FLiP: Towards understanding and interpreting multimodal multilingual sentence embeddings

Online hate speech and abusive language pose a growing challenge for content moderation, especially in multilingual settings and for low-resource languages such as Lithuanian. This paper investigates to what extent modern multilingual sentence embedding models can support accurate hate speech detection in Lithuanian, Russian, and English, and how their performance depends on downstream modeling choices and feature dimensionality. We introduce LtHate, a new Lithuanian hate speech corpus derived from news portals and social networks, and benchmark six modern multilingual encoders (potion, gemma, bge, snow, jina, e5) on LtHate, RuToxic, and EnSuperset using a unified Python pipeline. For each embedding, we train both a one class HBOS anomaly detector and a two class CatBoost classifier, with and without principal component analysis (PCA) compression to 64-dimensional feature vectors. Across all datasets, two class supervised models consistently and substantially outperform one class anomaly detection, with the best configurations achieving up to 80.96% accuracy and AUC ROC of 0.887 in Lithuanian (jina), 92.19% accuracy and AUC ROC of 0.978 in Russian (e5), and 77.21% accuracy and AUC ROC of 0.859 in English (e5 with PCA). PCA compression preserves almost all discriminative power in the supervised setting, while showing some negative impact for the unsupervised anomaly detection case. These results demonstrate how modern multilingual sentence embeddings combined with gradient boosted decision trees provide robust soft-computing solutions for multilingual hate speech detection applications.

Multilingual assessment systems commonly rely on translation for scoring and quality-control processes. We evaluate whether multilingual sentence embeddings can replace translated English input for Linguistic-Integrated Reliability Auditing (LiRA) across 11 PIRLS constructed-response items and three embedding models. Native-language embeddings reproduced translation-based reliability estimates closely while recovering responses excluded after translation failure, with no meaningful change in reliability.

Recent multilingual vision--language encoders cover hundreds of languages in a single model, yet on two state-of-the-art instances retrieval on low-resource languages (LRL; e.g. Swahili) trails high-resource ones (HRL; e.g. English) by $30^+$,pp. We ask where in the trained encoder this gap is located. Prior modality-gap and cross-lingual subspace work suggests a linear language direction at the output crowds out alignment-relevant geometry. We falsify this: LEACE drives the linear language classifier from $>99\%$ to near chance and iterated INLP to $37$--$50\%$ while LRL retrieval moves within $\pm 1.5$,pp and all tier means within $2.2$,pp, tracking random controls. The linear bias is a \emph{symptom}, not the cause. Instead, the alignment-causal factor lies along the encoder's forward path: the EOS (end-of-sequence) hidden state's per-language trajectory diverges with depth. Substituting the EOS with its parallel English value three blocks before the projector lifts Swahili from $22.1\%$ to $69.1\%$ on one encoder (and reproduces on the other); three controls rule out pooled-position tautology and English specificity. A front-layer trunk that pulls each language's projection toward the parallel-content centroid corroborates the diagnosis at training time, recovering $+9.6$ / $+17.1$,pp on LRL XM3600 retrieval (1{,}000-image subset), with consistent gains across three further benchmarks while preserving HRL performance.