Embedding-Based Techniques MATRICES, TENSORS, AND NEURAL NETWORKS

Transcription

1 Embedding-Based Techniques MATRICES, TENSORS, AND NEURAL NETWORKS

2 Probabilistic Models: Downsides Limitation to Logical Relations Embeddings Representation restricted by manual design Clustering? Assymetric implications? Information flows through these s Difficult to generalize to unseen entities/s Everything as dense vectors Can capture many s Learned from data Computational Complexity of Algorithms Complexity depends on explicit dimensionality Often NP-Hard, in size of data More rules, more expensive inference Query-time inference is sometimes NP-Hard Not trivial to parallelize, or use GPUs Complexity depends on latent dimensions Learning using stochastic gradient, back-propagation Querying is often cheap GPU-parallelism friendly 2

3 Two Related Tasks surface pattern Relation Extraction surface pattern Graph Completion 3

4 Two Related Tasks surface pattern Relation Extraction surface pattern Graph Completion 4

5 Relation Extraction From Text John was born in Liverpool, to Julia and Alfred Lennon. born in, to Alfred Lennon Liverpool was born in John Lennon was born to and was born to born in, to Julia Lennon 5

6 Relation Extraction From Text John was born in Liverpool, to Julia and Alfred Lennon. livedin Alfred Lennon born in, to Liverpool was born in birthplace John Lennon childof was born to childof was born to and born in, to livedin Julia Lennon 6

7 Distant Supervision Liverpool was born in birthplace John Lennon No direct supervision gives us this information. Supervised: Too expensive to label sentences Rule-based: Too much variety in language Both only work for a small set of s, i.e. 10s, not 100s Honolulu is native to was born in birthplace visited met the senator from Barack Obama 7

8 Relation Extraction as a Matrix John was born in Liverpool, to Julia and Alfred Lennon. John Lennon, Liverpool 1? John Lennon, Julia Lennon 1 Entity Pairs John Lennon, Alfred Lennon Julia Lennon, Alfred Lennon Barack Obama, Hawaii ? Barack Obama, Michelle Obama 1 1 Universal Schema, Riedel et al, NAACL (2013) 8

9 Matrix Factorization n m s n k k m s pairs pairs X bornin(john,liverpool) Universal Schema, Riedel et al, NAACL (2013) 9

10 Training: Stochastic Updates s s pairs R 0 (x, y) R(i, j) pairs Pick an observed cell, R(i, j) : Update p ij & r R such that R(i, j) is higher Pick any random cell, assume it is negative: Update & such that is lower p xy r R 0 R 0 (x, y) 10

11 Relation Embeddings 11

12 Embeddings ~ Logical Relations Relation Embeddings, w Similar embedding for 2 s denote they are paraphrases is married to, spouseof(x,y), /person/spouse One embedding can be contained by another w(topemployeeof) w(employeeof) topemployeeof(x,y) employeeof(x,y) Can capture logical patterns, without needing to specify them! Entity Pair Embeddings, v Similar entity pairs denote similar s between them Entity pairs may describe multiple s independent foundedby and employeeof s From Sebastian Riedel 12

13 Similar Embeddings similar underlying embedding X own percentage of Y X buy stake in Y similar embedding Time, Inc Amer. Tel. and Comm. Volvo Scania A.B. Campeau Federated Dept Stores Apple HP Successfully predicts Volvo owns percentage of Scania A.B. from Volvo bought a stake in Scania A.B. From Sebastian Riedel 13

14 Implications X historian at Y X professor at Y X professor at Y X historian at Y (Freeman,Harvard) (Boyle,OhioState) Kevin Boyle Ohio State R. Freeman Harvard 1 1 Learns asymmetric entailment: PER historian at UNIV PER professor at UNIV But, PER professor at UNIV PER historian at UNIV From Sebastian Riedel 14

