Modeling Noisiness to Recognize Named Entities using Multitask Neural Networks on Social Media

Proceedings of NAACL-HLT 2018, pages 1401–1412 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics Modeling Noisiness to Recognize Named Entities using Multitask Neural Networks on Social Media Gustavo Aguilar⋆ A. Pastor L´opez-Monroy⋆ Fabio A. Gonz´alez† Thamar Solorio⋆ ⋆Department of Computer Science, University of Houston †Systems and Computer Engineering Department, Universidad Nacional de Colombia {gaguilaralas, alopezmonroy}@uh.edu fagonzalezo@unal.edu.co, solorio@cs.uh.edu Abstract Recognizing named entities in a document is a key task in many NLP applications. Although current state-of-the-art approaches to this task reach a high performance on clean text (e.g. newswire genres), those algorithms dramati- cally degrade when they are moved to noisy environments such as social media domains. We present two systems that address the chal- lenges of processing social media data using character-level phonetics and phonology, word embeddings, and Part-of-Speech tags as fea- tures. The first model is a multitask end-to- end Bidirectional Long Short-Term Memory (BLSTM)-Conditional Random Field (CRF) network whose output layer contains two CRF classifiers. The second model uses a multi- task BLSTM network as feature extractor that transfers the learning to a CRF classifier for the final prediction. Our systems outperform the current F1 scores of the state of the art on the Workshop on Noisy User-generated Text 2017 dataset by 2.45% and 3.69%, establish- ing a more suitable approach for social media environments. 1 Introduction One of the core tasks in Natural Language Processing (NLP) is Named Entity Recognition (NER). NER is a sequence tagging task that con- sists in selecting the words that describe entities and recognizing their types (e.g., a person, lo- cation, company, etc.). Figure 1 shows exam- ples of sentences from different domains that con- tain named entities. Recognizing entities in run- ning text is typically one of the first tasks in the pipeline of many NLP applications, including ma- chine translation, summarization, sentiment anal- ysis, and question answering. Traditional machine learning systems have proven to be effective in formal text, where gram- matical errors are minimal and writers stick to CoNLL 2003 [Spanish]MISC Farm Minister [Loyola de Palacio]PER had earlier accused [Fischler]PER at an [EU]ORG farm ministers ’ meeting of causing unjustified alarm through ” dangerous generalisation . ” WNUT 2017, Twitter domain been listenin to [trey]PER alllll week ... can u luv someone u never met ?? bcuz i think im in luv yeeuuuuppp !!! Figure 1: Examples from the CoNLL 2003 and the WNUT 2017 datasets. The noise from the WNUT dataset makes a clear difference from one text to the other, establishing new challenges to the current state- of-the-art systems on formal text. The words in bold are grouped to described the entities. the rules of the written language (Florian et al., 2003a; Chieu and Ng, 2003a). However, those tra- ditional systems dramatically fail on informal text, where improper grammatical structures, spelling inconsistencies, and slang vocabulary prevail (Rit- ter et al., 2011). For instance, Table 1 shows a snapshot of NER systems’ performance during the last years, where the results drop from 96.49% …

Content truncated to 500 words