Finding Optimal Probabilistic Generators for XML Collections
Serge Abiteboul, Yael Amsterdamer,
Daniel Deutch, Tova Milo, Pierre Senellart
Adding probabilities to an XML Schema
• Given a collection of XML documents, we sometimes have a schema the documents conform to.
– E.g., DTD or XSD
– Restricts the structure, mostly parent-child node relations (using regular expressions)
• The schema may be very general (e.g., xhtml, RSS)
• We want to add probabilities that reflect the likelihood of different parts of the schema
– We will use the probabilities to turn the schema into a probabilistic generative model for XML documents
– In particular, we want them to maximize the likelihood of a given XML document or document collection
Motivation
One Application: XML Auto- Completion [SIGMOD 2012]
• Based on previous document versions / corpus of example documents –
• Suggest nodes / sub-trees / node values to the user
• For example:
• Challenges:
– Allow editing in every part of the document
– What kind of completion to suggest?
– Finding the top-k best completions
Motivation
<MyPapers>
<Paper>
<title>XML for Beginners</title>
<author>M. Jones<author>
<author>H. Q. David</author>
<author>L. Martin</author>
<author>S. Smith</author>
</Paper>
<Paper>
<title>Advanced XML</title>
<author>M. Jones</author>
<author>J. E. Peterson</author>
<author>G. L. Williams</author>
</Paper>
<Paper>
<title> </title>
<author> </author>
<author> </author>
<author> </author>
</Paper>
</MyPapers>
Many Other Usages for a Probabilistic Schema
• Testing – e.g., generating many XML messages to
simulate network load and test system performance.
• Explaining – e.g., a probabilistic schema for DBLP may show which types of publications are rarely used,
which kinds of attributes are not filled for BibTex, etc.
• Schema Evaluation – how well a given schema describes a given corpus.
• …
Motivation
Our solution - An Outline
Preliminaries – Tree Automata
Generators for Schemas without Constraints
Restart Generators
Continuation-Test Generators Leaf Values
Adding Constraints
Schema as a Deterministic Tree Automaton
Preliminari es
q
0b q
1q
2a c
$
An XML document is
modeled as an ordered tree.
Document d 0 :
Schema validation: the children of an a-labeled node are accepted by DFA A a
Automaton A r : (L(A
r) = a*bc*$)
Validation is performed for the children of every inner node.
abcd 532 abcd
$
r
a b c
Using the Schema as a Generator
• Recall that we want to turn the schema from an acceptor into a probabilistic generative model.
• Straightforward nondeterministic generator: repeatedly choose an accepting run for a node's automaton, and generate children accordingly.
• Adding probabilities: we consider two problem settings
1. Generating documents that are accepted by the schema, while maximizing the likelihood of a corpus.
2. Additionally, imposing integrity constraints on the documents (e.g., key constraints)
Preliminari
es
Probabilistic Generator
• Each transition is assigned a probability
• We assume independent choices, (a Markovian process) thus the document probability is the product.
• In this case, Pr(d)=p
a∙ p
a∙ p
b∙ p
$• The schema and generator ignore leaf values (for now!)
Without Constraints
b
a c
$ p a p c
p b p $
q
0q
1q
2$
r
a a b
Formal Problem Definition
• Given a corpus D of documents ,
• and a deterministic schema S that accepts every document in D
• We want to find an optimal generator based on S:
– Find probabilities for the transitions of S that maximize the probability of generating D,
– i.e., the maximum likelihood estimator (MLE).
Without
Constraints
A Learning Algorithm
Without Constraints
b
a c
$
$
The frequency of using each
transition during the corpus verification process is recorded.
(q
0, a) (q
0, b) (q
1, c) (q
1, $)
1 1 1
1 q 0 q 1 q 2
r
a b c
An Algorithm for Probabilities Learning (Cont.)
This is repeated for every node in every corpus document.
We set the probability of each transition to be its relative frequency.
Without Constraints
(q
0, a) 1 (q
0, b) 1 (q
1, c) 1 (q
1, $) 1
/2 /2 /2
Theorem: This efficient algorithm /2
learns the MLE probabilities – finds
an optimal probabilistic generator
An Additional Result
• Theorem: generation terminates with probability 1.
