Enhancing Named Entity Recognition in Twitter Messages Using Entity Linking Ikuya Yamada1 2 3 Hideaki Takeda2 Yoshiyasu Takefuji3 [email protected] [email protected] [email protected] 1Studio Ousia, 4489-105-221 Endo, Fujisawa, Kanagawa, Japan 2National Institute of Informatics, 2-1-2 Hitotsubashi, Chiyoda, Tokyo, Japan 3Keio University, 5322 Endo, Fujisawa, Kanagawa, Japan Abstract Freebase (Bollacker et al., 2008)). Because of the recent emergence of large online knowledge In this paper, we describe our approach bases (KB), EL has recently gained significant at- for Named Entity Recognition in Twitter, tention. It is evident that the performance of EL a shared task for ACL 2015 Workshop also degrades when analyzing tweets (Derczynski on Noisy User-generated Text (Baldwin et et al., 2015; Meij et al., 2012). However, Guo al., 2015). Because of the noisy, short, et al. (Guo et al., 2013) recently revealed that and colloquial nature of Twitter, the per- the main failures of Twitter EL are caused while formance of Named Entity Recognition detecting entity mentions from text, because ex- (NER) degrades significantly. To address isting EL methods usually address the mention this problem, we propose a novel method detection task by using external NER software to enhance the performance of the Twitter whose performance is unreliable when processing NER task by using Entity Linking which tweets. Consequently, several approaches (Guo et is a method for detecting entity mentions al., 2013; Yamada et al., 2015) have been proposed in text and resolving them to correspond- with enhanced abilities that address the task in an ing entries in knowledge bases such as end-to-end manner without completely depending Wikipedia. Our method is based on super- on NER software. vised machine-learning and uses the high- The main objective of this study is to investigate quality knowledge obtained from several the possibility of enhancing the performance of open knowledge bases. In comparison Twitter NER by using an end-to-end EL. Although with the other systems proposed for this EL is typically performed after NER in most of shared task, our method achieved the best the existing methods, our approach performs EL performance. before NER and uses the EL results to enhance the 1 Introduction NER performance. Resolving the entity mentions to the KB entries enables us to use the high-quality Named Entity Recognition (NER) refers to the knowledge in KB for enhancing the NER perfor- task of identifying mentions of entities (e.g., per- mance. This knowledge includes things such as sons, locations, organizations) within text. Be- the popularity of the entity, the classes of the en- cause of the noisy, short, and colloquial nature of tity, and the likelihood that the entity appears in Twitter messages (or tweets), the performance of the given context. standard NER software significantly suffers. For We begin by briefly introducing our end-to-end example, Derczynski et al. (Derczynski et al., EL method that specifically focuses on tweets. 2015) recently demonstrated that the performance Our EL method is based on supervised machine- of various state-of-the-art NER software (e.g., learning and addresses the task in an end-to-end Stanford NER and ANNIE) is typically lower than manner. It considers every possible n-gram as a 50% F11 for tweets. candidate entity mention and detects the mention Entity Linking (EL) refers to the task of de- with a corresponding link to a KB entry if the men- tecting textual entity mentions and linking them tion exists in the KB. Furthermore, it can handle to corresponding entries within knowledge bases mentions that appear as irregular forms (e.g., mis- (e.g., Wikipedia, DBpedia (Auer et al., 2007), spellings, abbreviations, acronyms) using several 1The harmonic mean of precision and recall. approximate string matching algorithms. 136 Proceedings of the ACL 2015 Workshop on Noisy User-generated Text, pages 136–140, Beijing, China, July 31, 2015. c 2015 Association for Computational Linguistics The NER task is split into two separate sub- 2.2.1 Mention Candidate Generation tasks: segmentation and classification. During Our system first generates a set of candidate en- segmentation, the entity mentions are detected tity mentions with the set of corresponding refer- from tweets. Then, the entity mentions are clas- ent entities. The system takes all the n-grams of sified into the predefined entity types. Both tasks n 10 and looks up each n-gram in a dictionary, ≤ involve supervised machine-learning with various treats an n-gram as a candidate mention if it exists features. in the dictionary, and finally, generates an output For the segmentation task, we use data obtained of pairs of mentions and their associated possible from the KB of the corresponding entity men- referent entities. tion detected by the EL and the output of a NER software as the main machine-learning features. Mention-Entity Dictionary: The system uses a Furthermore, we include several common