Detect Missing Attributes for Entities in Knowledge Bases via Hierarchical Clustering

Transcription

1 Detect Missing Attributes for Entities in Knowledge Bases via Hierarchical Clustering Bingfeng Luo, Huanquan Lu, Yigang Diao, Yansong Feng and Dongyan Zhao ICST, Peking University

2 Motivations Entities often have missing attributes Human-made KBs: human negligence Auto-constructed KBs: incompleteness of source data, imperfectness of algorithm Detecting missing attributes is useful Present possible missing attributes to open KB (like Wikipedia) editors Rescore candidate triples proposed by relation extraction tools Taylor Swift (Wikipedia) Occupation: singer Instrument: vocal, guitar Genre: pop, country, rock Born: Pennsylvania The attribute Record Label is missing! (She works for Big Machine)

3 Overview Basic Idea Entities in the same category may share some common attributes Algorithm Framework Build a cluster system over the entities in KB Apply our basic idea in each cluster to find missing attributes Taylor Swift Occupation: singer Instrument: vocal, guitar Genre: pop, country, rock Born: Pennsylvania Michael Jackson Justin Bieber Pop singer The attribute Record Label is missing Lady Gaga Robin Thicke Most of Pop Singers have the attribute Record Label Taylor Swift should also have that attribute Record Label

4 Building Clustering System Entity Representation Each entity has several (attribute, value) pairs The value of each attribute can be represented as a vector (explain later) Each entity can be represented as a set of vectors, and each vector is an attribute value Taylor Swift Occupation: singer Instrument: vocal, guitar Genre: pop, country, rock Born: Pennsylvania Taylor Swift Occupation: [1.1, -0.3, 0.4] Instrument: [0.1, 0.2, -0.4], [-0.4, 1.0, 0.9] Genre: [0.2, 0.3, -0.5], [0.4, 0.8, 0.9], [0.3, 0.9, -0.8] Born: [0.1, 0,3, 0,2]

5 Building Clustering System How to Acquire Attribute Value Vector? Numeric values and date values are not very useful We only consider string values when clustering Use word2vec to convert words or phrases in the attribute value into vectors If a value have several words or phrases, will simple average them up Not Useful Clusters: 1.70, 1.71, 1.80 Useful Clusters: High, Medium, Short Useful clusters need human assistance Birth Date: Height: 1.75 Birth Place: Beijing word2vec [1.32, 0.43, -0.83,, 0.55]

6 Building Clustering System How to cluster entities? Clustering entities directly is hard, since it contains so many vectors Instead, we cluster attribute values within each attribute singer director SCAN Clustering Algorithm (allow overlap) [1.0, 0.8, 0.9] artist [0.95, 0.7, 0.9] = 1 2 [0.9, 0.6, 0.9] ([1.0, 0.8, 0.9] + [0.9, 0.6, 0.9]) Occupation physicist [-1.0, -0.5, 0.1] chemist [-1.0, -0.4, 0.1] scientist [-1.0, 0.4, 0.07] = 1 3 biologist [-1.0, -0.3, 0.0] ([-1.0, -0.5, 0.1] + [-1.0, -0.4, 0.1] + [-1.0, -0.3, 0.0])

7 Building Clustering System Clustering Within Attribute Keep clustering to form hierarchical structure The average height of the clustering system is about 4 layers singer director [1.0, 0.8, 0.9] [0.9, 0.6, 0.9] Clustering Clustering [0.95, 0.7, 0.9] [0.75, 0.72, 0.33] Clustering Occupation physicist [-1.0, -0.5, 0.1] [0.15, 0.1, 0.2] chemist [-1.0, -0.4, 0.1] [-1.0, 0.4, 0.07] biologist [-1.0, -0.3, 0.0] [-0.15, 0.11, 0.11] [0.01, 0.12, 0.33]

8 Building Clustering System Clustering Within Attribute Build clustering hierarchy within each attribute [1.0, 0.8, 0.9] [0.95, 0.7, 0.9] [0.75, 0.72, 0.33] [0.9, 0.6, 0.9] [0.15, 0.1, 0.2] [-1.0, -0.5, 0.1] Occupation [-1.0, -0.4, 0.1] [-1.0, 0.4, 0.07] [0.01, 0.12, 0.33] [-1.0, -0.3, 0.0] [-0.15, 0.11, 0.11] [-0.1, 0.2, 0.3] [-0.05, 0.15, 0.3] [0.0, 0.1, 0.3] Instrument [-1.1, -0.5, 0.0] [-1.0, 0.4, 0.07]

