Enhancing Entity Linking by Combining NER Models
2 likes909 views
The document describes enhancements made to the ADEL entity linking framework. ADEL combines multiple named entity recognition models and uses a combination of linguistic and dictionary-based approaches. New features in ADEL include using a generic API to interface with NLP tools, combining multiple CRF models for entity extraction, clustering nil entities, and developing a new backend using Elasticsearch and Couchbase. The document compares the performance of the original 2015 version of ADEL to the new 2016 version on standard entity linking tasks and datasets.
More Related Content
Similar to Enhancing Entity Linking by Combining NER Models(20)
Enhancing Entity Linking by Combining NER Models
- 1. Julien Plu, GiuseppeRizzo, Raphaël Troncy [email protected] @julienplu Enhancing Entity Linking by Combining NER Models
- 2. ADEL in anutshell § ADaptive Entity Linking Framework: http://multimediasemantics.github.io/adel/ § OKE2015 Challenge winner § ADEL (in 2015): Ø Use Stanford POS and Stanford NER for extracting named entities Ø Train and use one single CRF model via Stanford NER Ø NIL entities were all distinct 2016/05/31 - OKE Challenge at ESWC2016 – Heraklion, Crete - 2
- 3. What’s new inADEL (2016)? § Use a generic NIF wrapper over NLP Core systems (Stanford CoreNLP, OpenNLP, etc.) as an external API § Use an algorithm to combine multiple CRFs model § Cluster NIL entities § Propose a generic index format and develop a new backend (Elasticsearch + Couchbase) § Filter candidate links based on types 2016/05/31 - OKE Challenge at ESWC2016 – Heraklion, Crete - 3
- 4. Different Approaches E2E approaches: Adictionary of mentions and links is built from a referent KB. A text is split in n-grams that are used to look up candidate links from the dictionary. A selection function is used to pick up the best match Linguistic-based approaches: A text is parsed by a NER classifier. Entity mentions are used to look up resources in a referent KB. A ranking function is used to select the best match ADEL is a combination of both to make a hybrid approach 2016/05/31 - OKE Challenge at ESWC2016 – Heraklion, Crete - 4
- 5. ADEL from 30,000foots 2016/05/31 - OKE Challenge at ESWC2016 – Heraklion, Crete ADEL Entity Extraction Entity Linking Index - 5
- 6. § POS Tagger: Øbidirectional CMM (left to right and right to left) § NER Combiner: Ø Use a combination of CRF with Gibbs sampling (Monte Carlo as graph inference method) models. A simple CRF model could be: PER PER PERO OOO X X X X XX XXXX X set of features for the current word: word capitalized, previous word is “de”, next word is a NNP, … Suppose P(PER | X, PER, O, LOC) = P(PER | X, neighbors(PER)) then X with PER is a CRF Jimmy Page , knowing the professionalism of John Paul Jones Entity Extraction: Extractors Module PER PERO 2016/05/31 - OKE Challenge at ESWC2016 – Heraklion, Crete - 6
- 7. Entity Extraction: OverlapResolution § Detect overlaps among boundaries of entities coming from the extractors § Different heuristics can be applied: Ø Merge: (“United States” and “States of America” => “United States of America”) default behavior Ø Simple Substring: (“Florence” and “Florence May Harding” => ”Florence” and “May Harding”) Ø Smart Substring: (”Giants of New York” and “New York” => “Giants” and “New York”) 2016/05/31 - OKE Challenge at ESWC2016 – Heraklion, Crete - 7
- 8. Index: Indexing § UseDBpedia and Wikipedia as knowledge bases § Integrate external data such as PageRank scores from Hasso Platner Institute § Backend sytem with Elasticsearch and Couchbase § Turn DBpedia and Wikipedia into a CSV-based generic format 2016/05/31 - OKE Challenge at ESWC2016 – Heraklion, Crete - 8
- 9. Entity Linking: Linkingtasks § Generate candidate links for all extracted mentions: Ø If any, they go to the linking method Ø If not, they are linked to NIL via NIL Clustering module § Linking method: Ø Filter out candidates that have different types than the one given by NER Ø ADEL linear formula: 𝑟 𝑙 = 𝑎. 𝐿 𝑚, 𝑡𝑖𝑡𝑙𝑒 + 𝑏. max 𝐿 𝑚, 𝑅 + 𝑐. max 𝐿 𝑚, 𝐷 . 𝑃𝑅(𝑙) r(l): the score of the candidate l L: the Levenshtein distance m: the extracted mention title: the title of the candidate l R: the set of redirect pages associated to the candidate l D: the set of disambiguation pages associated to the candidate l PR: Pagerank associated to the candidate l a, b and c are weights following the properties: a > b > c and a + b + c = 1 2016/05/31 - OKE Challenge at ESWC2016 – Heraklion, Crete - 9
- 10. ADEL 2015 vsADEL 2016 in OKE § ADEL (2015 version) over OKE2015 test set (after adjudication) § ADEL (2016 version) over OKE2015 test set § ADEL (2016 version) over OKE2016 training set (4-fold validation) 2016/05/31 - OKE Challenge at ESWC2016 – Heraklion, Crete - 10 Precision Recall F-measure extraction 78.2 65.4 71.2 recognition 65.8 54.8 59.8 linking 49.4 46.6 48 Precision Recall F-measure extraction 85.1 89.7 87.3 recognition 75.3 59 66.2 linking 85.4 42.7 57 Precision Recall F-measure extraction 81 88.7 84.7 recognition 78.1 85.4 81.6 linking 57.4 55.7 56.5
- 11. Questions? Thank you forlistening! 2016/05/31 - OKE Challenge at ESWC2016 – Heraklion, Crete - 11 http://multimediasemantics.github.io/adel http://jplu.github.io [email protected] @julienplu http://www.slideshare.net/julienplu