View
1.456
Download
3
Category
Preview:
DESCRIPTION
Named entity recognition and disambiguation are important for information extraction and populating knowledge bases. Detecting and classifying named entities has traditionally been taken on by the natural language processing community, whilst linking of entities to external resources, such as DBpedia and GeoNames, has been the domain of the Semantic Web community. As these tasks are treated in different communities, it is difficult to assess the performance of these tasks combined. We present results on an evaluation of the NERD-ML approach on newswire and tweets for both Named Entity Recognition and Named Entity Disambiguation. Presented at CLIN 24: http://clin24.inl.nl/ http://nerd.eurecom.fr https://github.com/giusepperizzo/nerdml
Citation preview
EVALUATING NAMED ENTITY RECOGNITION AND DISAMBIGUATION IN NEWS AND TWEETS
Giuseppe Rizzo Università degli studi di Torino Marieke van Erp VU University Amsterdam Raphaël Troncy EURECOM
EVALUATING NER & NED• NER typically an NLP task (MUC, CoNLL, ACE)
• NED took flight with availability of large structured resources (Wikipedia, DBpedia, Freebase)
• Tools for NER and NED have started popping up outside regular research outlets (TextRazor, DBpedia Spotlight, AlchemyAPI)
• Unclear how well these tools perform
THIS WORK
• Evaluation & comparison of 10 out-of-the-box NER and NED tools through NERD API as well as a combination of the tools in NERD-ML
• Two types of data: Newswire & Tweets
• http://nerd.eurecom.fr
• Ontology, REST API & UI
• Uniform access to 12 different extractors/linkers: AlchemyAPI, DBpedia Spotlight, Extractiv, Lupedia, OpenCalais, Saplo, SemiTags, TextRazor, THD, Wikimeta, Yahoo! Content Analysis, Zemanta
NERD-ML
• The aim of NERD-ML is to combine the knowledge of the different extractors into a better named entity recogniser
• Uses NERD predictions, Stanford NER & extra features
• Naive Bayes, k-NN, SMO
DATA
• CoNLL 2003 English NER with AIDA CoNLL-YAGO links to Wikipedia (5,648 NEs/4,485 links in test set)
• Making Sense of Microposts 2013 (MSM’13) for NER on Twitter domain + 62 randomly selected tweets from Ritter et al.’s corpus with links to DBpedia resources (MSM: 1,538 NEs/Ritter : 177 links in test set)
020
4060
80100
F1
PER LOC ORG MISC OVERALL
010
2030
4050
6070
8090
100
020
4060
80100
Precision
PER LOC ORG MISC OVERALL
010
2030
4050
6070
8090
100
RESULTS NER NEWSWIRE
Precision F1AlchemyAPIDBpedia SpotlightExtractivLupediaOpenCalaisSaploTextrazorYahooWikimetaZemantaStanford NERNERD-ML Run01NERD-ML Run02NERD-ML Run03Upper Limit
020
4060
80100
Recall
PER LOC ORG MISC OVERALL
010
2030
4050
6070
8090
100
Recall
RESULTS NER MSM
020
4060
80100
Precision
PER LOC ORG MISC OVERALL
010
2030
4050
6070
8090
100
020
4060
80100
Precision
PER LOC ORG MISC OVERALL
010
2030
4050
6070
8090
100
020
4060
80100
Precision
PER LOC ORG MISC OVERALL
010
2030
4050
6070
8090
100
Precision Recall F1 AlchemyAPIDBpedia SpotlightExtractivLupediaOpenCalaisSaploTextrazorWikimetaZemantaRitter et al.Stanford NERNERD-ML Run01NERD-ML Run02NERD-ML Run03Upper Limit
RESULTS NED
AlchemyAPI DBpedia Spotlight Extractiv Lupedia Textrazor Yahoo Zemanta
AIDA-YAGO 70.63 26.93 51.31 57.98 49.21 0.0 35.58
TWEETS 53.85 25.13 74.07 65.38 58.14 76.00 48.57
DISCUSSION
• Still a ways to go, but for certain classes NER is getting close to really good results
• MISC class is (and probably always will be?) hard
• Bigger datasets needed (for tweets and NED)
• NED task can use standardisation
THANK YOU FOR LISTENING
• Try out our code at: https://github.com/giusepperizzo/nerdml
ACKNOWLEDGEMENTS
This research is funded through the LinkedTV and NewsReader projects, both funded by the European Union’s 7th Framework Programme grants GA 287911 and ICT-316404).
Recommended