15 Two Related Tasks surface pattern Relation Extraction surface pattern Graph Completion 15

16 Graph Completion livedin Alfred Lennon born in, to Liverpool was born in birthplace John Lennon childof was born to childof was born to and born in, to livedin Julia Lennon 16

17 Graph Completion livedin Alfred Lennon childof Liverpool birthplace John Lennon childof spouse spouse livedin Julia Lennon 17

18 Tensor Formulation of KG R Does an unseen exist? E e1 r e2 E 18

19 Factorize that Tensor E R k k k E E R E S(r(a, b)) = f(v r, v a, v b ) 19

20 Many Different Factorizations CANDECOMP/PARAFAC-Decomposition S (r(a, b)) = X k R r,k e a,k e b,k Tucker2 and RESCAL Decompositions Model E S (r(a, b)) = (R r e a ) e b Holographic Embeddings S (r(a, b)) = R r,1 e a + R r,2 e b Not tensor factorization (per se) S(r(a, b)) = R r (e a? e b ) HOLE: Nickel et al, AAAI (2016), Model E: Riedel et al, NAACL (2013), RESCAL: Nickel et al, WWW (2012), CP: Harshman (1970), Tucker2: Tucker (1966) 20

21 Translation Embeddings TransE birthplace r Honolulu e2 S (r(a, b)) = ke a + R r e b k 2 2 TransH e1 Barack Obama Liverpool S (r(a, b)) = ke? a + R r e? b k 2 2 e? a = e a w T r e a w r birthplace TransR John Lennon S (r(a, b)) = ke a M r + R r e b M r k 2 2 TransE: Bordes et al. XXX (2011), TransH: Bordes et al. XXX (2011), TransR: Bordes et al. XXX (2011) 21

22 Parameter Estimation R Observed cell: increase score S (r(a, b)) E e1 r e2 E Unobserved cell: decrease score S (r 0 (x, y)) 22

23 Matrix vs Tensor Factorization Vectors for each entity pair Can only predict for entity pairs that appear in text together No sharing for same entity in different entity pairs Vectors for each entity Assume entity pairs are low-rank But many s are not! Spouse: you can have only ~1 Cannot learn pair specific information 23

24 What they can, and can t, do.. Red: deterministically implied by Black - needs pair-specific embedding - Only F is able to generalize Green: needs to estimate entity types - needs entity-specific embedding - Tensor factorization generalizes, F doesn't Blue: implied by Red and Green - Nothing works much better than random From Singh et al. VSM (2015), 24

25 Joint Extraction+Completion surface pattern Relation Extraction surface pattern Joint Model Graph Completion 25

26 Compositional Neural Models So far, we re learning vectors for each entity/surface pattern/.. But learning vectors independently ignores composition Composition in Surface Patterns Every surface pattern is not unique Synonymy: A is B s spouse. A is married to B. Composition in Relation Paths Every path is not unique Explicit: A parent B, B parent C A grandparent C Inverse: X is Y s parent. Y is one of X s children. Implicit: X bornincity Y, Y cityinstate Z X borninstate Z Can the representation learn this? Can the representation capture this? 26

27 Composing Dependency Paths was born to s parents are \parentsof (never appears in training data) But we don t need linked data to know they mean similar things Use neural networks to produce the embeddings from text! NN NN was born to s parents are \parentsof Verga et al (2016), 27

28 Composing Relational Paths countrybasedin statebasedin NN NN isbasedin statelocatedin countrylocatedin Microsoft Seattle Washington USA Neelakantan et al (2015), Lin et al, EMNLP (2015), 28

29 Review: Embedding Techniques Two Related Tasks: Relation Extraction from Text Graph (or Link) Completion Relation Extraction: Matrix Factorization Approaches Graph Completion: Tensor Factorization Approaches Compositional Neural Models Compose over dependency paths Compose over paths 29

30 Using Tutorial Embeddings Overview in MLNs Part 1: Knowledge Graphs Part 2: Knowledge Extraction Part 3: Graph Construction Part 4: Critical Analysis 302