Incorporating Social Context and Domain Knowledge for Entity Recognition

Transcription

1 Incorporating Social Context and Domain Knowledge for Entity Recognition Jie Tang, Zhanpeng Fang Department of Computer Science, Tsinghua University Jimeng Sun College of Computing, Georgia Institute of Technology 1

2 Entity Recognition in Social Media People use blogs, forums, and review sites to share opinions on politicians or products. One fundamental analytic issue is to recognize entity instances from the UGC short documents. However, the problem is very challenging S4 vs. Samsung Galaxy S4 Fruit company vs. Apple Inc. Peace West King vs. Xilai Bo (a sensitive Chinese politician) 2

3 A Concrete Example Social Network Documents Knowledge Base A Both Disease 1 and Disease 2 have symptom 1 reply Symptom Health B Re: Remember D2 also has symptom 3. Disease retweet C Treatment RT: S1 can be resolved by treatment 1 //@A: Both Disease 1 and Disease 2... Challenges: short text + social networks + domain knowledge =? 3

4 Related Work 4 Entity recognition Modeling as a ranking problem based on boosting and voted perceptron (Collins [9]) Incorporating long-distance dependency (Finkel et al. [13]) Use Labeled LDA [26] to exploit Freebase to help extraction (Ritter et al. [27]) Entity morph (Huang et al. [17]) Entity resolution A collective method for entity resolution in relational data (Bhattacharya and Getoor [4]) A hierarchical topic model for resolving name ambiguity (Kataria et al. [18]) Name disambiguation in digital libraries (Tang et al. [32])

5 Approach Framework SOCINST 5

6 Preliminary: Sequential Labeling OTH OTH OTH OTH OTH The label results y LOC LOC LOC LOC LOC POL POL POL POL POL The input text x 6 Peace-West King from Chongqing fell y * = max y p(y x; f,θ) where f represents features and Θ are model parameters.

7 Sequential Labeling with CRFs y POL POL OTH LOC OTH x Peace-West King from Chongqing fell p(y x,λ,µ) = 1 Z exp( λ f (x, y ) + µ k k i i j f j (x, y i, y i+1 )) i µ and λ are parameters to be learned from the training data. k i j f k denotes the k-th feature defined for token x i f j denotes the j-th feature defined for two consecutive tokens x i ; and x j ; 7

8 Sequential Labeling with CRFs y POL POL OTH LOC OTH x Peace-West King from Chongqing fell p(y x,λ,µ) = 1 Z exp( λ f (x, y ) + µ k k i i j f j (x, y i, y i+1 )) i µ and λ are parameters to be learned from the training data. Performance of the model will be bad when dealing with short-text due to sparsity k i j f k denotes the k-th feature defined for token x i f j denotes the j-th feature defined for two consecutive tokens x i ; and x j ; 8

9 Sequential Labeling Incorporating Topics y θ x P(z y) P(x z) POL POL OTH LOC OTH z 1 z 2 z 3 z T Peace-West King from Chongqing fell p(y x,θ,λ,µ) = 1 Z exp( λ f (x,θ, y ) + µ k k i i i j f j (x,θ, y i, y i+1 )) i k i j 9

10 Latent Dirichlet Allocation Distribution of document over topics β ϕ k k [1,K] Distribution of topic over words α θ m z m,n x m,n Word n [1,N m ] m [1,M] α, β : Prior distributions (Dirichlet distribution) Document Topic K p(x,z,θ,φ α,β) = p(φ z β) p(θ d α ) p(x i φ z ) p(z θ d ) z=1 [5] D. Blei, A. Y. Ng, and M. I. Jordan. Latent dirichlet allocation. JMLR, 3: , M d=1 N d i=1

11 Extend to Model Authorship and Categories TM DS Generative process DS Shafiei TM Milios P(c z) P(w z) P(c z) P(w z) disease 0.23 sympton disease sympton health treatment 0.23 operation treatment operation 0.19 disease Article Liberia Declared Free of Ebola Shafiei and Milios Disease Treatment After the West African nation goes more than a month with no new reported cases of viral infection, the World Health Organization says the country is Ebola-free. [35] J. Tang, J. Zhang, L. Yao, J. Li, L. Zhang, and Z. Su. Arnetminer: Extraction and mining of academic social networks. In 11 KDD 08, pages , 2008

12 ACT Model Generative process: authors words category tag Topic ACT category [35] J. Tang, J. Zhang, L. Yao, J. Li, L. Zhang, and Z. Su. Arnetminer: Extraction and mining of academic social networks. In 12 KDD 08, pages , 2008

13 Still challenges However, we still cannot model domain knowledge and social context! SOCINST: Modeling Domain Knowledge and Social Context Simultaneously 13

14 Modeling Domain Knowledge root β β c 1 c 2 ηβ ηβ θ~dirichlettree(β, η) w 1 w 2... w k β θ~dirichlet(β) ηβ w j [1] D. Andrzejewski, X. Zhu, and M. Craven. Incorporating domain knowledge into topic modeling via dirichlet forest priors. In 14 ICML 09, pages 25 32, 2009.

