A SEMANTIC APPROACH TO DISCOVERING SCHEMA MAPPING

Yuan An, Alex Borgida, Renee J. Miller, and John Mylopoulos

Presented by:

Kristine Monteith

OVERVIEW

Goal of the paper: Matching schemas with more than just simple element correspondence

(e.g. Can we improve on a naïve mapping?)

OVERVIEW

Approach: Derive a conceptual model for the semantics in a table and match the conceptual model in the source schema to the conceptual model in the target schema

e.g. Can we figure out that a source schema like this:

can match a target schema like this: hasBookSoldAt(aname,sid)

EXAMPLE 1

BASELINE SOLUTION: REFERENTIAL INTEGRITY CONSTRAINTS

Find correspondences v1: connect person.pname to hasBookAt.aname v2: connect bookstore.sid and hasBookSoldAt.sid

Create logical relations using referential constraints S1: person(pname) |X| writes(pname, bid) |X| book(bid) S2: book(bid) |X| soldAt(bid,sid) |X| bookstore(sid) S3: person(name) S4: bookstore(sid)

Look at target T1: hasBookSoldAt(aname,sid)

Look at each pair of source and target relations and check to see which are “covered” <S1,T1,v1> <S2,T1,v2> <S3,T1,v1> <S4,T1,v2>

ASK THE USER ABOUT THE FOLLOWING:

Doesn’t present an entire tuple to match the target query: hasBookSoldAt(aname,sid)

WHAT THIS PAPER SEEKS TO ACCOMPLISH:

Generate the following:

compose “writes” and “soldAt” to produce a new semantic connection between “person” and “bookstore”

APPROACH:REPRESENTING SEMANTICS OF SCHEMAS

Create a Conceptual Model (CM) graph Create nodes for classes and attributes Create directed edges for relationships and

inverses

C1 ---ISA--- C2 subclassesC ---p--- D relationshipsC ---p->-- D functional relationships

o Duplicate concept nodes to represent recursive relationships

GENERATING MAPPING CANDIDATES

Problem description Inputs:

A source relational schema S and a target relational schema T

A concept model (GS and GT respectively) associated with each relational schema via table semantic mappings

A set of correspondences L linking a set L(S) of columns in S to a set L(T) of columns in T

Goal: A pair of expressions <E1,E2> which are

“semantically similar” in terms of modeling the subject matter

MARKED NODES

The set L(S) of columns gives rise to a set CS of marked class nodes in the graph GS

Likewise, the set L(T) gives rise to a set CT of marked class nodes in the graph GT

BASIC ALGORITHM

Create conceptual subgraphs find a subgraph D1 connecting concept nodes in

CS, and a subgraph D2 connecting concept nodes in CT such that D1 and D2 are “semantically similar

Suggest possible mapping candidates translate D1 and D2 into algebraic expressions E1

and E2 and return the triple < E1,E2,LM> as a mapping candidate

CREATING CONCEPTUAL SUBGRAPHS

Notice simple matches a node v in CS corresponds to a node u in CT when v and

u have attributes that are associated with corresponding columns via the table semantics

More complicated rules The connections (v1,v2) and (u1,u2) should be

“semantically similar” or at least “compatible” (cardinality constraints, relationships like “is-a” or “part of”)

Use edges from pre-selected trees Represent “intuitively meaningful” concepts Favor smaller trees (Occam’s razor)

Other considerations Favor lossless joins Reject contradictions

EXAMPLE

Looking for a functional tree with a root corresponding to the anchor Proj

EXAMPLE Notice simple matches Find a tree with minimal cost (edges in pre-selected

trees don’t contribute to cost) Find a tree containing the most number of edges in

the pre-selected trees

Project ---controlledBy->-- Department --hasManager->-- Employee

MORE COMPLICATED EXAMPLE

Same Answer:Project ---controlledBy->-- Department --hasManager->-- Employee

Still looking for low-cost, minimal trees to connect Employee to Project

DEALING WITH N-ARY RELATIONS

StoreSells(Person, Product)

CONSIDERATIONS FOR REIFIED RELATIONSHIPS

A path of length 2 passing through a reified relationship node should be considered to be length 1

The semantic category of a target tree rooted at a reified relationship induces preferences for similarly rooted (minimal) functional trees in the source (cardinality restrictions, number of roles, subclass relationship to top level ontology concept)

OBTAINING RELATIONAL EXPRESSIONS

EXPERIMENTAL RESULTS

AVERAGE PRECISION

AVERAGE RECALL

CONCLUSIONS

Semantic approach performs at least as well as the RIC-based approach on datasets studied

These approaches made significant improvements in some cases

Many of the datasets did not have complicated schema; a semantic approach didn’t provide as much benefit in those cases

STRENGTHS/WEAKNESSES

Strengths Lots of examples Provides a useful solution to a common problem

Weaknesses Formalism sometimes made things more

complicated rather than more clear Assumes a lot of background knowledge

FUTURE WORK

Embed this functionality into pre-existing mapping tools (they suggest Clio since a lot of their work is based off of this)

Add negation to semantic representation Investigate more complex semantic

mappings

QUESTIONS???