This work investigates whether modern speech models are sensitive to prosodic emphasis - whether they encode emphasized and neutral words in systematically different ways. Prior work typically relies on isolated acoustic correlates (e.g., pitch, duration) or label prediction, both of which miss the relational structure of emphasis. This paper proposes a residual-based framework, defining emphasis as the difference between paired neutral and emphasized word representations. Analysis on self-supervised speech models shows that these residuals correlate strongly with duration changes and perform poorly at word identity prediction, indicating a structured, relational encoding of prosodic emphasis. In ASR fine-tuned models, residuals occupy a subspace up to 50% more compact than in pre-trained models, further suggesting that emphasis is encoded as a consistent, low-dimensional transformation that becomes more structured with task-specific learning.

Abstract meaning representation (AMR) is a semantic formalism used to represent the meaning of sentences as directed acyclic graphs. In this paper, we describe how real digital dictionaries can be embedded into AMR directed graphs (digraphs), using state-of-the-art pre-trained large language models. Then, we reduce those graphs in a confluent manner, i.e. with transformations that preserve their circuit space. Finally, the properties of these reduces digraphs are analyzed and discussed in relation to the symbol grounding problem.

Most pre-trained Vision-Language (VL) models and training data for the downstream tasks are only available in English. Therefore, multilingual VL tasks are solved using cross-lingual transfer: fine-tune a multilingual pre-trained model or transfer the text encoder using parallel data. We study the alternative approach: transferring an already trained encoder using parallel data. We investigate the effect of parallel data: domain and the number of languages, which were out of focus in previous work. Our results show that even machine-translated task data are the best on average, caption-like authentic parallel data outperformed it in some languages. Further, we show that most languages benefit from multilingual training.

We thank all the reviewers for their time and insightful feedback about our work. Many of the recent few-shot learning works focus on computer vision compared to NLU tasks. We leverage self-training with several advances to bridge this gap. Similar baselines reported for active learning [Gal et al., 2017] and preference learning [Houlsby et al., UDA [Xie et al., 2019] and self-training with noisy student [Xie et al., 2020] show these techniques to work best with Additionally, for IMDB longer sequence length plays a big role. Sample mixing based on easy and hard examples is an interesting idea.