CPECSC 580 Knowledge Management Dr Franz J Kurfess

Course Overview u u Introduction Knowledge Processing u u Knowledge Organization u u u

Overview Knowledge Retrieval u u u Motivation Objectives Finding Out About u u Data

Logistics u Term Project u APIs u Lab and Homework Assignments u Deadline HW

Finding Out About © 2001 -2005 Franz J. Kurfess [Belew 2000] Knowledge Retrieval 5

Pre-Test © 2001 -2005 Franz J. Kurfess Knowledge Retrieval 6

Motivation © 2001 -2005 Franz J. Kurfess Knowledge Retrieval 7

Objectives © 2001 -2005 Franz J. Kurfess Knowledge Retrieval 8

Finding Out About u Keywords u Queries u Documents u Indexing © 2001 -2005

Keywords u linguistic atoms used to characterize the subject or content of a document

Queries u formulated u natural language u u interaction with human information providers artificial

Documents u general u any u interpretation document that can be represented digitally text,

Aboutness of Documents u describes the suitability of a document as answer to a

Structural Aspects of Documents u documents may be composed of documents u paragraphs, subsections,

Document Proxies u surrogates u abridged u for the real document representations catalog, abstract

Indexing ua vocabulary of keywords is assigned to all documents of a corpus u

FOA Conversation Loop © 2001 -2005 Franz J. Kurfess [Belew 2000] Knowledge Retrieval 18

Data Retrieval u access to specific data items u access via address, field, name

Information Retrieval (IR) u access u also to documents referred to as document retrieval

Diagram Search Engine © 2001 -2005 Franz J. Kurfess [Belew 2000] Knowledge Retrieval 21

Parsing u extraction u mostly of lexical features from documents words u may require

Parsing Tools u Montytagger python and Java u u fn. TBL u (C++) http:

Matching Against Indices u identification of documents that are relevant for a particular query

Retrieved and Relevant Documents recall precision © 2001 -2005 Franz J. Kurfess |retrieved relevant|

Specificity vs. Exhaustivity © 2001 -2005 Franz J. Kurfess [Belew 2000] Knowledge Retrieval 26

Vector Space u interpretation u relates u can of the index matrix documents and

Vector Space Diagram © 2001 -2005 Franz J. Kurfess [Belew 2000] Knowledge Retrieval 28

Document Retrieval © 2001 -2005 Franz J. Kurfess [Belew 2000] Knowledge Retrieval 29

Retrieval Assessment u subjective assessment u how well do the retrieved documents satisfy the

Retrieval Assessment Diagram © 2001 -2005 Franz J. Kurfess [Belew 2000] Knowledge Retrieval 31

Relevance Feedback u subjective assessment of retrieval results u often used to iteratively improve

Relevance Feedback in Vector Space u relevance feedback is used to move the query

Query Session Example © 2001 -2005 Franz J. Kurfess [Belew 2000] Knowledge Retrieval 34

Consensual Relevance u relevance feedback from multiple users u identifies documents that many users

IR Diagram Index Query Documents Term 2 Term 3 Term 4 Keywords Term 1

Query Term 3 Term 2 Term 4 Term 1 Corpus Documents Doc. 5 Doc.

Knowledge Retrieval u Context u Usage © 2001 -2005 Franz J. Kurfess Knowledge Retrieval

Context in Knowledge Retrieval u in addition to keywords, relationships between keywords and documents

Inference beyond the Index u determines relationships between documents u citations are explicit references

Additional Information Sources © 2001 -2005 Franz J. Kurfess [Belew 2000, after Kochen 1975]

Hypertext u inter-document links provide explicit relationships between documents u can be used to

Adaptive Retrieval Techniques u fine-tuning the matching between queries and retrieved documents u learning

Document Classification © 2001 -2005 Franz J. Kurfess Knowledge Retrieval 48

Query Model u query types (templates) u frequently v used types of queries e.

