Amharic neural coreference resolution with multi-head attention and named entity recognition

Coreference resolution is a fundamental task in natural language processing (NLP) that identifies all textual expressions referring to the same real-world entity. While crucial for downstream NLP applications, this task presents particular challenges for Amharic, a morphologically rich Semitic language with limited annotated resources compared to widely studied languages. This study presents a neural coreference resolution system for Amharic, which integrates multi-head attention (MHA) and a named entity recognition (NER) module. MHA is utilized to capture intricate contextual relationships within and between mentions, enhancing the model's ability to discern coreferent links. The NER module integrates a bidirectional long short-term memory (BiLSTM) network with a conditional random field, enhancing mention detection by leveraging semantic information from named entities to guide candidate span generation. In addition, we proposed a comprehensive model architecture comprising preprocessing and morphological analysis steps, followed by contextualization, span generation and representation, mention scoring and pruning, and a dedicated coreference resolution layer. Experimental results on a custom-built Amharic coreference dataset demonstrate the effectiveness of our approach, showcasing competitive performance and highlighting the synergistic benefits of combining multi-head attention with explicit NER for improving coreference resolution in a low-resource setting.