Why Study IR? Many reasons, but if you want a one-word answer:
UIT, VNU-HCM Social Media Analysis 3
Google … ❖ Examines billions of web pages ❖ Returns results in less than half a second ❖ Valued at gazillions of dollars by the public market
UIT, VNU-HCM Social Media Analysis 4
How Does Google Work? ❖ Only Google know, but . . . ❖ Uses hundreds of thousands of machines ❖ Uses some sophisticated computer science (efficient storage and searching of large datasets) ❖ Uses an innovative ranking algorithm (based on the hypertext structure of the web)
UIT, VNU-HCM Social Media Analysis 5
How Does Google Work? ❖ Underlying Google is basic IR technology. ❖ The Web is indexed. ▪ an index links terms with pages. ❖ A user’s information need is represented as a query. ❖ Queries are matched against web pages. ▪ Google attempts to return pages which are relevant to the information need.
UIT, VNU-HCM Social Media Analysis 6
IR is More Than Web Search ❖ IR is much older than the Web (1950s –) ❖ The Web has some unique characteristics which make it a special case. ❖ IR deals with tasks other than searching: ▪ Categorizing Documents. ▪ Summarizing Documents. ▪ Answering Questions. ▪ …
UIT, VNU-HCM Social Media Analysis 7
Motivation for IR ❖ Searching literature databases. ❖ Web search. ❖ Volume of information stored electronically is growing at ever faster rates ▪ need to search it ▪ categorize it ▪ filter it ▪ translate it ▪ summarize it ▪ ... UIT, VNU-HCM Social Media Analysis 8 Biomedical Information ❖ Biomedical literature is growing at a startling rate ▪ Around 1,000,000 new articles are added to Medline each year ❖ Tasks: ▪ Literature search ▪ Creation and maintenance of biological databases ▪ Knowledge discovery from text mining
UIT, VNU-HCM Social Media Analysis 9
Document Retrieval IR is often used to mean Document Retrieval Primary task of an IR system: retrieve documents with content that is relevant to a user’s information need How do we represent content? How do we represent information need? How do we decide on relevance?
UIT, VNU-HCM Social Media Analysis 10
Document Retrieval Representation/Indexing ◼ Representation of documents and requests: ▪ bag of words? ▪ stop words, upper/lower case, . . . ∗ ▪ query language ◼ Storing the documents, building the index Searching ◼ Is a document relevant to the query? ▪ models of IR: Boolean, vector-space, probabilistic ◼ Efficient algorithms for searching large datasets
UIT, VNU-HCM Social Media Analysis 11
What IR is Not An IR system is not a Database Management System A DBMS stores and processes well-defined data A search in a DBMS is exact / deterministic Search in an IR system is probabilistic ◼ inherent uncertainty exists at all stages of IR: information need, formulating query, searching
UIT, VNU-HCM Social Media Analysis 12
A Simple Retrieval Model Bag of Words approach ◼ A document is represented as a bag of words – Word order is ignored ◼ Syntactic structure is ignored ◼ ... Relevance is determined by comparing the words in the document with the words in a query Simple approach has been very effective UIT, VNU-HCM Social Media Analysis 13 Vector Space Model Provides a ranking of documents with respect to a query Documents and queries are vectors in a multi- dimensional information space Key questions: ◼ What forms the dimensions of the space? ∗ terms, concepts, . .. ◼ How are document and query vectors compared?
