Multi-Domain Named Entity Recognition with Genre-Aware and Agnostic Inference

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8476–8488 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 8476 Multi-Domain Named Entity Recognition with Genre-Aware and Agnostic Inference Jing Wang∗, Mayank Kulkarni∗, Daniel Preot¸iuc-Pietro Bloomberg New York, New York, USA {jwang1621,mkulkarni24,dpreotiucpie}@bloomberg.net Abstract Named entity recognition is a key component of many text processing pipelines and it is thus essential for this component to be robust to dif- ferent types of input. However, domain trans- fer of NER models with data from multiple genres has not been widely studied. To this end, we conduct NER experiments in three pre- dictive setups on data from: a) multiple do- mains; b) multiple domains where the genre la- bel is unknown at inference time; c) domains not encountered in training. We introduce a new architecture tailored to this task by us- ing shared and private domain parameters and multi-task learning. This consistently outper- forms all other baseline and competitive meth- ods on all three experimental setups, with dif- ferences ranging between +1.95 to +3.11 aver- age F1 across multiple genres when compared to standard approaches. These results illustrate the challenges that need to be taken into ac- count when building real-world NLP applica- tions that are robust to various types of text and the methods that can help, at least partially, al- leviate these issues. 1 Introduction Accurately identifying named entities and their type in texts is a key processing step for many NLP applications. Named entity recognition (NER) is an important component in several tasks in- cluding named entity linking (Cucerzan, 2007), co-reference resolution (Ng and Cardie, 2002), question answering (Krishnamurthy and Mitchell, 2015), relation extraction (Culotta and Sorensen, 2004) and usually sits upstream of analytics such as sentiment (Pang and Lee, 2004) or stance (Moham- mad et al., 2016). Building robust NER models to accurately tag and adapt to heterogeneous types of text is thus paramount. Recent research focused on ∗*Equal Contribution improving the overall performance of NER models on specific data sets. Yet NER models show rela- tively high variance even when trained on the same data (Reimers and Gurevych, 2017) and poorly gen- eralize when tested on data from different genres1, especially if these contain entity mentions unseen in the test data (Augenstein et al., 2017; Agarwal et al., 2020). Despite this, research on NER models robust to different types of input is usually limited to the stan- dard domain adaptation scenario: a single source domain rich in training data and a single target do- main with limited or no training data (Lin and Lu, 2018). We argue that this is an over-simplified ex- perimental setup that is not typical for how NER models are used in real-world applications. Ide- ally, NER models use all available data, regardless of genre, and perform inference on data from any genre, even if this was not encountered in training. In this scenario, simply pooling all the available data is likely sub-optimal as genre-specific differ- ences in named …

Content truncated to 500 words