IMPLEMENTATION OFINFORMATION RETRIEVAL SYSTEMS VIA RDBMS
Relational Database: Definitions Relational database: a set of relations Relation: made up of 2 parts: Instance : a table, with rows and columns. #Rows = cardinality, #fields = degree / arity. Schema : specifies name of relation, plus name and type of each column. E.G. Students(sid: string, name: string, login: string, age: integer, gpa: real). Can think of a relation as a set of rows or tuples (i.e., all rows are distinct).
Example Instance of Students Relation sid name login age gpa 53666 Jones jones@cs 18 3.4 53688 Smith smith@eecs 18 3.2 53650 Smith smith@math 19 3.8Cardinality = 3, degree = 5, all rows distinct
Relational Query Languages A major strength of the relational model: supports simple, powerful querying of data. Queries can be written intuitively, and the DBMS is responsible for efficient evaluation.
The SQL Query Language Developed by IBM (system R) in the 1970s Need for a standard since it is used by many vendors Standards: SQL-86 SQL-89 (minor revision) SQL-92 (major revision, current standard) SQL-99 (major extensions)
The SQL Query Language To find all 18 year old students, we can write: SELECT * sid name login age gpa FROM Students S 53666 Jones jones@cs 18 3.4 WHERE S.age=18 53688 Smith smith@ee 18 3.2 •To find just names and logins, replace the first line: SELECT S.name, S.login
Querying Multiple Relations sid cid grade 53831 Carnatic101 C 53831 Reggae203 B 53650 Topology112 A 53666 History105 B SELECT S.name, E.cid FROM Students S, Enrolled E WHERE S.sid=E.sid AND E.grade=“A” S.name E.cid Smith Topology112
Creating Relations in SQL Creates the Students relation. Observe that the type (domain) of each field CREATE TABLE Students is specified, and enforced by the DBMS (sid: CHAR(20), whenever tuples are added or modified. name: CHAR(20), As another example, the Enrolled table login: CHAR(10), holds information about courses that students take. age: INTEGER, gpa: REAL) CREATE TABLE Enrolled (sid: CHAR(20), cid: CHAR(20), grade: CHAR(2))
Combining Separate Systems Use an IR and RDBMS systems which are independent. Divide the query into two: Structured part for the RDBMS Unstructured (text) part for the IR Combine the results from IR and RDBMS Good for letting each vendor develop its own system Bad for data integrity, recovery, portability, and performance
User Defined Operators Allow users to modify SQL by adding their own functions Some vendors used this approach (such as IBM DB2 text extender) Lynch and Stonebreaker defined “user defined operators” to implement information retrieval in 1988 //Retrieves documents that contain term1, term2, term3 SELECT Doc_Id FROM Doc WHERE SEARCH-TERM(Text, Term1, Term 2, Term3) //Retrieves documents that contain term1, term2, term3 // within a window of 5 terms SELECT Doc_Id FROM Doc WHERE PROXIMITY(Text,5, Term1, Term 2, Term3)
Non-First Normal Form Approaches Capture the many-to-many relationships into sets via nested relations Hard to implement ad-hoc queries No standard yet
Using RDBMS for IR Benefits: Recovery Performance Data migration Concurrency Control Access control mechanism Logical and physical data independence
Using RDBMS for IR Example: A bibliography that includes both structured and unstructured information DIRECTORY (name, institution) : affiliation of the author AUTHOR(name,DocId) :authorship information INDEX (name, DocId) :terms that are used to index a document
Using RDBMS for IR Preprocessing SGML can be used as a starting point which is a standard for defining parts of documents <DOC> <DOCNO> WSJ834234234 </DOCNO> <HL> How to make students suffer in IR Course </HL> <DD> 03/23/87</DD> <DATELINE> Sabanci, Turkey </DATELINE> <TEXT> Crawler HW, Inverted Index, Querying </TEXT> </DOC>