15 Modeling Social Context v 1 v 2 v 3 θ v1 =<0.1, 0.5,...> θ v2 =<0.3, 0.2,...> θ v3 v θ B θ A B A C θ C User A s Social context is defined as a mixture of topic distributions of neighbors, i.e. j NB(vi ) γ j θ j multinomial mixture! v1v2 =θ v1 +θ v2 v

16 Theoretical Basis Aggregation property of Dirichlet distribution If then Inverse of the aggregation property If then (θ 1,,θ i,θ i+1,,θ K ) Dirichlet(α 1,,α i,α i+1,,α K ) (θ 1,,θ i +θ i+1,,θ K ) Dirichlet(α 1,,α i + α i+1,,α K ) (θ 1,,θ K ) Dirichlet(α 1,,α K ) (θ 1,,τθ i,(1 τ )θ i,,θ K ) Dirichlet(α 1,,τα i,(1 τ )α i,,α K ) 16

17 17 Model Learning

18 Sequential Labeling Incorporating Topics θ v1 =<0.1, 0.5,...> v 1 v 2 v 3 v 12 multinomial mixture! v1v2 =θ v1 +θ v2 θ v2 =<0.3, 0.2,...>... v 123 θ v3 root β β c 1 c 2 ηβ ηβ β θ~dirichlettree(β, η) w 1 w 2 w k... θ~dirichlet(β) ηβ w j p(y x,θ,λ,µ) = 1 Z exp( λ f (x,θ, y ) + µ k k i i i j f j (x,θ, y i, y i+1 )) i k i j 18

19 19 Experiments

20 20 All codes and datasets can be downloaded here Dataset Data Sets Domain #documents #instances #relationships Weibo 1, ,763 I2B ,400 27,175 ICDM 12 Contest 2, NA Goal: Weibo: Our goal is to extract real morph instances in the dataset. I2B2: Our goal here is to extract private health information instances in the dataset. ICDM 12 Contest: Our goal is to recognize product mentions in the dataset.

21 I2B2 HISTORY OF PRESENT ILLNESS : Mr. Blind is a 79-year-old white male with a history of diabetes mellitus, inferior myocardial infarction, who underwent open repair of his increased diverticulum November 13th at Sephsandpot Center. The patient developed hematemesis November 15th and was intubated for respiratory distress. He was transferred to the Valtawnprinceel Community Memorial Hospital for endoscopy and esophagoscopy on the 16th of November which showed a 2 cm linear tear of the esophagus at 30 to 32 cm. Patient Doctor Date Location Hospital 21

22 22 ICDM 12 Contest

23 Results F1-Measure SM RT CRF CRF+AT SOINST Weibo I2B2 ICDM'12 23 SM: Simply extracts all the terms/symbols that are annotated RT: Recognizes target instances from the test data by a set of rule templates CRF: Trains a CRF model using features associated with each token CRF+AT: Uses Author-Topic (AT) [30] to train a model and then it use the learned topics as features for CRF for instance recognition SOCINST: Our proposed model

24 Results SM: Simply extracts all the terms/symbols that are annotated RT: Recognizes target instances from the test data by a set of rule templates. CRF: Trains a CRF model using features associated with each token CRF+AT: Uses Author-Topic (AT) [30] to train a model and then it use the learned topics as features for CRF for instance recognition SOCINST: Our proposed model 24

25 More Results ICDM 12 Contest Performance comparison of SOCINST and the first place [38] in ICDM 12 Contest. By incorporating the modeling results into the CRF model [38] 25 S. Wu, Z. Fang, and J. Tang. Accurate product name recognition from user generated content. In ICDM 12 Contest.

26 Effects of Social Context and Domain Knowledge SOCINST base we removed both social context and domain knowledge from our method; SOCINST-SC we removed social context from our method; SOCINST-DK we removed domain knowledge from our method; 26

27 27 Parameter Analysis

28 Parameter Analysis (cont.) * All the other hyperparameters fixed The number of topics is set to K = 15 28

29 29 AMiner (

30 Conclusion Study the problem of instance recognition by incorporating social context and domain knowledge Propose a topic modeling approach to learn topics by considering social relationships between users and context information from a domain knowledge base Experimental results on three different datasets validate the effectiveness and the efficiency of the proposed method. 30

31 Future work The general idea of incorporating social context and domain knowledge for entity recognition represents a new research direction Combining the sequential labeling model and the proposed SOCINST into a unified model should be beneficial Further incorporating other social interactions, such as social influence, to help instance recognition is an intriguing direction 31

32 Thank you! Collaborators: Jimeng Sun (Georgia Tech) Zhanpeng Fang (THU) Jie Tang, KEG, Tsinghua U, Download all data & Codes,

33 Modeling Short Text with Topics p d (x) = λ B p(x θ B ) + (1 λ) π d,k p(x θ k ) K k=1 log p(d) = n(x,d)log[λ B p(x θ B ) + (1 λ) π d,k p(x θ k )] x V K k=1 Topic Topic Topic θ 1 θ 2 θ 3 Background B warning 0.3 system Aid 0.1 donation 0.05 support statistics 0.2 loss 0.1 dead Is 0.05 the 0.04 a Document d θ 1 θ 2 θ k B θ d,1 θ d,2 θ d,k θ B Generating word x in doc d in the collection 1 - θ B Parameters: θ B = noise-level (manually set) θ 1 and π are estimated with Maximum Likelihood x 33

34 α θ x β ϕ k k [1,K] x m,n a m x m,n z m,n c m,n n [1,N m ] m [1,M] 34