Chapter 6. Inference beyond the index 2007 년 1 월 30 일 부산대학교 인공지능연구실 김민호 Text : FINDING OUT ABOUT Page. 182 ~ 251
Introduce(1/4) - Knowledge representation AI is primary contribution to computer science! Related to: Programming language’s “abstract data types” Database (logical!) modeling -eg, ‘ontology’ building
Introduce(2/4) - Traditionally (GOFAI) serving deductive goals Valid inference Man (x) -> Mortal(x) Man(Socrates) Mortal(Socrates) Expressiveness Even first-order logic offers tradeoffs wrt/propositional
Introduce(3/4) - Machine learning: inductive sources of knowledge Data-mining Statistical analysis of large datasets Searching for patterns Inferring semantics (meaning) from syntactic cues from word statistics from bibliographic citations Even from capitalization -Proper names → Global reference!
Introduce(4/4) - Exploiting other (non-index) information
Subsection 6.1 Citation: Interdocument Links 6.2 Hypertext, Intradocument Links 6.3 Keyword Structures 6.4 Social Relations among Authors 6.5 Modes of Inference 6.6 Deep Interfaces 6.7 FOA(The Law) 6.8 FOA(Evolution) 6.9 Text-Based Inteligence
6.1 Citation: Interdocument Links Citation is a pointer, from a document to a document. how accurately do we know the location of the citation in the citing paper? how precisely is its pointer into the cited paper?
6.1 Citation: Interdocument Links Document similarity based on shared bibliographies Coupling Overlap between two document’s bibliographis Co-citation Degree to which two documents are both referenced by other document’s bibliographies
6.1 Citation: Interdocument Links
Common law depends on rule of precedence Stare decisis Prior decisions applied to new factual situations Hierarchical local jurisdictions limit interpretation Dialectic debate (rationale, justice, change, etc.) NB: Same corpus used by both adversaries References to history of O(10 year)
6.1 Citation: Interdocument Links Unambiguous
6.1 Citation: Interdocument Links Eigen-structure of citation graphs Authority: analogous to bibliometric ‘impact’ Hubs: Pull together authorities Citation-expanded hitlist
Summary Writings do not exist in isolation Author explicit references to other’s documents provides excellent evidence concerning the ARGUMENTS they each make
Google’s Page rank Simulate stationary distribution of Markov process with incremental update of page weight
Hierarchic structure Visualizing references
Pedagogical structure Prerequisite lattice Reading-level analysis – against well – tested vocabularies Level of coverage
Argument relationships
thesaurus BT/NT/RT relations aot / AI “ontologies”
WordNet
Classification taxonomies Institutionalized Myopic discipline focus
Neural networks - basics Query Retrieval Relevance Feedback
Construction of initial NNet
Query: “Case-based approach to the law” Morphological processing of tokens High-frequency “noise” words elided
SAS in IR Initial query may refer to many “features” Descriptive keywords are only one type Retrieval becomes a process of completion
Type of relation less important than fact of association
Initial retrieval
Most highly ranked document
Goal document
3rd-order transitive associations
4th-order transitive associations
Swanson's Arrowsmith