Presentation on theme: "CILC2011 A framework for structured knowledge extraction and representation from natural language via deep sentence analysis Stefania Costantini Niva Florio."— Presentation transcript:
CILC2011 A framework for structured knowledge extraction and representation from natural language via deep sentence analysis Stefania Costantini Niva Florio Alessio Paolucci Università Degli Studi Dell’Aquila
“overcome the knowledge acquisition bottleneck” Motivation
Structured data from plain text The more interesting one: Ontology population (Semantic Web) …but endless possibilities!!!
Our Proposal Our framework allows us to: Extract knowledge from natural language sentences using a deep analysis technique based on linguistic dependencies and phrase syntactic structure. Use OOLOT (Ontology Oriented Language of Thought) an intermediate language based on ASP (Answer Set Programming), specifically designed for the representation of the distinctive features of the knowledge extracted from natural language. Easily Integrate our framework in the context of the Semantic Web. OOLOT lets us exploit the non monotonic reasoning (through ASP) to deal with common sense reasoning and other typical aspects of the knowledge encoded through the Natural Language.
Parsing Syntactic Parsing: It can determine the syntactic structure of a sentence Chomsky’s constituent analysis It builds up the elements in their hierarchical order Syntactic parsers decompose a text into tokens and attribute them their grammatical function Statistical Parsing: It is based on a corpus of training annotated data It gathers information about the frequency with which the elements are needed in specific contexts Only statistic may be not enough to determine when to split a symbol in sub- symbols Probabilistic Context Free Grammar (PCFG): More than one production rule may apply to a sequence of words, thus resulting in a conflict It uses the frequency of various productions to order them
Parsing Statistical parsing is useful to solve problems like ambiguity and efficiency We lose part of the semantic information BUT Dependency Grammar: words in a sentence are connected by means of binary, asymmetrical governor-dependent relationships
Context Disambiguation Given a (finite) set of contexts, assign each lexical item to one (or more) context(s) including a score. Context_1Context_2Context… Context_m Lexical Item 0.7 0.3 We use a simple, frequency-based, disambiguation algorithm.
Resolution Car Each lexical item (a word, or a set of), is resolved against popular ontologies, including DBPedia, YAGO, GeoNames, WordNet 3 OWL, …
OOLOT The language of thought is an intermediate format mainly inspired by Kowalski’s LoT. It has been introduced to represent the extracted knowledge in a way that is totally independent from original lexical items and, therefore, from original language. Our LOT is itself a language, but its lexicon is ontology oriented, so we adopted the acronym OOLOT (Ontology Oriented Language Of Thought). OOLOT is used to represent the knowledge extracted from natural language sentences, so basically the bricks of OOLOT (lexicons) are ontological identifier related to concepts (in the ontology), and they are not a translation at lexical level.
And, finally, after applying apple to the previous partial expression, we have:
RDF/OWL Exporting Since OOLOT is designed to have a representation very close to RDF, it's possible to export toward RDF/OWL. In many cases, when is possible to maintain the semantic, there is a 1:1 mapping, otherwise we're starting using RDF/OWL syntactic approximations through reification (when you can’t preserve the original semantic) OOLOT: predicate(subject, object) RDF: Best case:
Framework In Action “Ferrari is an Italian sports car manufacturer based in Maranello.”
Conclusion & Future Works OOLOT Deep Analysis RDF Exporting Dependency Parsing ASP Further exploit: OOLOT language ASP to RDF/OWL Exporting This is a quite new framework, so many aspects need to be refined and improved.