Entity Typing Using Distributional Semantics and DBpedia

Entity Typing Using Distributional Semantics and DBpediaMarieke van Erp and Piek Vossen

Conclusions

• Finegrained entity typing is necessary for semantic queries over text

• Search space for word2vec is large, topics help

• Combining Distributional Semantics with DBpedia can help overcome NIL and Dark Entities

https://github.com/MvanErp/entity-typing/

Dark entities: little or no information available in KB

https://github.com/MvanErp/entity-typing/

Dark entities: little or no information available in KB

https://github.com/MvanErp/entity-typing/

Distributional Semantics

• Similar concepts (denoted by words) occur in similar contexts

• Word2Vec (Mikolov et al., 2013) explores this notion in a popular implementation

SushiTeriyakiUdon

Okonomiyaki

SobaSashimi

KimonoYukataNemakiKFC

Steak

HamburgerMcDonald’s

JeansT-shirt

Skirt

Research Question:

• Can we predict the type of the concept ‘Sushi’ by modelling it in a distributional semantics space and comparing its vector to the vectors of concepts for which we do know the type?

SushiTeriyakiUdon

Okonomiyaki

SobaSashimi

KimonoYukataNemakiKFC

Steak

HamburgerMcDonald’s

JeansT-shirt

Skirt

Setup

• 7 Named Entity Linking Benchmark datasets (AIDA-YAGO, 2014 NEEL, 2015 NEEL, OKE2015, RSS500, WES2015, Wikinews)

• 3 Word2Vec models: GoogleNews, English Wikipedia, Reuters RCV1*

• Compare all entities within datasets to each other and return highest ranking type (as taken from DBpedia)

* AIDA-YAGO is part of Reuters RCV1

https://github.com/MvanErp/entity-typing/

Initial results

• Not so great?

https://github.com/MvanErp/entity-typing/

Initial results (some footnotes)

• Ranking approach favours fine-grained entity types

• The Word2Vec corpus matters! NEEL2014&2015 are derived from Tweets, typically low coverage when querying news

• Smaller datasets (Wikinews, WES2015, OKE2015) do better?

https://github.com/MvanErp/entity-typing/

Let’s zoom in on topics

• Initially, all entities within a benchmark dataset were compared to all other entities.

• What if we only compare entities from sports documents to other entities from sports documents?

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22

AIDA−YAGO Coarsegrained Categories GoogleNews Fine

20

40

60

80

1001510

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22

AIDA−YAGO Coarsegrained Categories RCV1 Fine

20

40

60

80

1001510

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22

AIDA−YAGO Coarsegrained Categories Wikipedia Fine

20

40

60

80

1001510

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69

AIDA−YAGO Finegrained Categories GoogleNews Fine

20

40

60

80

1001510

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69

AIDA−YAGO Finegrained Categories RCV1 Fine

20

40

60

80

1001510

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69

AIDA−YAGO Finegrained Categories Wikipedia Fine

20

40

60

80

1001510

https://github.com/MvanErp/entity-typing/

Conclusions and Future Work

• Difficult task, but topics help

• Ranking needs to be improved

• Multi-class classification (KFC: food & organisation, Arnold Schwarzenegger: Actor & Politician)

• What else can we discover beyond type?

https://github.com/MvanErp/entity-typing/

Thank you!

https://github.com/MvanErp/entity-typing/

This research was made possible by the CLARIAH-CORE project financed by NWO.

http://www.clariah.nl