Terminology Model u individual terms are connected to related terms u thesaurus/ontology v synonyms,

Matching u categorizer u determines the categories to be selected for the grouping of

Results u retrieved documents are grouped into hierarchically arranged categories meaningful for the user

Dyna. Cat u knowledge-based approach to the organization of search results u categorizes results

Dyan. Cat Results © 2001 -2005 Franz J. Kurfess [Dyna. Cat, 2000] Knowledge Retrieval

Dyna. Cat Query Types © 2001 -2005 Franz J. Kurfess [Dyna. Cat, 2000] Knowledge

Dyna. Cat Search © 2001 -2005 Franz J. Kurfess [Dyna. Cat, 2000] Knowledge Retrieval

Information vs. Knowledge Retrieval u IR u u u KR keywords as main components

KR Diagram © 2001 -2005 Franz J. Kurfess Knowledge Retrieval 58

Knowledge Discovery u Data Mining u Rule Extraction © 2001 -2005 Franz J. Kurfess

Post-Test © 2001 -2005 Franz J. Kurfess Knowledge Retrieval 60

Important Concepts and Terms u u agent information retrieval knowledge representation knowledge retrieval ©

Summary Knowledge Retrieval © 2001 -2005 Franz J. Kurfess Knowledge Retrieval 63

© 2001 -2005 Franz J. Kurfess Knowledge Retrieval 64

Slides: 62

Download presentation

Course Overview u u Introduction Knowledge Processing u u Knowledge Organization u u u Classification, Categorization Ontologies, Taxonomies, Thesauri Knowledge Retrieval u u u Knowledge Acquisition, Representation and Manipulation Information Retrieval Knowledge Navigation Knowledge Presentation u Knowledge Visualization © 2001 -2005 Franz J. Kurfess u Knowledge Capture, Transfer, and Distribution u Usage u of Knowledge Access Patterns, User Feedback u Knowledge Techniques u Exchange Management Topic Maps, Agents u Knowledge Management Tools u Knowledge Management in Organizations Knowledge Retrieval 2

Overview Knowledge Retrieval u u u Motivation Objectives Finding Out About u u Data Retrieval u u Keywords and Queries Documents Indexing Access via Address, Field, Name Information Retrieval u u Access via Content (Values) Parsing Matching Against Indices Retrieval Assessment © 2001 -2005 Franz J. Kurfess u Knowledge u u Access via Structure Meaning Context Usage u Knowledge u u Retrieval Discovery Data Mining Rule Extraction u Important Concepts and Terms u Chapter Summary Knowledge Retrieval 3

Keywords u linguistic atoms used to characterize the subject or content of a document u words u pieces of words (stems) u phrases u provide the basis for a match between u the user’s characterization of information need u the contents of the document u problems u ambiguity u choice of keywords © 2001 -2005 Franz J. Kurfess [Belew 2000] Knowledge Retrieval 11

Queries u formulated u natural language u u interaction with human information providers artificial language u interaction with computers u u especially search engines vocabulary u controlled u u limited set of keywords may be used uncontrolled u u in a query language any keywords may be used syntax u u often Boolean operators (AND, OR) sometimes regular expressions © 2001 -2005 Franz J. Kurfess [Belew 2000] Knowledge Retrieval 12

Documents u general u any u interpretation document that can be represented digitally text, image, music, video, program, etc. u practical interpretation u passage of text strings of characters in an alphabet u written natural language u length may vary u u longer documents may be composed of shorter ones © 2001 -2005 Franz J. Kurfess Knowledge Retrieval 13

Aboutness of Documents u describes the suitability of a document as answer to a query u assumptions u all documents have equal aboutness the probability of any document in a corpus to be considered relevant is equal for all documents u simplistic; not valid in reality u ua paragraph is the smallest unit of text with appreciable aboutness © 2001 -2005 Franz J. Kurfess [Belew 2000] Knowledge Retrieval 14

Structural Aspects of Documents u documents may be composed of documents u paragraphs, subsections, chapters, parts u footnotes, references u documents may contain meta-data u information about the document u not part of the content of the document itself u may be used for organization and retrieval purposes u can be abused by creators u usually to increase the perceived relevance © 2001 -2005 Franz J. Kurfess Knowledge Retrieval 15

