How hard is it to set the heating to 19 degrees or turn on the light with your voice? Harder than it sounds: spoken home-automation commands must be transcribed and mapped to an unambiguous, machine-executable representation, but recognition errors and off-schema parsing often break this chain and open speech-to-action benchmarks remain scarce. We introduce the Home Automation Speech Ensemble (HASE), an open dataset of 700 spoken commands with gold transcripts and gold sequences, and a fully open two-stage pipeline that combines local automatic speech recognition with an instruction-tuned language model to produce a single schema-valid JSON command. We benchmark seven ASR back ends and ten instruction-tuned LLMs (up to 8B parameters) under few-shot prompting, and compare the best open stack (Whisper large-v2 + Qwen3-4B-Instruct-2507) against a representative commercial baseline (GPT-4o-transcribe + GPT-4o-mini) using deployment-oriented metrics including EEM and SMR. On HASE, the best open stack reaches 0.778 EEM and 0.888 SMR, compared to 0.818 EEM and 0.910 SMR for the commercial baseline.
This demo presents ExPeerience, an AI-assisted learnersourcing system that supports students in co-creating contextualized worked-out examples for database programming. Learnersourcing is a pedagogical approach in which students act as contributors by creating learning materials as part of the learning process. The system addresses two challenges in learnersourcing: (1) the high cognitive burden students face during content creation, and (2) the difficulty of balancing student learning with the production of reusable, high-quality sourced learning materials. ExPeerience structures student-AI collaboration by assigning the AI distinct roles across ideation, content generation, and solution evaluation, while keeping students responsible for decision-making and problem-solving.
Numeric ratings compress heterogeneous emotional experiences into a single scalar, obscuring why users assign similar scores for different reasons. We argue that this limitation stems from how emotions are represented rather than from the absence of emotion modeling. We introduce a polarity-aware decomposition of anticipation and show that treating anticipation as a single undifferentiated emotion produces incoherent and misleading relationships with ratings. Building on this insight, we model emotion strength as a continuous explanatory representation rather than predicting emotion occurrence probabilities. To make these interpretable representations usable with text encoders, we train a neural regressor to approximate lexicon-derived emotion intensity vectors while preserving interpretability. Experiments on two large review datasets demonstrate that the proposed representation improves rating prediction accuracy and yields coherent emotion to rating relationships that are invisible under probability-based emotion models. These findings highlight representational design as a key factor in explainable recommendation systems.
Frequent switching between keyboard and mouse disrupts work-flow, slows down text editing, and increases physical fatigue for experienced typists. This repetitive context switching also contributes to repetitive strain injuries. In this paper, we introduce WASDHead, a system that combines a head mouse with keyboard input to reposition the cursor and perform actions, eliminating keyboard-mouse context switching. WASDHead enables users to select text editing positions, perform copy-paste actions, and interact with common functions without lifting their hands from the home row. Our technique uses head movement with simple and memorable keys (W, A, S, D for navigation and J, K, L for mouse clicking) and offers a low learning curve. It also requires no specialized hardware or additional sensors by relying solely on a standard RGB camera.
Intelligent interfaces increasingly rely on large language models (LLMs) to summarize user-generated content. While efective at condensing what is explicitly mentioned, known as presence bias, such summaries systematically overlook what is expected yet missing. This presence bias can mislead users by obscuring diagnostic gaps that matter for decision making. We present a demonstration of DiSCo (Domain-informed Summarization through Contrast), a method for review summarization that explicitly surfaces both salient and absent information. The system constructs domain-level topical expectations from comparable entities and contrasts them with individual review profles to identify over-represented and missing aspects. These contrasts are then integrated into summaries using LLMs, allowing users to explore what is present and stands out and what is missing compared to other entities in the domain. The demo showcases how DiSCo summaries change user interaction with review content across multiple accommodation domains. Participants can interactively toggle expectation views, inspect missing aspects, and compare standard LLM summaries with DiSCo summaries in real time. The demonstration illustrates how expectation-aware interfaces support transparency, refection, and better-informed decision making.