9 Building Cluster System Assign entities to clusters Assign entities to clusters according to its attribute value guitar Taylor Swift Occupation: singer instrument: guitar, vocal genre: pop, country, rock born: Pennsylvania Occupation singer director physicist biologist artist scientist instrument violin erhu piano accordion string instrument clavier

10 Building Cluster System Intersection of different clusters The intersection of different clusters is also meaningful We will keep the new cluster only when its size is large enough (contains a fair number of entities) Within attribute: Genre: country country pop Between attribute: Genre: pop pop pop singer Occupation: singer

11 Detecting missing attributes Old entities Exist in the KB when building the cluster system Already assigned clusters to them Simply apply our basic idea Taylor Swift Occupation: singer Instrument: vocal, guitar Genre: pop, country, rock Born: Pennsylvania Michael Jackson Pop singer The attribute Record Label is missing Lady Gaga Justin Bieber Robin Thicke Most of Pop Singers have the attribute Record Label Taylor Swift should also have that attribute Record Label

12 Detecting missing attributes New entities Not exist in the KB when building the cluster system Find clusters for them first Vector Similarity Match Justin Bieber Occupation: singer Instrument: vocal, guitar Genre: pop, R&B Born: London Justin Bieber Occupation: [1.1, -0.3, 0.4] Instrument: [0.1, 0.2, -0.4], [-0.4, 1.0, 0.9] Genre: [0.2, 0.3, -0.5], [0.1, 0.8, 0.2] Born: [-0.1, 0,8, 0,9] Within Occupation: singer [1.1, -0.3, 0.4] Should also belong to cluster artist, and other clusters contains cluster singer

13 Summary Represent entity as a set of vectors Taylor Swift Occupation: singer Instrument: vocal, guitar Genre: pop, country, rock Born: Pennsylvania Taylor Swift Occupation: [1.1, -0.3, 0.4] Instrument: [0.1, 0.2, -0.4], [-0.4, 1.0, 0.9] Genre: [0.2, 0.3, -0.5], [0.4, 0.8, 0.9], [0.3, 0.9, -0.8] Born: [0.1, 0,3, 0,2] Assign clusters to entities Michael Jackson Justin Bieber Pop singer Generated From Lady Gaga Robin Thicke Occupation Instrument Build clustering system within attributes AND interest these clusters [1.0, 0.8, 0.9] [0.95, 0.7, 0.9] [0.9, 0.6, 0.9] [0.15, 0.1, 0.2] [-1.0, -0.5, 0.1] [-1.0, -0.4, 0.1] [-1.0, 0.4, 0.07] [-1.0, -0.3, 0.0] [-0.15, 0.11, 0.11] [-0.1, 0.2, 0.3] [0.0, 0.1, 0.3] [-1.0, 0.4, 0.07] [-1.1, -0.5, 0.0] Detect missing attributes based on the Basic Idea The attribute record label is missing

14 Experiment Dataset 20,000 randomly sampled person entities in DBpedia Evaluation First Method: randomly delete one the attribute of an entity, see if our method can find this attribute back or not Second Method: human evaluation, see if the proposed missing attributes are reasonable or not First Method Taylor Swift Occupation: singer Instrument: vocal, guitar Genre: pop, country, rock Born: Pennsylvania Record Label: Big Machine DELETE! Can our algorithm find the attribute Record Label back? Second Method Missing Attribute Proposed for Taylor Swift Record Label Birth date Allegiance (for military people) Are these proposed missing attributes reasonable?

15 Experiment Comparison method: Z. Abedjan and F. Naumann.(2013) First evaluation method Top 5 means the algorithm proposes 5 most probable missing attributes, see if the deleted attribute is contained in The metric is precision (if the deleted attribute is contained, then a match) Top1 Top5 top10 Our Method 84.43% 95.36% 96.05% Abedjan & Naumann NA 51.00% 71.40%

16 Experiment Comparison of old entities and new entities First evaluation method May be not fair when most of the proposed missing attributes are reasonable, except that the deleted one has a lower rank Top1 Top5 top10 Old Entities 84.43% 95.36% 96.05% New Entities 33.86% 45.45% 46.07% Not Fair Case: Proposed Missing Attributes (Top 5): 1. (good proposal) 2. (good proposal) 3. (good proposal) 4. (not good) 5. (good proposal) 18. Deleted Attribute

17 Experiment Human Evaluation Randomly choose 1000 old entities and 1000 new entities Propose all the attributes with a score higher than the threshold (no more than 10) Old Entity New Entity Precision 96.72% 97.09%

18 Conclusion Performance Our method has a high precision, more than 95% suggested attributes are reasonable Our method is good enough to be used in real world Future Work Try Chinese data Combine our method with relation extraction

19 Q & A