Search for: All records

Despite their extensive application in language understanding tasks, large language models (LLMs) still encounter challenges including hallucinations - occasional fabrication of information - and alignment issues - lack of associations with human-curated world models (e.g., intuitive physics or common-sense knowledge). Moreover, the black-box nature of LLMs presents significant obstacles in training them effectively to achieve desired behaviors. In particular, modifying the concept embedding spaces of LLMs can be highly intractable. This process involves analyzing the implicit impact of such adjustments on the myriad parameters within LLMs and the resulting inductive biases. We propose a novel architecture that wraps powerful function approximation architectures within an outer, interpretable read-out layer. This read-out layer can be scrutinized to explicitly observe the effects of concept modeling during the training of the LLM. Our method stands in contrast with gradient-based implicit mechanisms, which depend solely on adjustments to the LLM parameters and thus evade scrutiny. By conducting extensive experiments across both generative and discriminative language modeling tasks, we evaluate the capabilities of our proposed architecture relative to state-of-the-art LLMs of similar sizes. Additionally, we offer a qualitative examination of the interpretable read-out layer and visualize the concepts it captures. The results demonstrate the potential of our approach for effectively controlling LLM hallucinations and enhancing the alignment with human expectations. 

Large Language Models have excelled at encoding and leveraging language patterns in large text-based corpora for various tasks, including spatiotemporal event-based question answering (QA). However, due to encoding a text-based projection of the world, they have also been shown to lack a full bodied understanding of such events, e.g., a sense of intuitive physics, and cause-and-effect relationships among events. In this work, we propose using causal event graphs (CEGs) to enhance language understanding of spatiotemporal events in language models, using a novel approach that also provides proofs for the model’s capture of the CEGs. A CEG consists of events denoted by nodes, and edges that denote cause and effect relationships among the events. We perform experimentation and evaluation of our approach for benchmark spatiotemporal QA tasks and show effective performance, both quantitative and qualitative, over state-of-the-art baseline methods. 

Predicting anomalies in manufacturing assembly lines is crucial for reducing time and labor costs and improving processes. For instance, in rocket assembly, premature part failures can lead to significant financial losses and labor inefficiencies. With the abundance of sensor data in the Industry 4.0 era, machine learning (ML) offers potential for early anomaly detection. However, current ML methods for anomaly prediction have limitations, with F1 measure scores of only 50% and 66% for prediction and detection, respectively. This is due to challenges like the rarity of anomalous events, scarcity of high-fidelity simulation data (actual data are expensive), and the complex relationships between anomalies not easily captured using traditional ML approaches. Specifically, these challenges relate to two dimensions of anomaly prediction: predicting when anomalies will occur and understanding the dependencies between them. This paper introduces a new method called Robust and Interpretable 2D Anomaly Prediction (RI2AP) designed to address both dimensions effectively. RI2AP is demonstrated on a rocket assembly simulation, showing up to a 30-point improvement in F1 measure compared to current ML methods. This highlights its potential to enhance automated anomaly prediction in manufacturing. Additionally, RI2AP includes a novel interpretation mechanism inspired by a causal-influence framework, providing domain experts with valuable insights into sensor readings and their impact on predictions. Finally, the RI2AP model was deployed in a real manufacturing setting for assembling rocket parts. Results and insights from this deployment demonstrate the promise of RI2AP for anomaly prediction in manufacturing assembly pipelines. 

Language models have the potential to assess mental health using social media data. By analyzing online posts and conversations, these models can detect patterns indicating mental health conditions like depression, anxiety, or suicidal thoughts. They examine keywords, language markers, and sentiment to gain insights into an individual’s mental well-being. This information is crucial for early detection, intervention, and support, improving mental health care and prevention strategies. However, using language models for mental health assessments from social media has two limitations: (1) They do not compare posts against clinicians’ diagnostic processes, and (2) It’s challenging to explain language model outputs using concepts that the clinician can understand, i.e., clinician-friendly explanations. In this study, we introduce Process Knowledge-infused Learning (PK-iL), a new learning paradigm that layers clinical process knowledge structures on language model outputs, enabling clinician-friendly explanations of the underlying language model predictions. We rigorously test our methods on existing benchmark datasets, augmented with such clinical process knowledge, and release a new dataset for assessing suicidality. PKiL performs competitively, achieving a 70% agreement with users, while other XAI methods only achieve 47% agreement (average inter-rater agreement of 0.72). Our evaluations demonstrate that PK-iL effectively explains model predictions to clinicians. 

Conversational Agents (CAs) powered with deep language models (DLMs) have shown tremendous promise in the domain of mental health. Prominently, the CAs have been used to provide informational or therapeutic services (e.g., cognitive behavioral therapy) to patients. However, the utility of CAs to assist in mental health triaging has not been explored in the existing work as it requires a controlled generation of follow-up questions (FQs), which are often initiated and guided by the mental health professionals (MHPs) in clinical settings. In the context of ‘depression’, our experiments show that DLMs coupled with process knowledge in a mental health questionnaire generate 12.54% and 9.37% better FQs based on similarity and longest common subsequence matches to questions in the PHQ-9 dataset respectively, when compared with DLMs without process knowledge support.Despite coupling with process knowledge, we find that DLMs are still prone to hallucination, i.e., generating redundant, irrelevant, and unsafe FQs. We demonstrate the challenge of using existing datasets to train a DLM for generating FQs that adhere to clinical process knowledge. To address this limitation, we prepared an extended PHQ-9 based dataset, PRIMATE, in collaboration with MHPs. PRIMATE contains annotations regarding whether a particular question in the PHQ-9 dataset has already been answered in the user’s initial description of the mental health condition. We used PRIMATE to train a DLM in a supervised setting to identify which of the PHQ-9 questions can be answered directly from the user’s post and which ones would require more information from the user. Using performance analysis based on MCC scores, we show that PRIMATE is appropriate for identifying questions in PHQ-9 that could guide generative DLMs towards controlled FQ generation (with minimal hallucination) suitable for aiding triaging. The dataset created as a part of this research can be obtained from https://github.com/primate-mh/Primate2022