From Static to Contextual: A Survey of Embedding Advances in NLP
Hussein Alkaabi, Ali Najm Jasim, Ali Darroudi
Abstract
Embedding techniques have been a cornerstone of Natural Language Processing (NLP), enabling machines to represent textual data in a form that captures semantic and syntactic relationships. Over the years, the field has witnessed a significant evolution—from static word embeddings, such as Word2Vec and GloVe, which represent words as fixed vectors, to dynamic, contextualized embeddings like BERT and GPT, which generate word representations based on their surrounding context. This survey provides a comprehensive overview of embedding techniques, tracing their development from early methods to state-of-the-art approaches. We discuss the strengths and limitations of each paradigm, their applications across various NLP tasks, and the challenges they address, such as polysemy and out-of-vocabulary words. Furthermore, we highlight emerging trends, including multimodal embeddings, domain-specific representations, and efforts to mitigate embedding bias. By synthesizing the advancements in this rapidly evolving field, this paper aims to serve as a valuable resource for researchers and practitioners while identifying open challenges and future directions for embedding research in NLP.