features mention-entity dictionary that maps a mention sur- used in traditional NER methods. face (e.g., apple) to the possible referent entities For the classification task, the following three (e.g., Apple Inc., Apple (food)). The possible types of features are used as primary features: 1) mention surfaces of an entity are extracted from the KB types of the entity detected by the EL, 2) the corresponding Wikipedia page title, the page the entity types detected by the NER software, and titles of the Wikipedia pages that redirect to the 3) the vector representation of the entity mention page of the entity, and anchor texts in Wikipedia derived from word embeddings. The entity’s KB articles that point to the page of the entity. We types are extracted from the corresponding entries constructed this dictionary using the January 2015 in DBpedia and Freebase. Furthermore, the vec- dump of Wikipedia. tor representation of the entity mention is derived Approximate Candidate Generation: One using GloVe word embeddings (Pennington et al., major problem of the mention candidate gen- 2014). eration task is that many entity mentions in To train and evaluate our system, we used the tweets cannot be detected because they appear as dataset given by the Named Entity Recognition in irregular forms (e.g., misspellings, abbreviations). Twitter shared task. Our proposed method signifi- In order to address this problem, we introduce cantly outperformed the second ranked system by the following three approximate string-matching a wide margin; 10.3% F1 at the segmentation task, methods to improve the ability of this task: and 5.0% F1 at the end-to-end (both the segmen- tation and the classification) task. Fuzzy match searches the mention candidates • that have text surfaces within a certain dis- 2 The Proposed System tance of the surface of the n-gram measured by edit distance. 2.1 Preprocessing Approximate token search obtains mention • The system first assigns part-of-speech tags to candidates whose text surfaces have a signif- the resulting tokens using ARK Twitter Part-of- icant ratio of words in common with the sur- Speech Tagger (Gimpel et al., 2011). It also to- face of the n-gram. kenizes Twitter hashtags using our enhanced im- Acronym search retrieves mention candidates plementation of the hashtag tokenization. • with possible acronyms2 that include the sur- face of the n-gram. 2.2 Entity Linking We formalize our EL task as follows: Given a When using the above methods, we observed tweet, our goal is to recognize a set of entity that the number of mention candidates becomes mentions (e.g., Obama, President Obama, Barack very large. To deal with this, we use a simple fil- Obama) that appear in a tweet, and then resolve tering method based on soft tf-idf (Cohen et al., the mentions into entities (e.g., Barack Obama) 2003); we simply use only the mention candidates in Wikipedia if they exist. that have a similarity greater than a threshold mea- Our EL system addresses the task using the fol- sured by the soft tf-idf. We use 0.9 as the threshold lowing two steps; mention candidate generation 2We generate acronyms by tokenizing the mention surface and mention detection and disambiguation. and simply taking the first characters of the resulting tokens. 137 because this achieves the best performance in our 2015). Here, we adopt two models of Stanford experiments of EL. NER to enhance the performance: 1) the standard three-class model which is included in the soft- 2.2.2 Mention Detection and Disambiguation ware and 2) a model that does not use capitaliza- Given a pair of a mention and its possible refer- tion as a feature, in order to deal with the unrelia- ent entity, it needs to be determined if the possible bility of capitalization in tweets. referent entity is indeed the correct one for its as- The results of the NER and the KB knowledge sociated mention. of the corresponding entity mention detected by In this system, we use a supervised machine- the EL are used as the primary machine-learning learning algorithm to assign a relevance score to features. We also include features that are tradi- each of the pairs and select the entity mention with tionally used in NER such as part-of-speech tags the highest score. We use random forest as the and the capitalization features. Furthermore, the machine-learning algorithm. ratio of the capitalized words in the tweet is also Here, we use machine-learning features that are used as an indicator of the reliability of the capi- mostly identical to the method proposed previ- talization. ously (Yamada et al., 2015). Basically, we use var- The machine-learning features for this step in- ious features that are commonly observed in EL clude: studies and enhance the performance further by introducing two new features: 1) the entity pop- EL relevance score∗: The relevance score of • ularity knowledge extracted from Wikipedia page the entity mention assigned by the previous views3, and 2) the contextual similarity between EL step.