Document Proxies u surrogates u abridged u for the real document representations catalog, abstract u pointers u bibliographical citation, URL u different media microfiches u digital representations u © 2001 -2005 Franz J. Kurfess Knowledge Retrieval 16

Indexing ua vocabulary of keywords is assigned to all documents of a corpus u an index maps each document doci to the set of keywords {kwj} it is about Index : doci about {kwj} Index-1 : {kwj} describes doci u indexing of a document / corpus u manual: humans select appropriate keywords u automatic: a computer program selects the keywords u building the index relation between documents and sets of keywords is critical for information retrieval © 2001 -2005 Franz J. Kurfess [Belew 2000] Knowledge Retrieval 17

Data Retrieval u access to specific data items u access via address, field, name u typically used in data bases u user asks for items with specific features u absence or presence of features u values u system u no returns data items irrelevant items u deterministic © 2001 -2005 Franz J. Kurfess retrieval method Knowledge Retrieval 19

Information Retrieval (IR) u access u also to documents referred to as document retrieval u access via keywords u IR aspects u parsing u matching against indices u retrieval assessment © 2001 -2005 Franz J. Kurfess Knowledge Retrieval 20

Parsing u extraction u mostly of lexical features from documents words u may require some manipulation of the extracted features u e. g. stemming of words u used as the basis for automatic compilation of indices © 2001 -2005 Franz J. Kurfess [Belew 2000] Knowledge Retrieval 22

Parsing Tools u Montytagger python and Java u u fn. TBL u (C++) http: //nlp. cs. jhu. edu/~rflorian/fntbl/ fast u Brill u http: //web. media. mit. edu/~hugo/montytagger/ Tagger (C) http: //www. cs. jhu. edu/~brill/ the original; influenced several later ones u out there is the Natural Language Toolkit: http: //nltk. sourceforge. net/ u good starting point for basics of NLP algorithms [recommended by Dan Miller, who is working on his Master’s thesis on automated generation of abstracts] © 2001 -2005 Franz J. Kurfess Knowledge Retrieval 23

Matching Against Indices u identification of documents that are relevant for a particular query u keywords of the query are compared against the keywords that appear in the document u either in the data or meta-data of the document u in addition to queries, other features of documents may be used u descriptive v features provided by the author or cataloger usually meta-data u derived features computed from the contents of the document © 2001 -2005 Franz J. Kurfess [Belew 2000] Knowledge Retrieval 24

Vector Space u interpretation u relates u can of the index matrix documents and keywords grow extremely large u binary matrix of 100, 000 words * 1, 000 documents u sparsely populated: most entries will be 0 u can be used to determine similarity of documents u overlap in keywords u proximity in the (virtual) vector space u associative memories can be used as hardware implementation u extremely fast, but expensive to build © 2001 -2005 Franz J. Kurfess [Belew 2000] Knowledge Retrieval 27

Retrieval Assessment u subjective assessment u how well do the retrieved documents satisfy the request of the user u objective assessment u idealized omniscient expert determines the quality of the response © 2001 -2005 Franz J. Kurfess [Belew 2000] Knowledge Retrieval 30

Relevance Feedback u subjective assessment of retrieval results u often used to iteratively improve retrieval results u may be collected by the retrieval system for statistical evaluation u can be viewed as a variant of object recognition u the object to be recognized is the prototypical document the user is looking for v this document may or may not exist u the difference between the retrieved document(s) and the idealized prototype indicates the quality of the retrieval results © 2001 -2005 Franz J. Kurfess [Belew 2000] Knowledge Retrieval 32

Relevance Feedback in Vector Space u relevance feedback is used to move the query towards the cluster of positive documents u moving away from bad documents does not necessarily improve the results u it can also be used as a filter for a constant stream of documents u as in news channels or similar situations © 2001 -2005 Franz J. Kurfess [Belew 2000] Knowledge Retrieval 33

Consensual Relevance u relevance feedback from multiple users u identifies documents that many users found useful or interesting u used by some Web sites u related to collaborative filtering u can also be used as an evaluation method for search engines v performance criteria must be carefully considered v precision and recall, plus many others © 2001 -2005 Franz J. Kurfess [Belew 2000] Knowledge Retrieval 35

