Sentence embeddings represent meaning as a single fixed-size vector, regardless of the complexity of the input sentence. This design compresses all propositional structure into a single point, losing part of the internal structure of the sentence in the process. To address this, we introduce Atomic Proposition Embeddings (APE) to dynamically discover and encode atomic propositions in sentences as variable-size bags of vectors. We use a slot mechanism, inspired by object detection in computer vision, over an encoder backbone. APE is trained on a proposition-decomposed version of WikiSplit, and matches or outperforms sentence encoders up to 15x its size on PAWS-Wiki and SNLI with no task-specific fine-tuning, using proposition-level scoring rules at inference time. We further introduce PropSearch, a retrieval benchmark for proposition-level embeddings in which models must retrieve source sentences given atomic propositions. APE leads on PropSearch over proposition-level alternatives, with its advantage growing with sentence complexity. Ablations confirm the gains are architectural rather than data-driven and that APE can generalize over its training signal. We release our code, data and models for research purposes at https://github.com/SapienzaNLP/APE.

The uncontrolled diffusion of hate speech on social media requires robust detection mechanisms to measure its harmful impact. Analyzing texts from X (formerly Twitter) is challenging due to slang, neologisms, and sarcasm, which require advanced and intelligent detection approaches. While sophisticated models like large language models (LLMs) demonstrate impressive accuracy, their prohibitive inference times make it impractical to process millions of tweets. Therefore, we propose a mixed approach using a bidirectional long short-term memory model with an added attention mechanism (AT-BiLSTM) for improved natural language understanding. We benchmark this model against a standard BiLSTM model and a fine-tuned multilingual robustly optimized BERT (RoBERTa).The task of hate speech detection has been extensively explored in the EVALITA campaigns, which have achieved impressive results. Building on this foundation, we aim to develop a robust classifier to predict the content of approximately 20 million tweets related to immigration. The performance of our models is comparable to the top entries from the EVALITA campaigns, and we show the effects of training different networks on the dynamics of the Hate Speech Index (HSI). We also utilize a custom labeled dataset for benchmarking and training.