InferNER: an attentive model leveraging the sentence-level information for Named Entity Recognition in Microblogs

Abstract

We investigate the problem of named entity recognition in the user-generated text such as social media posts. This task is rendered particularly difficult by the restricted length and limited grammatical coherence of this data type. Current state-of-the-art approaches rely on external sources such as gazetteers to alleviate some of these restrictions. We present a neural model able to outperform state of the art on this task without recurring to gazetteers or similar external sources of information. Our approach relies on word-, character-, and sentence-level information for NER in short-text. Social media posts like tweets often have associated images that may provide auxiliary context relevant to understand these texts. Hence, we also incorporate visual information and introduce an attention component which computes attention weight probabilities over textual and text-relevant visual contexts separately. Our model outperforms the current state of the art on various NER datasets. On WNUT 2016 and 2017, our model achieved 53.48% and 50.52% F1 score, respectively. With Multimodal model, our system also outperforms the current SOTA with an F1 score of 74% on the multimodal dataset.
Our evaluation further suggests that our model also goes beyond the current state-of-the-art on newswire data, hence corroborating its suitability for various NER tasks.