USDC: A Dataset of $\underline{U}$ser $\underline{S}$tance and $\underline{D}$ogmatism in Long $\underline{C}$onversations
Keywords: large language models, annotators, user opinions, stance, dogmatism, human-llm alignment, open-source llms, closed-source llms
Abstract: Although prior studies have explored Stance and Dogmatism in user conversations, their datasets are constructed at the post level, treating each post as independent and randomly sampling posts from conversation threads. Thus, Stance and Dogmatism labels in these datasets cannot capture the user's opinion fluctuations expressed throughout the entire conversation context. However, identifying user's opinion fluctuations in long conversation threads on various topics can be extremely critical for enhanced personalization, market research, political campaigns, customer service, conflict resolution, targeted advertising, and content moderation. Hence, training language models to automate this task is critical. However, to train such models, gathering manual annotations has multiple challenges: 1) It is time-consuming and costly; 2) Conversation threads could be very long, increasing chances of noisy annotations; and 3) Interpreting instances where a user changes their opinion within a conversation is difficult because often such transitions are subtle and not expressed explicitly. Inspired by the recent success of large language models (LLMs) for complex natural language processing tasks, we leverage Mistral Large and GPT-4 to automate the human annotation process on the following two tasks while also providing reasoning: i) User Stance classification, which involves labeling a user's stance of a post in a conversation on a five-point scale; ii) User Dogmatism classification, which deals with labeling a user's overall opinion in the conversation on a four-point scale. The majority voting on zero-shot, one-shot, and few-shot annotations from these two LLMs on 764 multi-user Reddit conversations helps us curate the USDC dataset. USDC is then used to finetune and instruction-tune multiple deployable small language models for the 5-class stance and 4-class dogmatism classification tasks. Additionally, human annotations on 200 test conversations achieved inter-annotator agreement scores of 0.49 for stance and 0.50 for dogmatism, indicating a reasonable level of consistency between human and LLM annotations. We make the code and dataset publicly available [https://anonymous.4open.science/r/USDC-0F7F].
Primary Area: datasets and benchmarks
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 4859
Loading