Using RDBMS for IR Preprocessing SGML can be used as a starting point which is a standard for defining parts of documents Use a parser together with a hash function to identify terms Use STOP_TERM table for referencing stop words Produce three output tables INDEX (DocId, Term, TermFrequency) : Models the inverted index DOC (DocId, DocName, PubDate, DateLine) : Document metadata TERM (Term, Idf) : stored the weights of each term //Construct TERM table, N is the total number of documents INSERT INTO TERM SELECT Term,log(N/Count(*)) FROM INDEX GROUP BY Term
Using RDBMS for IR An offset can be added together with the term to be able to answer proximity queries. For example “Vice President” should occur together in the same document for relevant documents etc. INDEX_PROX (DocId, Term, OffSet) //Construct TERM table, N is the total number of documents INSERT INTO INDEX SELECT DocId, Term, COUNT(*) FROM INDEX_PROX GROUP BY DocId, Term
Using RDBMS for IR Query can be modeled as a relation as well when it is a long document QUERY(Term,TermFreq) Ex: “Find all news documents written on 03/03/2005 about Sabanci University Data will be extracted from the structured fields Terms will be extracted using the inverted indexSELECT d.DocIdFROM DOC d, INDEX iWHERE i.Term IN (“Sabanci”, “University”) AND d.PubDate = “03/03/2005” AND d.DocId = i.DocId
Using RDBMS for IR Boolean Queries: Consists of terms with boolean operators (AND, OR, and NOT) For a single inputTerm: retrieve the document texts that contain that termSELECT d.TextFROM DOC d,WHERE d.DocId IN (SELECT DISTINCT (i.DocId) FROM INDEX i WHERE i.Term = inputTerm)Note that we can store the text part of a document using BLOB or CLOG (Binary or Character Large Object)
Using RDBMS for IR Boolean Queries that contain ORSELECT DISTINCT (i.DocId)FROM INDEX iWHERE i.Term = inputTerm1 OR i.Term = inputTerm2 OR ….. i.Term = inputTermn OR
Using RDBMS for IR Boolean Queries that contain ANDSELECT DISTINCT (i.DocId)FROM INDEX iWHERE i.Term = inputTerm1 AND i.Term = inputTerm2 AND ….. i.Term = inputTermn AND??
Using RDBMS for IR Boolean Queries that contain AND (Previous Answer Was Wrong)SELECT DISTINCT (i.DocId)FROM INDEX i1, INDEX i2, INDEX i3, …. INDEX inWHERE i1.Term = inputTerm1 AND i2.Term = inputTerm2 AND ….. in.Term = inputTermn AND i1.DocID = i2.DocId AND i2.DocID = i3.DocId AND … in-1 = in.DocIDOR YOU CAN USE INTERSECTION
Using RDBMS for IR Boolean Queries that contain AND Commercial DBMSs are not able to process more than a fixed number of joins. Solution SELECT i.DocId FROM INDEX i, Query q WHERE i.Term = q.term GROUP BY i.DocId HAVING COUNT(i.Term) = (SELECT COUNT(*) FROM QUERY) Works only when the INDEX contains only one occurrence of a given term Together with its frequency. No Proximity is recorded.
Using RDBMS for IR Boolean Queries that contain AND Commercial DBMSs are not able to process more than a fixed number of joins. Solution for terms appearing more than once in the INDEX SELECT i.DocId FROM INDEX i, Query q WHERE i.Term = q.term GROUP BY i.DocId HAVING COUNT(DISTINCT(i.Term)) = (SELECT COUNT(*) FROM QUERY) This is slower since DISTINC requires a sort for duplicate elimination.
Using RDBMS for IR Boolean Queries that contain AND Commercial DBMSs are not able to process more than a fixed number of joins. Implementation of TAND (Threshold AND) is also simple SELECT i.DocId FROM INDEX i, Query q WHERE i.Term = q.term GROUP BY i.DocId HAVING COUNT(DISTINCT(i.Term)) > k
Using RDBMS for IR Proximity Queries for terms within a specific window width SELECT a.DocId FROM INDEX_PROX a, INDEX_PROX b WHERE a.Term IN (SELECT q.Term FROM QUERY q) AND b.Term IN (SELECT q.Term FROM QUERY q) AND a.DocId = b.DocId AND (a.offset –b.offset) BETWEEN 0 AND (width-1) GROUP BY a.DocId, b.DocId, a.Term, a.offset HAVING COUNT(DISTINCT(b.Term)) = SELECT (COUNT(*) FROM QUERY)
Using RDBMS for IR Calculating Relevance SELECT i.DocId, SUM(q.tf*t.idf*t.tf*t.idf) FROM QUERY q, INDEX i, TERM t WHERE q.Term = t.term AND i.Term = t.Term GROUP BY i.DocId ORDER BY 2 DESC