Multilingual lexicons improve sentiment control for low-resource languages
Steering with Multilingual Lexicons for Sentiment Control in Low-Resource Languages
Computation and Language
Summary
Many languages lack large datasets needed for teaching computers to recognize sentiment in sentences. The authors found that using multilingual sentiment word lists can guide AI models to better understand sentiment in these languages. They also discovered that these word lists work better than English-only ones, especially for languages not heavily used in training. However, some technical challenges can limit how well this works, like when English signals dominate others. Their work explains why and how these effects happen and shows that word lists are a practical way to steer AI sentiment understanding across languages.
What this means in practice
- •For natural language processing engineers: Enhance sentiment analysis models for languages with limited training data by using aggregated multilingual word sentiment lists as control signals.
- •For machine translation developers: Improve sentiment preservation in translation systems by integrating multilingual sentiment lexicons to better steer model outputs for low-resource languages.
Authors
Lucinda Laurent, Muhammad Dehan Al Kautsar, Alfan Farizki Wicaksono, Fajri Koto
Abstract
Activation steering offers a lightweight mechanism for controlling model behavior, but existing methods depend on sentence-level contrastive data that is unavailable for most of the world's languages. We show that multilingual sentiment lexicons are a viable substitute across three model families: aggregated multilingual lexicons outperform English-derived steering vectors for mid- and low-resource languages, while monolingual lexicons remain stronger for languages a model has specialized in during pre-training. The steering efficacy is mechanistically dictated by both representational density and pre-training priors. Large-magnitude English vectors overwhelm the compact representations of low-resource languages, causing "linguistic collapse." This fragility is mitigated by the norm-dampening of multilingual aggregation. Architectural biases, including bilingual specialization penalty and latent numerical rigidity, also heavily constrain cross-lingual steering responsiveness. Ultimately, our work establishes lexicons as a practical steering signal and offers a comprehensive mechanistic account of cross-lingual steering dynamics.