The relentless pace of video production exacerbates the digital accessibility gap that individuals who are blind or low vision (BLV) face on a daily basis, resulting in disproportionate exclusion from community opportunities and risk management. Whereas previous automated audio description (AD) systems provide single-tool approaches for delivering minimum viable description (MVD) or delivering on-demand visual question answering (VQA), we present a tandem AI-based AD tool that combines MVD and on-demand VQA. A user study with 26 BLV individuals explored how the tandem system may be used under the conditions of delivering MVD and/or on-demand VQA with AI-only or human-in-the-loop support. When each tool was used in isolation, AI-only conditions scored significantly lower in both user enjoyment and comprehension. When used in tandem, AI-only conditions matched outcomes delivered with human-in-the-loop, which suggests that AI-only AD tools may be most effective when both types of tools are used in tandem. A multimodal analysis of interactions with the tandem system revealed areas for system improvement in terms of the timing of AD delivery and accurate content delivery. We discuss how the use of both types of tools in a tandem system can mitigate some of the digital frictions that have plagued efforts in machine learning and automated tools for accessibility.

We present a new high-quality validated image caption rating (VICR) dataset. How well a caption fits an image can be difficult to assess due to the subjective nature of caption quality. How do we evaluate whether a caption is good? We generated a new dataset to help answer this question by using our new image caption rating system, which consists of a novel robust rating scale and gamified approach to gathering human ratings. We show that our approach is consistent and teachable. 113 participants were involved in generating the dataset, which is composed of 68,217 ratings among 15,646 image-caption pairs. Our new dataset has greater inter-rater agreement than the state of the art, and custom machine learning rating predictors that were trained on our dataset outperform previous metrics. We improve over Flickr8k-Expert in Kendall's $W$ by 12\% and in Fleiss' $\kappa$ by 19\%, and thus provide a new benchmark dataset for image caption rating.

Given a graph G=(V,E) with V={1,...,n}, we place on every vertex a token T_1,...,T_n. A swap is an exchange of tokens on adjacent vertices. We consider the algorithmic question of finding a shortest sequence of swaps such that token T_i is on vertex i. We are able to achieve essentially matching upper and lower bounds, for exact algorithms and approximation algorithms. For exact algorithms, we rule out any 2^{o(n)} algorithm under the ETH. This is matched with a simple 2^{O(n*log(n))} algorithm based on a breadth-first search in an auxiliary graph. We show one general 4-approximation and show APX-hardness. Thus, there is a small constant delta > 1 such that every polynomial time approximation algorithm has approximation factor at least delta. Our results also hold for a generalized version, where tokens and vertices are colored. In this generalized version each token must go to a vertex with the same color.