Ontology Learning and Population from Text: Algorithms, Evaluation and Applications Chapters 1 - 5 Presented by Sole
Introduction Artificial intelligence Build systems that incorporate knowledge about a domain to reason on the basis of this knowledge and solve problems not encountered before Include explicit and symbolic representation of knowledge about a domain Symbolic representation and procedural aspects are separated so that it can be reused across systems Example for Yoyo the cat Which symbols to use and what they stand for?
Introduction Ontology Defines what is important in a domain and how concepts are related Knowledge-based system: determine which symbols are needed and how they are interpreted Logical level: interpretation can be constraint according to the ontology by axiomatizing symbols Issues Costly to construct Time-consuming Significant coverage of domain is needed Meaning and consistent generalization are required The trade-off between modeling a large amount of knowledge and providing as many abstractions as possible to keep the model concise makes ontology engineering indeed a challenging Knowledge Acquisition Bottleneck
Introduction Solution Automatically learn ontologies from data Goal: bridging the gap between World of symbols (words used in natural language) World of concepts (abstractions of human thought) Challenge Correctness and consistency of the model can not be guaranteed Human post-processing definitely necessary Automatically learned ontologies need to be inspected, validated, and modified by humans before they can be applied for applications relying on logical reasoning Rest of book will be about explaining in more details ontology learning
Ontologies Definition Philosophical discipline Computer Science Science of existence or the study of being Computer Science Formal specifications of a conceptualization Resources representing the conceptual model underlying a certain domain, describing it in a declarative fashion and thus cleanly separating it from procedural aspects
Ontologies Example
Learning from Text Ontology learning Acquire a domain model from data Lifting : XML-DTDs, UML diagrams, databases Semi-structured sources: HTML, XML Unstructured sources: ontology learning from text The task is inherently complex and challenging mainly due to two reasons. * There is typically only a small part of the authors' domain knowledge involved in the creation process, such that the process of reverse engineering can, at best, only partially reconstruct the authors' model. *World knowledge – unless we are considering a text book or dictionary - is rarely mentioned explicitly.
Learning from Text Meaning triangle Every language has symbols that evoke a concept that refers to a concrete individual in the world
Learning from Text Ontology population Learning concepts and relations Knowledge markup or annotation: select text fragments and assign them to an ontological concept Applications Several methods have been developed in recent years Challenge No consensus within ontology learning community on concrete tasks for ontology learning Comparison between approaches is difficult
Learning from Text Ontology learning tasks (layer cake)
Learning from Text Terms: Task: find a set of relevant concepts and relations E.g., words, multi-word compounds State-of-the-art IR methods NLP methods: POS tagger, statistical approaches Linguistic realization of domain specific concepts (keywords in a domain) Input: text Output: concepts
Learning from Text Synonyms: Task: find words which denote the same concept E.g., synsets on WordNet State-of-the-art Semantically-similar words Sense disambiguation and synonym discovery Latent Semantic Indexing (LSI) Statistical information measures defined over the Web to detect synonyms
Learning from Text Concepts: Task: find intentional definitions of concept, their extension, and lexical signs used to refer to them State-of-the-art Clusters of related terms LSI-based techniques Discovery of hierarchies of named entities Know-it-all system OntoLearn system Intension: natural language description of concept The Know-It-All system [Etzioni et al., 2004a] also aims at learning the extension of given concepts, such as, for example, all the actors appearing on the Web. In the approach of Evans [Evans, 2003], the concepts as well as their extensions are thus derived automatically, while Etzioni et al. [Etzioni et al., 2004a] essentially learn the extension of existing concepts. Finally, other systems learn concepts intentionally. The OntoLearn system [Velardi et al., 2005], for example, derives WordNet-like glosses for domain specific concepts on the basis of a compositional interpretation of the meaning of compounds.