Query Term 3 Term 2 Term 4 Term 1 Corpus Documents Doc. 5 Doc. 5 Doc. 4 Doc. 4 Doc. 3 Doc. 3 Doc. 2 Doc. 2 Doc. 1 Doc. 1 Keywords KR Diagram Index Term A Term B Term C Term E Term D Ontology Term F Term H © 2001 -2005 Franz J. Kurfess Term I Term J Term K Term L Term G Term M Knowledge Retrieval 41

Context in Knowledge Retrieval u in addition to keywords, relationships between keywords and documents are exploited u explicit v hypertext u related v links concepts thesaurus, ontology u proximity spatial: place, directory v temporal: creation date/time v u intermediate relations author/creator v organization v project v © 2001 -2005 Franz J. Kurfess Knowledge Retrieval 43

Inference beyond the Index u determines relationships between documents u citations are explicit references to relevant documents u bibliographic references u legal citations u hypertext u examples u NEC Cite. Seer <http: //citeseer. nj. nec. com> u Google Scholar http: //scholar. google. com © 2001 -2005 Franz J. Kurfess Knowledge Retrieval 44

Hypertext u inter-document links provide explicit relationships between documents u can be used to determine the relevance of a document for a query u example: Google <http: //www. google. com> u intra-document links may offer additional context information for some terms u footnotes, glossaries, related terms © 2001 -2005 Franz J. Kurfess Knowledge Retrieval 46

Adaptive Retrieval Techniques u fine-tuning the matching between queries and retrieved documents u learning of relationships between terms training with term pairs (thesaurus) v pattern detection in past queries v automatic grouping of documents according to common features v u clustering of similar documents pre-defined categories v metadata v overlap in keywords v consensual relevance v source v © 2001 -2005 Franz J. Kurfess Knowledge Retrieval 47

Query Model u query types (templates) u frequently v used types of queries e. g. problem/solution, symptoms/diagnosis, problem/further checks, . . . u category types u abstractions of query types u used to determine categories or topics for the grouping of search results u context information u current working document/directory u previous queries © 2001 -2005 Franz J. Kurfess [Pratt, Hearst, Fagan 2000] Knowledge Retrieval 49

Terminology Model u individual terms are connected to related terms u thesaurus/ontology v synonyms, super-/sub-classes, related terms u identifies labels for the category types © 2001 -2005 Franz J. Kurfess [Pratt, Hearst, Fagan 2000] Knowledge Retrieval 50

Matching u categorizer u determines the categories to be selected for the grouping of results u assigns retrieved documents to the categories u organizer u arranges v categories into a hierarchy should be balanced and easy to browse by the user u depends on the distribution of the search results © 2001 -2005 Franz J. Kurfess [Pratt, Hearst, Fagan 2000] Knowledge Retrieval 51

Results u retrieved documents are grouped into hierarchically arranged categories meaningful for the user u the categories are related to the query u the categories are related to each other u all categories have similar size v not always achievable due to the distribution of documents u reduced search times u higher user satisfaction © 2001 -2005 Franz J. Kurfess [Pratt, Hearst, Fagan 2000] Knowledge Retrieval 52

Dyna. Cat u knowledge-based approach to the organization of search results u categorizes results into meaningful groups that correspond to the user’s query u uses knowledge of query types and of the domain terminology to generate hierarchical categories u applied to the domain of medicine u MEDLINE is an on-line repository of medical abstracts 9. 2 million bibliographic entries from 3800 journals v Pub. Med is a web-based search tool v v v returns titles as an relevance-ranked list links to “related articles” © 2001 -2005 Franz J. Kurfess [Dyna. Cat, 2000] Knowledge Retrieval 53

Information vs. Knowledge Retrieval u IR u u u KR keywords as main components of the query index as match-making facility statistical basis for selection of relevant documents (ordered) list of results u u © 2001 -2005 Franz J. Kurfess keywords plus context information for the query index plus ontology for matching query and documents relationships between keywords and documents influence the selection of relevant documents results are grouped into meaningful categories Knowledge Retrieval 57