In the past decade a surge in the amount of electronic health record (EHR) data in the United States, attributed to a favorable policy environment created by the Health Information Technology for Economic and Clinical Health (HITECH) Act of 2009 and the 21st Century Cures Act of 2016. Clinical notes for patients' assessments, diagnoses, and treatments are captured in these EHRs in free-form text by physicians, who spend a considerable amount of time entering and editing them. Manually writing clinical notes takes a considerable amount of a doctor's valuable time, increasing the patient's waiting time and possibly delaying diagnoses. Large language models (LLMs) possess the ability to generate news articles that closely resemble human-written ones. We investigate the usage of Chain-of-Thought (CoT) prompt engineering to improve the LLM's response in clinical note generation. In our prompts, we use as input International Classification of Diseases (ICD) codes and basic patient information. We investigate a strategy that combines the traditional CoT with semantic search results to improve the quality of generated clinical notes. Additionally, we infuse a knowledge graph (KG) built from clinical ontology to further enrich the domain-specific knowledge of generated clinical notes. We test our prompting technique on six clinical cases from the CodiEsp test dataset using GPT-4 and our results show that it outperformed the clinical notes generated by standard one-shot prompts.

One way to mitigate risks in vision-language models (VLMs) is to remove dangerous samples in their training data. However, such data moderation can be easily bypassed when harmful images are split into small, benign-looking patches, scattered across many training samples. VLMs may then learn to piece these fragments together during training and generate harmful responses at inference, either from full images or text references. For instance, if trained on image patches from a bloody scene paired with the descriptions "safe," VLMs may later describe, the full image or a text reference to the scene, as "safe." We define the core ability of VLMs enabling this attack as $\textit{visual stitching}$ -- the ability to integrate visual information spread across multiple training samples that share the same textual descriptions. In our work, we first demonstrate visual stitching abilities in common open-source VLMs on three datasets where each image is labeled with a unique synthetic ID: we split each $(\texttt{image}, \texttt{ID})$ pair into $\{(\texttt{patch}, \texttt{ID})\}$ pairs at different granularity for finetuning, and we find that tuned models can verbalize the correct IDs from full images or text reference. Building on this, we simulate the adversarial data poisoning scenario mentioned above by using patches from dangerous images and replacing IDs with text descriptions like ``safe'' or ``unsafe'', demonstrating how harmful content can evade moderation in patches and later be reconstructed through visual stitching, posing serious VLM safety risks. Code is available at https://github.com/ZHZisZZ/visual-stitching.

High-quality data labeling from specific domains is costly and human time-consuming. In this work, we propose a self-supervised domain adaptation method, based upon an iterative pseudo-forced alignment algorithm. The produced alignments are employed to customize an end-to-end Automatic Speech Recognition (ASR) and iteratively refined. The algorithm is fed with frame-wise character posteriors produced by a seed ASR, trained with out-of-domain data, and optimized throughout a Connectionist Temporal Classification (CTC) loss. The alignments are computed iteratively upon a corpus of broadcast TV. The process is repeated by reducing the quantity of text to be aligned or expanding the alignment window until finding the best possible audio-text alignment. The starting timestamps, or temporal anchors, are produced uniquely based on the confidence score of the last aligned utterance. This score is computed with the paths of the CTC-alignment matrix. With this methodology, no human-revised text references are required. Alignments from long audio files with low-quality transcriptions, like TV captions, are filtered out by confidence score and ready for further ASR adaptation. The obtained results, on both the Spanish RTVE2022 and CommonVoice databases, underpin the feasibility of using CTC-based systems to perform: highly accurate audio-text alignments, domain adaptation and semi-supervised training of end-to-end ASR.

Reference resolution, which aims to identify entities being referred to by a speaker, is more complex in real world settings: new referents may be created by processes the agents engage in and/or be salient only because they belong to the shared physical setting. Our focus is on resolving references to visualizations on a large screen display in multimodal dialogue; crucially, reference resolution is directly involved in the process of creating new visualizations. We describe our annotations for user references to visualizations appearing on a large screen via language and hand gesture and also new entity establishment, which results from executing the user request to create a new visualization. We also describe our reference resolution pipeline which relies on an information-state architecture to maintain dialogue context. We report results on detecting and resolving references, effectiveness of contextual information on the model, and under-specified requests for creating visualizations. We also experiment with conventional CRF and deep learning / transformer models (BiLSTM-CRF and BERT-CRF) for tagging references in user utterance text. Our results show that transfer learning significantly boost performance of the deep learning methods, although CRF still out-performs them, suggesting that conventional methods may generalize better for low resource data.

Prompting inputs with natural language task descriptions has emerged as a popular mechanism to elicit reasonably accurate outputs from large-scale generative language models with little to no in-context supervision. This also helps gain insight into how well language models capture the semantics of a wide range of downstream tasks purely from self-supervised pre-training on massive corpora of unlabeled text. Such models have naturally also been exposed to a lot of undesirable content like racist and sexist language and there is limited work on awareness of models along these dimensions. In this paper, we define and comprehensively evaluate how well such language models capture the semantics of four tasks for bias: diagnosis, identification, extraction and rephrasing. We define three broad classes of task descriptions for these tasks: statement, question, and completion, with numerous lexical variants within each class. We study the efficacy of prompting for each task using these classes and the null task description across several decoding methods and few-shot examples. Our analyses indicate that language models are capable of performing these tasks to widely varying degrees across different bias dimensions, such as gender and political affiliation. We believe our work is an important step towards unbiased language models by quantifying the limits of current self-supervision objectives at accomplishing such sociologically challenging tasks.

A major challenge in visually grounded language generation is to build robust benchmark datasets and models that can generalize well in real-world settings. To do this, it is critical to ensure that our evaluation protocols are correct, and benchmarks are reliable. In this work, we set forth to design a set of experiments to understand an important but often ignored problem in visually grounded language generation: given that humans have different utilities and visual attention, how will the sample variance in multi-reference datasets affect the models' performance? Empirically, we study several multi-reference datasets and corresponding vision-and-language tasks. We show that it is of paramount importance to report variance in experiments; that human-generated references could vary drastically in different datasets/tasks, revealing the nature of each task; that metric-wise, CIDEr has shown systematically larger variances than others. Our evaluations on reference-per-instance shed light on the design of reliable datasets in the future.