Language models have shown promise in various tasks but can be affected by undesired data during training, fine-tuning, or alignment. For example, if some unsafe conversations are wrongly annotated as safe ones, the model fine-tuned on these samples may be harmful. Therefore, the correctness of annotations, i.e., the credibility of the dataset, is important. This study focuses on the credibility of real-world datasets, including the popular benchmarks Jigsaw Civil Comments, Anthropic Harmless & Red Team, PKU BeaverTails & SafeRLHF, that can be used for training a harmless language model. Given the cost and difficulty of cleaning these datasets by humans, we introduce a systematic framework for evaluating the credibility of datasets, identifying label errors, and evaluating the influence of noisy labels in the curated language data, specifically focusing on unsafe comments and conversation classification. With the framework, we find and fix an average of 6.16% label errors in 11 datasets constructed from the above benchmarks. The data credibility and downstream learning performance can be remarkably improved by directly fixing label errors, indicating the significance of cleaning existing real-world datasets.
more »
« less
This content will become publicly available on August 11, 2025
UNIWIZ: A Unified Large Language Model Orchestrated Wizard for Safe Knowledge Grounded Conversations
Large Language Models (LLMs) have made significant progress in integrating safety and knowledge alignment. However, adversarial actors can manipulate these models into generating unsafe responses, and excessive safety alignment can lead to unintended hallucinations. To address these challenges, we introduce UniWiz, a novel 2-step data orchestration framework that unifies safety and knowledge data generation. We propose a “safety-priming” method to generate synthetic safety data and overcome safety bottlenecks. We also inject relevant knowledge into conversations by retrieving factual information from curated sources. UniWiz dataset consists of 17,638 quality-controlled conversations and 10,000 augmented preference data. Pretrained models fine-tuned on UniWiz show improvements across various metrics and outperform state-of-the-art instruction-tuned models trained on much larger datasets.
more »
« less
- Award ID(s):
- 2214070
- PAR ID:
- 10543965
- Publisher / Repository:
- Association for Computational Linguistics
- Date Published:
- Page Range / eLocation ID:
- 1749 to 1762
- Format(s):
- Medium: X
- Location:
- Bangkok, Thailand and virtual meeting
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
We collected Instagram data from 150 adolescents (ages 13-21) that included 15,547 private message conversations of which 326 conversations were flagged as sexually risky by participants. Based on this data, we leveraged a human-centered machine learning approach to create sexual risk detection classifiers for youth social media conversations. Our Convolutional Neural Network (CNN) and Random Forest models outperformed in identifying sexual risks at the conversation-level (AUC=0.88), and CNN outperformed at the message-level (AUC=0.85). We also trained classifiers to detect the severity risk level (i.e., safe, low, medium-high) of a given message with CNN outperforming other models (AUC=0.88). A feature analysis yielded deeper insights into patterns found within sexually safe versus unsafe conversations. We found that contextual features (e.g., age, gender, and relationship type) and Linguistic Inquiry and Word Count (LIWC) contributed the most for accurately detecting sexual conversations that made youth feel uncomfortable or unsafe. Our analysis provides insights into the important factors and contextual features that enhance automated detection of sexual risks within youths' private conversations. As such, we make valuable contributions to the computational risk detection and adolescent online safety literature through our human-centered approach of collecting and ground truth coding private social media conversations of youth for the purpose of risk classification.more » « less
-
Yvette Graham, Matthew Purver (Ed.)Understanding the dynamics of counseling conversations is an important task, yet it is a challenging NLP problem regardless of the recent advance of Transformer-based pre-trained language models. This paper proposes a systematic approach to examine the efficacy of domain knowledge and large language models (LLMs) in better representing conversations between a crisis counselor and a help seeker. We empirically show that state-of-the-art language models such as Transformer-based models and GPT models fail to predict the conversation outcome. To provide richer context to conversations, we incorporate human-annotated domain knowledge and LLM-generated features; simple integration of domain knowledge and LLM features improves the model performance by approximately 15%. We argue that both domain knowledge and LLM-generated features can be exploited to better characterize counseling conversations when they are used as an additional context to conversations.more » « less
-
Virtual Mental Health Assistants (VMHAs) are utilized in health care to provide patient services such as counseling and suggestive care. They are not used for patient diagnostic assistance because they cannot adhere to safety constraints and specialized clinical process knowledge ( ProKnow ) used to obtain clinical diagnoses. In this work, we define ProKnow as an ordered set of information that maps to evidence-based guidelines or categories of conceptual understanding to experts in a domain. We also introduce a new dataset of diagnostic conversations guided by safety constraints and ProKnow that healthcare professionals use ( ProKnow - data ). We develop a method for natural language question generation (NLG) that collects diagnostic information from the patient interactively ( ProKnow - algo ). We demonstrate the limitations of using state-of-the-art large-scale language models (LMs) on this dataset. ProKnow - algo incorporates the process knowledge through explicitly modeling safety, knowledge capture, and explainability. As computational metrics for evaluation do not directly translate to clinical settings, we involve expert clinicians in designing evaluation metrics that test four properties: safety, logical coherence, and knowledge capture for explainability while minimizing the standard cross entropy loss to preserve distribution semantics-based similarity to the ground truth. LMs with ProKnow - algo generated 89% safer questions in the depression and anxiety domain (tested property: safety ). Further, without ProKnow - algo generations question did not adhere to clinical process knowledge in ProKnow - data (tested property: knowledge capture ). In comparison, ProKnow - algo -based generations yield a 96% reduction in our metrics to measure knowledge capture. The explainability of the generated question is assessed by computing similarity with concepts in depression and anxiety knowledge bases. Overall, irrespective of the type of LMs, ProKnow - algo achieved an averaged 82% improvement over simple pre-trained LMs on safety, explainability, and process-guided question generation. For reproducibility, we will make ProKnow - data and the code repository of ProKnow - algo publicly available upon acceptance.more » « less
-
Instagram, one of the most popular social media platforms among youth, has recently come under scrutiny for potentially being harmful to the safety and well-being of our younger generations. Automated approaches for risk detection may be one way to help mitigate some of these risks if such algorithms are both accurate and contextual to the types of online harms youth face on social media platforms. However, the imminent switch by Instagram to end-to-end encryption for private conversations will limit the type of data that will be available to the platform to detect and mitigate such risks. In this paper, we investigate which indicators are most helpful in automatically detecting risk in Instagram private conversations, with an eye on high-level metadata, which will still be available in the scenario of end-to-end encryption. Toward this end, we collected Instagram data from 172 youth (ages 13-21) and asked them to identify private message conversations that made them feel uncomfortable or unsafe. Our participants risk-flagged 28,725 conversations that contained 4,181,970 direct messages, including textual posts and images. Based on this rich and multimodal dataset, we tested multiple feature sets (metadata, linguistic cues, and image features) and trained classifiers to detect risky conversations. Overall, we found that the metadata features (e.g., conversation length, a proxy for participant engagement) were the best predictors of risky conversations. However, for distinguishing between risk types, the different linguistic and media cues were the best predictors. Based on our findings, we provide design implications for AI risk detection systems in the presence of end-to-end encryption. More broadly, our work contributes to the literature on adolescent online safety by moving toward more robust solutions for risk detection that directly takes into account the lived risk experiences of youth.more » « less
