categorical or identifier fields such as event_type, event_people, event_time_reference, and turn_embedding_id are not converted directly into scalar memory values in this formulation.Interaction-Dynamics VariablesThe dataset also contains variables characterizing turn-level changes and interaction dynamics, including emotion shifts, emotion-transition information, personality-related change variables, and other derived annotations.Examples include:angry_shift fear_shift happy_shift neutral_shift sad_shift surprise_shift dominant_emotion_shift emotion_delta_magnitude emotion_transition_type These variables are primarily intended for retrospective analysis of conversational trajectories.Important Distinction: Annotation and Causal GenerationThe released CSV files contain a broader set of annotations than those used as causal inputs for response generation.This distinction is important.Some variables describe the current assistant response or are retrospectively derived from the completed dialogue turn.

Such variables must not be treated as information available before generating that response.In the response-generation experiments reported in the paper, target-derived current-turn information was excluded from the conditioning input. Examples include current assistant-response style, response-derived emotion transitions, bot_strategy, and retrospective or target-dependent variables.The generation experiments therefore use only the social-cognitive information considered causally available at the corresponding generation step.Conceptually,(ut,Dt−1,Qtgen)⟶r^t,(u_t,D_{t-1},Q_t^{\mathrm{gen}}) \longrightarrow \hat r_t,where utu_t is the current user utterance, Dt−1D_{t-1} is the observed preceding dialogue history, and QtgenQ_t^{\mathrm{gen}} denotes the causally available social-cognitive conditioning information.Researchers using this dataset for generative modeling should independently preserve this distinction to avoid target leakage.Social Cognitive EnergyThe released CSV files provide the component annotations rather than a precomputed SCET energy column.For retrospective analysis, the reference operational formulation used in the paper isSt=0.2Et+0.2Rt+0.2Mt+0.2ItC+0.2ItP,S_t = 0.2E_t + 0.2R_t + 0.2M_t + 0.2I_t^C + 0.2I_t^P,where EtE_t, RtR_t, and MtM_t represent emotional activation, relationship state, and memory activation, respectively, while ItCI_t^C and ItPI_t^P represent conversational-style and personality deviation.The equal weights constitute an operational reference setting rather than empirically established psychological coefficients.SCET energy should therefore be interpreted as a computational aggregate of social-cognitive variables, not as physical energy or a direct psychological measurement.Intended UsesThe dataset may support research on:socially conditioned dialogue generation;emotion- and personality-aware text generation;conversational-style modeling;relationship-aware dialogue modeling;memory-enhanced conversational systems;multi-turn social-cognitive trajectory analysis;computational modeling of conversational dynamics;reproducibility studies related to the accompanying paper.The annotations may also support alternative operationalizations beyond those proposed in the accompanying study.Important LimitationsThe social-cognitive variables contained in this dataset are automatically constructed operational annotations.

They should not be interpreted as human-annotated psychological ground truth.In particular, personality, relationship, emotion, conversational style, and memory-related variables represent computational operationalizations used for dialogue research. They do not constitute clinical, diagnostic, or direct psychological measurements of the speakers.The underlying data are derived from a static dialogue corpus. Consequently, the dataset does not provide evidence of unrestricted longitudinal human–AI coevolution or closed-loop adaptation.Researchers should also distinguish observed dialogue trajectories from counterfactual trajectories produced by generative models.

A generated response may alter subsequent conversational development in a real closed-loop interaction, whereas the released dataset records fixed observed dialogue sequences.ReproducibilityThe released train.csv, valid.csv, and test.csv files correspond to the dataset splits used in the accompanying study.The dataset contains:3,300 dialogues3,300\ \text{dialogues}and26,928 turn-level observations.26,928\ \text{turn-level observations}.The experimental training pipeline contains 14,699 supervised training instances, rather than 14,700, because one training observation contains an empty target response and is excluded from supervised response-generation training.This distinction does not alter the released dataset statistics: train.csv retains all 14,700 turn-level observations.License and AttributionThis resource is derived from the DuLeMon dataset.

Users must comply with the license and terms applicable to the original DuLeMon resource.The original dialogue content remains attributable to the DuLeMon authors and source. The present release introduces additional preprocessing, dataset organization, and automatically constructed social-cognitive annotations.The applicable original license and copyright notice should be retained when redistributing modified versions of the underlying data.

Citation

If you use the underlying DuLeMon dialogue resource, please cite the original DuLeMon paper: @inproceedings{xu2022long, title = {Long Time No See! Open-Domain Conversation with Long-Term Persona Memory}, author = {Xu, Xinchao and Gou, Zhen and Wu, Wenquan and Niu, Zheng-Yu and Wu, Hua and Wang, Haifeng and Wang, Shijin}, booktitle = {Findings of the Association for Computational Linguistics: ACL 2022}, pages = {2639--2650}, year = {2022} } If you use the social-cognitive annotations introduced in this repository, please also cite the accompanying study:[Full citation will be added after publication.]