Recent advancements in retrieval-augmented generation (RAG) have demonstrated impressive performance in the question-answering (QA) task. However, most previous works predominantly focus on text-based answers. While some studies address multimodal data, they still fall short in generating comprehensive multimodal answers, particularly for explaining concepts or providing step-by-step tutorials on how to accomplish specific goals. This capability is especially valuable for applications such as enterprise chatbots and settings such as customer service and educational systems, where the answers are sourced from multimodal data. In this paper, we introduce a simple and effective framework named MuRAR (Multimodal Retrieval and Answer Refinement). MuRAR enhances text-based answers by retrieving relevant multimodal data and refining the responses to create coherent multimodal answers. This framework can be easily extended to support multimodal answers in enterprise chatbots with minimal modifications. Human evaluation results indicate that multimodal answers generated by MuRAR are more useful and readable compared to plain text answers.

Multiple choice questions (MCQs) are commonly used to evaluate the capabilities of large language models (LLMs). One common way to evaluate the model response is to rank the candidate answers based on the log probability of the first token prediction. An alternative way is to examine the text output. Prior work has shown that first token probabilities lack robustness to changes in MCQ phrasing, and that first token probabilities do not match text answers for instruction-tuned models. Therefore, in this paper, we investigate the robustness of text answers. We show that the text answers are more robust to question perturbations than the first token probabilities, when the first token answers mismatch the text answers. The difference in robustness increases as the mismatch rate becomes greater. As the mismatch reaches over 50\%, the text answer is more robust to option order changes than the debiased first token probabilities using state-of-the-art debiasing methods such as PriDe. Our findings provide further evidence for the benefits of text answer evaluation over first token probability evaluation.

Human-spoken questions are critical to evaluating the performance of spoken question answering (SQA) systems that serve several real-world use cases including digital assistants. We present a new large-scale community-shared SQA dataset, HeySQuAD that consists of 76k human-spoken questions and 97k machine-generated questions and corresponding textual answers derived from the SQuAD QA dataset. The goal of HeySQuAD is to measure the ability of machines to understand noisy spoken questions and answer the questions accurately. To this end, we run extensive benchmarks on the human-spoken and machine-generated questions to quantify the differences in noise from both sources and its subsequent impact on the model and answering accuracy. Importantly, for the task of SQA, where we want to answer human-spoken questions, we observe that training using the transcribed human-spoken and original SQuAD questions leads to significant improvements (12.51%) over training using only the original SQuAD textual questions.

Open-ended questions in surveys are valuable because they do not constrain the respondent's answer, thereby avoiding biases. However, answers to open-ended questions are text data which are harder to analyze. Traditionally, answers were manually classified as specified in the coding manual. Most of the effort to automate coding has gone into the easier problem of single label prediction, where answers are classified into a single code. However, open-ends that require multi-label classification, i.e., that are assigned multiple codes, occur frequently. This paper focuses on multi-label classification of text answers to open-ended survey questions in social science surveys. We evaluate the performance of the transformer-based architecture BERT for the German language in comparison to traditional multi-label algorithms (Binary Relevance, Label Powerset, ECC) in a German social science survey, the GLES Panel (N=17,584, 55 labels). We find that classification with BERT (forcing at least one label) has the smallest 0/1 loss (13.1%) among methods considered (18.9%-21.6%). As expected, it is much easier to correctly predict answer texts that correspond to a single label (7.1% loss) than those that correspond to multiple labels ($\sim$50% loss). Because BERT predicts zero labels for only 1.5% of the answers, forcing at least one label, while recommended, ultimately does not lower the 0/1 loss by much. Our work has important implications for social scientists: 1) We have shown multi-label classification with BERT works in the German language for open-ends. 2) For mildly multi-label classification tasks, the loss now appears small enough to allow for fully automatic classification (as compared to semi-automatic approaches). 3) Multi-label classification with BERT requires only a single model. The leading competitor, ECC, iterates through individual single label predictions.