We describe our system for predicting sense plausibility in short narratives. Our approach centers on task decomposition: instead of predicting a score directly, we break the problem into simpler subtasks and combine their outputs. We further improve performance by ensembling complementary signals, including word sense disambiguation and fine-tuned embedding models. We also find empirical support for the one-homonym-per-translation principle of Hauer and Kondrak (2020a). Our best ensemble system achieves competitive performance in the official evaluation. Our code and data are available on GitHub.

Large language models (LLMs) continue to achieve impressive performance on reasoning benchmarks, yet it remains unclear how their predictions capture semantic consistency between sentences. We investigate the important open question of whether word-level surprisal correlates with sentence-level contradiction between a premise and a hypothesis. Specifically, we compute surprisal for hypothesis words across a diverse set of experimental variants, and analyze its association with contradiction labels over multiple datasets and open-source LLMs. Because modern LLMs operate on subword tokens and can not directly produce reliable surprisal estimates, we introduce a token-to-word decoding algorithm that extends theoretically grounded probability estimation to open-vocabulary settings. Experiments show a consistent and statistically significant positive correlation between surprisal and contradiction across models and domains. Our analysis also provides new insights into the capabilities and limitations of current LLMs. Together, our findings suggest that surprisal can localize sentence-level inconsistency at the word level, establishing a quantitative link between lexical uncertainty and sentential semantics. We plan to release our code and data upon publication.

We describe the methods used by our UAlberta team for the SemEval-2025 Task 2 on Entity-Aware Machine Translation (EA-MT). Our methods leverage large language models with prompt engineering strategies suited to this task, including retrieval augmented generation and in-context learning. Our best results overall are obtained with ensembles of multiple models, leveraging named entity knowledge in the dataset. Finally, we provide proof-of-concept experiments showing that lexico-semantic knowledge can be used to identify high-quality translations. We further demonstrate that our methods can function even without gold named entity translations, by using an alternative knowledge base such as BabelNet.