UIT, VNU-HCM Social Media Analysis 14
Coordinate Matching Document relevance measured by the number of query terms appearing in a document Terms provide the dimensions ◼ Large vocabulary ⇒ high dimensional space Similarity measure is the dot-product of the query and document vectors
UIT, VNU-HCM Social Media Analysis 15
Simple Example Term vocabulary: (England, Australia, Pietersen, Hoggard, run, wicket, catch, century, collapse) Documents: ◼ d1: Australia collapse as Hoggard takes 6 wickets ◼ d2: Pietersen’s century puts Australia on back foot Queries: ◼ q1: {Hoggard, Australia, wickets} Query, document similarity ◼ q1.d1 = (0,1,0,1,0,1,0,0,0) · (0,1,0,1,0,1,0,0,1) = 3 ◼ q1.d2 = (0,1,0,1,0,1,0,0,0) · (0,1,1,0,0,0,0,1,0) = 1
UIT, VNU-HCM Social Media Analysis 16
Term Frequency (TF) Coordinate matching does not consider the frequency of query terms in documents Term vocabulary: (England, Australia, Pietersen, Hoggard, run, wicket, catch, century, collapse) d1: Australia collapsed as Hoggard took 6 wickets. Flintoff praised Hoggard for his excellent line and length. q1: {Hoggard, Australia, wickets} q1 · d1 = (0,1,0,1,0,1,0,0,0) · (0,1,0,2,0,1,0,0,1) = 4
UIT, VNU-HCM Social Media Analysis 17
Term Frequency (TF) Coordinate matching does not consider the number of documents query terms appear in Term vocabulary: (England, Australia, Pietersen, Hoggard, run, wicket, catch, century, collapse) d2: Flintoff took the wicket of Australia’s Ponting, to give him 2 wickets for the innings and 5 wickets for the match. q1: {Hoggard, Australia, wickets} q1 · d2 = (0,1,0,1,0,1,0,0,0) · (0,1,0,0,0,3,0,0,0) = 4
UIT, VNU-HCM Social Media Analysis 18
Inverse Document Frequency (IDF) Assume wicket appears in 100 documents in total, Hoggard appears in 5, and Australia in 10 (ignoring IDF of other terms) d1: Australia collapsed as Hoggard took 6 wickets. Flintoff praised Hoggard for his excellent line and length. d2: Flintoff took the wicket of Australia’s Ponting, to give him 2 wickets for the innings and 5 wickets for the match. q1: {Hoggard, Australia, wickets} q1 · d1 = (0,1,0,1,0,1,0,0,0) · (0,1/10,0,2/5,0,1/100,0,0,1) = 0.411 q1 · d2 = (0,1,0,1,0,1,0,0,0) · (0,1/10,0,0/5,0,3/100,0,0,0) = 0.13
UIT, VNU-HCM Social Media Analysis 19
Document Length
UIT, VNU-HCM Social Media Analysis 20
Vector Space Similarity
UIT, VNU-HCM Social Media Analysis 21
Vector Space Similarity
UIT, VNU-HCM Social Media Analysis 22
Language Understanding? Want a system which “understands” documents and query and matches them? ◼ use semantic representation and logical inference Until recently such technology was not robust / did not scale to large unrestricted text collections But: ◼ useful for restricted domains ◼ now used for some large-scale tasks (QA, IE) Is a “deep” approach appropriate for document retrieval? ◼ Powerset (Natural Language Search) think so (see www.powerset.com)
UIT, VNU-HCM Social Media Analysis 24
Tasks in IR (broadly conceived) Document Retrieval (ad-hoc retrieval) Document Filtering or Routing Document Categorization Document Summarizing Information Extraction Question Answering
UIT, VNU-HCM Social Media Analysis 25
Other Topics Multimedia IR (images, sound, . . .) ◼ but text can be of different types (web pages, e- mails, . . .) User-system interaction (HCI) Browsing
UIT, VNU-HCM Social Media Analysis 26
Evaluation IR has largely been treated as an empirical, or engineering, task Evaluation has played an important role in the development of IR DARPA/NIST Text Retrieval Conference (TREC) ◼ began in 1992 ◼ has many participants ◼ uses large text databases ◼ considers many tasks in addition to document retrieval
UIT, VNU-HCM Social Media Analysis 27
IR in One Sentence “Indexing, retrieving and organizing text by probabilistic or statistical techniques that reflect semantics without actually understanding” (James Allan, Umass)
UIT, VNU-HCM Social Media Analysis 28
Brief History of IR 1960s ◼ development of basic techniques in automated indexing and searching 1970s ◼ Development of statistical methods / vector space models ◼ Split from NLP/AI ◼ Operational Boolean systems 1980s ◼ Increased computing power ◼ Spread of operational systems
UIT, VNU-HCM Social Media Analysis 29
Brief History of IR 1990s and 2000s Large-scale full text IR systems for retrieval and filtering Dominance of statistical ranking approaches Web search Multimedia and multilingual applications Question Answering TREC evaluations
UIT, VNU-HCM Social Media Analysis 30
Brief History of IR Course book ◼ Introduction to Information Retrieval Manning, Raghavan, Schu ̈tze http://nlp.stanford.edu/IR- book/html/htmledition/irbook.html
