Enhanced DNMF-graph integrated multilevel semantic factorization for contextual named entity recognition

Named Entity Recognition (NER) has undergone considerable changes in the recent past driven by the growing demand for models that are more robust, context aware and capable of generalizing across diverse textual domain. However, existing state of the art approaches primarily rely on local contextual representations, and often fails to capture corpus level semantic structures, particularly in noisy or low resource settings. To address these limitations, a unified NER system is proposed in the present work, that utilizes a Deep Non-Negative Matrix Factorization (DNMF) integrated with token level graph community discovery and a Transformer- Conditional Random Fields (CRF) model. The proposed model, named as DNMF-Graph-Transformer CRF, can utilize the strengths of various representation levels. By applying DNMF to the token-context matrix, several interpretable, low-rank latent semantic factors are obtained that not only complement neural representations but also help to improve the model's generalization when it has to work with a small amount of noisy data. Next, a token interaction graph is generated to represent the global relationships among tokens. Thus, community embeddings can be extracted that give structural priors at the corpus level. These features represent semantic neighbourhoods and functional groupings of tokens. Finally, the Transformer- CRF is adapted to learn the contextual aspects of the text. The performance of the proposed model is evaluated by conducting comprehensive experiments on CoNLL-2003, OntoNotes 5.0, and WNUT 2017. The results demonstrate that the proposed model significantly outperforms existing baseline models. It is observed that the performance gains are especially significant for noisy social media text.