AnalogySpace: Reducing the Dimensionality of Common Sense Knowledge Advisor: Hsin-His Chen Reporter: Chi-Hsin Yu Date: From AAAI 2008 Robert Speer, CSAIL, Massachusetts Institute of Technology Catherine Havasi, Laboratory for Linguistics and Computation, Brandeis University Henry Lieberman, Software Agents Group, MIT Media Lab
Outlines Introduction Common Sense Computing AnalogySpace Evaluation Related Work Conclusion
Introduction AnalogySpace ◦ As a component in ConceptNet3 ◦ As an dimensionality reduction (LSA) technique ◦ As an mechanism to “smooth” the interaction between the web contributors and ConceptNet system ConceptNet Users
Common Sense Computing ConceptNet ◦ Open Mind Common Sense project ◦ > 700k pieces of information ◦ > contributors ◦ > 250k assertions, 3.4% negative polarity
Common Sense Computing (cont.) CatDog a cat is a peta dog is a pet a cat has fura dog has fur a cat has a taila dog has a tail a cat has four legs ?? a dog has four legs ?? Learner System (Chklovski 2003) ◦ reasoning about commonsense by “cumulative analogy” ◦ Step Dividing statements into objects and features Hypothesizing new knowledge by analogy step
AnalogySpace Singular value decomposition, SVD ◦ Like LSA in IR ◦ A=U Σ V t, U, V: orthogonal matrix Σ : diagonal matrix Truncated SVD ◦ Row: concept ◦ Column: features
AnalogySpace (cont.) Assertion:“A trunk is part of car.” ◦ Feature (PartOf, “Car”) to the concept “trunk” ◦ Feature (“trunk”, PartOf) to the concept “car” Only included concepts that has min. 4 assertions Score and normalization ◦ Using confidence score as the value of matrix +n: for positive assertion, -n: for negative assertions, 0: for confidence score <= 0 ◦ Normalizing the rows of the matrix ◦ Scaling the rows down by Euclidean norm
AnalogySpace (cont.)
Evaluation AnalogySpace assures users that the system is learning from their input. Experiment ◦ 40 college students ◦ 60 assertions for each student ◦ 4 sources 25%: existing ConceptNet assertions (CS>2) 25%: produced by AnalogySpace 25%: from a modified AnalogySpace (Gutpa et.al., 2004) 25%: random combinations of concepts and features
Evaluation (cont.)
Assigning score (2=generaltrue,1,0,-1=not true) ◦ 1.315: existing assertions ◦ 1.025: new assertion by AnalogySpace ◦ 0.882: within-relation SVD ◦ : random
Related Work Extract general knowledge ◦ Suh, Halpin, & Klein (2006) Mining commonsense from wikipedia ◦ Eslick (2006) Uses data mining techniques to extract common sense from websites on the Internet ◦ KNEXT project (Schubert 2002) uses patterns to extract semantic relationships from the Penn Treebank
Conclusion AnalogySpace ◦ Has the ability to predict new assertions ◦ Helps to give users confidence that the system is learning from their inputs
Thanks!!
Questions: Knowledge extraction ◦ Types of commonsense knowledge ◦ Usage of commonsense knowledge in difference context Inference ◦ Used for extracting new knowledge ◦ Used for solving specific tasks