The Role of Global and Local Context in Named Entity Recognition

TL;DR

It is found that correctly retrieving global document context has a greater impact on performance than only leveraging local context, prompting for further research on how to better retrieve that context.

Abstract

Pre-trained transformer-based models have recently shown great performance when applied to Named Entity Recognition (NER). As the complexity of their self-attention mechanism prevents them from processing long documents at once, these models are usually applied in a sequential fashion. Such an approach unfortunately only incorporates local context and prevents leveraging global document context in long documents such as novels, which might hinder performance. In this article, we explore the impact of global document context, and its relationships with local context. We find that correctly retrieving global document context has a greater impact on performance than only leveraging local context, prompting for further research on how to better retrieve that context.

Figures (9)

• Figure 1: Mean F1 score versus max number of retrieved sentences for all retrieval heuristics across 3 runs.
• Figure 2: Mean F1 score versus max number of retrieved sentences across 3 runs for oracle versions of all retrieval heuristics.
• Figure 3: Distribution of the distance of retrieved sentences using the oracle versions of the samenoun and bm25 heuristics. The samenoun heuristic retrieves fewer sentences overall, since it is possible for some sentence to not have a common noun with any other sentence of its document.
• Figure 4: Mean F1 score versus number of retrieved sentences across 3 runs for the oracle version of the bm25 heuristic, and the same heuristic restricted to distant context.
• Figure 5: Distribution of the number of sentences in our enhanced version of the dataset from dekker-2019-evaluation_ner_social_networks_novels.