Learning from Text Hierarchies: Task: concept hierarchy induction, refinement and lexical extension State-of-the-art Lexico-syntactic patterns Clustering algorithm to automatically derive concept hierarchies Analysis of term co-occurrence in same sentence/document
Learning from Text Relations: Task: learn relations identifiers or labels as well as their appropriate domain and range State-of-the-art Association rules Syntactic-dependencies Very few approaches address the issue of learning ontology relations from text
Learning from Text Axiom schemata instantiations: General axioms Task: learn which concepts, relations, or pair of concepts the axioms in a given system apply to General axioms Task: derive more complex relationships and connections between concepts and relations Logical interpretations constraining the interpretation of concepts and relations
Learning from Text Population: Task: learn instances of concepts and relations State-of-the-art Associated to well-known tasks for which a variety of approaches have been developed Information extraction Named entity recognition
Basics Natural Language Processing Basic formalisms and techniques necessary for understanding the rest of the book NLP-> “green car” green is the color value to describe car Co-references John Adams J. Adams
Syntactic analysis: parsing Basics NLP Pre-processing steps Chunking, also called shallow or partial parsing, applies shallow processing techniques (typically regular expressions and finite automata) to group together words to larger syntactic and meaning-bearing constituents, typically with a head which is modified by other words in the unit. Chunking Syntactic analysis: parsing
Syntactic dependencies Basics NLP Pre-processing The museum houses an impressive collection of medieval and modern art. The building combines geometric abstraction with classical references that allude to the Roman influence on the region. Bank River Financial Institution Contextual features Syntactic dependencies
Basics Similarity measures NLP Context is often represented as vector in high dimensional space E", the dimensions corresponding to words found in the context of the word in question. This vector-based context representation constitutes the core of the so called vector space model used in information retrieval
Basics Similarity measures Binary similarity measures NLP Similarity measures Binary similarity measures Geometric similarity measures
Basics Similarity measures Measures based on probability distribution NLP Similarity measures Measures based on probability distribution Hypothesis testing When assessing the degree of association between words, the HQ hypothesis assumes that the probability of the two words is independent of each other, i.e. Piwi,W2) = Piwi) P{w2) The independence hypothesis is rejected in case the observed probability is found to significantly differ from P{wi,W2) as defined above.
Basics Term relevance Weight the importance of a term in a document NLP Term relevance Weight the importance of a term in a document
Basics NLP WordNet Lexical database for the English language
Basics Formal concept analysis Formal objects: concepts + Formal attributes: characteristics describing objects Incidence relation: information about which attributes hold for each object = Formal context
Basics FCA Example
Basics FCA Example Mention that there exist algorithms that can be used to create FC
Basics Machine learning Automatic recognition/detection of patterns and regularities within sample data Patterns can be used to understand/describe the data or to make predictions Learning process Supervised Predicts the appropriate category for an example from a set of categories represented by a set of labels Unsupervised Search for common and frequent structures within the data (data exploration)
Basics Supervised learning Regression Classification ML Numeric prediction (labels are continue values) Classification Assign proper category to a given example Target value Feature vector
Basics Classifiers Tools Bayesian Classifiers Decision Trees ML Classifiers Bayesian Classifiers Decision Trees Instance-Based Learning Support Vector Machines Artificial Neural Networks Tools WEKA RapidMiner
Basics ML Examples
Basics Unsupervised learning ML Unsupervised learning Clustering: find groups of similar objects in data There is no labeled data to train from Classification Hierarchical vs. non-hierarchical Non-hierarchical algorithms produce a set of groups Hierarchical algorithms order groups in a tree structure Hard vs. soft Hard: elements are assigned to distinct clusters Soft: elements are assigned to clusters with a certain degree of membership
Basics Algorithms K-means Hierarchical clustering ML Algorithms K-means Hierarchical clustering Hierarchical Agglomerative (Bottom-Up) Clustering Divisive (Top-Down) Clustering
Datasets Corpus description
Datasets Concept hierarchies