Improving entity linking by modeling latent relations between mentions

Proposed multi-relational entity linking model that encodes relations as latent variables and induce the relations without any supervision (or little supervision) and simultaneously optimise the entity-linking system in an end-to-end manner. This means that relations between mentions are induced in a way that benefits entity linking! This is different from previous work that relied on supervised systems or heuristics approach to predict relations.

Future Work

  1. Use syntactic and discourse structures to enable the model to discover a richer set of relations!

  2. Combine ment-norm and rel-norm!

  3. Examine whether induced latent relations can be helpful for relation extraction

Deep joint entity disambiguation with local neural attention

Proposed a deep learning model for joint document-level entity disambiguation, which has three main components: entity embeddings, local context attention, and collective disambiguation. Entity disambiguation usually focuses on two types of contextual information: local information based on words within context window and global information, exploiting document-level coherence of entities. This work, along with many SOTA, focuses on combining both sets of information. The additional of this paper is that it uses embeddings to represent entities to assess local and global evidence! Many systems still uses manually designed features!

Future Work

  1. Extend this system to perform nil detection, coreference resolution, and mention detection

A unified MRC framework for named entity recognition

Proposed a unified framework for tackling both nested and flat NER problem by treating the task as machine reading comprehension (MRC) problem. This means that we are formulating NER problem in a SQUAD style format where each entity type is characterised by a natural language query and entities are extracted by answering these queries given the contexts. This strategy can handle both flat and overlapping entities where we can extract two entities with different categories that overlap using two independent questions. This strategy also means all the queries are encoded with significant prior information about the entity category to extract, giving more information for the model to extract accurately!

Future Work

  1. Explore different variants of model architecture

Ryan

Ryan

Data Scientist

Leave a Reply