– Guaranteed only because of the choice of probabilities according to the corpus.
Without
Constraints
Integrity Constraints
• We want to support integrity constraints, which are used in XML schema languages.
• Key Constraint: the leaves of a-labeled leaves have unique values (unary key)
• Inclusion Constraint: the values of a-labeled leaves are contained in those of b-labeled leaves
• Domain Constraint: the values of a-labeled leaves belong to some (finite or infinite) domain
• Different types are considered in the literature [Fan & Libkin 2001; David Libkin & Tan 2011]
Adding
Constraints
New Problem
• We want to find optimal generators for XML schemas with constraints.
• Valid generator output: an XML document, which
1. is a accepted by the schema, and
2. there exists a valid leaf value assignment – which does not violate the constraints
– Example: each of a, b, c is unique, and contained the others
Adding Constraints
$ r
a a bc
r
a b
b
c
…
b
Restart Generators
• A simple idea:
– Use a probabilistic generator to generate a document
– Check if it has a value assignment valid w.r.t. the constraints – If not, 'restart' and try again until a valid document is generated
• Problem definition -- same as in the case without constraints (but now the schema includes constraints)
• Proposition: Given a document with no values, checking for the existence of a valid value assignment is in PTIME
– Proof: By translating the constraints to bounds on the number of unique values for each leaf label
• Bad news: number of restarts can be unboundedly large in an optimal generator
Adding
Constraints
Continuation-test Generators
• Never make choices that lead to a 'dead end', thus always generate a valid document.
• We use a binary test to check if a choice has a continuation.
• Example: add to the schema of d
0the constraints:
– c is included in a – c is unique
• The generation process:
Adding Constraints
b
a c
$ $
p a p c p b p $
q
0q
1q
2r
a b c
Pr( d ) = p a ∙ p b ∙ p c ∙1
Perform a continuation-test before taking the
transition
Implies |
c|≤|a|
Learning Algorithm for
Continuation-test Generators
• The probabilities are again relative frequencies, but –
only in cases where there was an alternative choice.
• The learned generator will generate as many c-s as a-s Adding
Constraints
(q
0, a) 1 (q
0, b) 1 (q
1, c) 1 (q
1, $) 0
/2 /2 /1
/1 (q
1, $) was chosen
only when (q
1, c)
was not available.
Results for Continuation-test Generators
• Theorem: The algorithm learns an optimal continuation-test generator, for automata with binary choices.
– Extensions to non-binary are discussed in the paper
• Theorem: Continuation-test is NP-Complete
– But only in the size of the schema; it is polynomial in the document size
– Both generation and finding the optimal generator are exponential in the schema size unless P=NP.
– Based on schema satisfiability test [David et al. 2011]
• Theorem: probability of termination for a continuation-test generator may be arbitrarily small!
– Proof – by construction of a simple, non-recursive schema – Can be handled by adding a constraint on the document size.
– Sub-classes of schemas that guarantee termination?
Adding
Constraints
Adding Values to the Structure
• So far our generators were used only for the document structure
• Leaf values may also have a distribution according to which they can be generated
– The distribution may be learned from the same document collection
• We will focus on the interesting case – generating leaf values for a schema with constraints
Leaf Values
Suggested Algorithm
• We start with a valid document skeleton
• Order labels by inclusion constraints (e.g., c, b, a)
• Choose a leaf from the 'smallest' (most included) label, and including leaves
• Draw a value (from the domain) according to a given distribution.
• Use PTIME test to verify validity, if not revert the step
• Improvements presented in the paper
Leaf Values
$
r
a b c
abcd
abcd efg
Related Work
• Schema Satisfiability tests [Fan & Libkin 2001; David, Libkin & Tan 2011]
• Probabilistic XML and Probabilistic Schemas [e.g., Benedikt, Kharlamov, Olteanu & Senellart 2010]
• Probabilistic XML generation [e.g., Antonopoulos, Geerts, Martens & Neven 2011]
• Schema Inference [e.g., Bex, Gelade, Neven & Vansummeren 2008]
• AXML [Abiteboul, Benjelloun & Milo 2008]
• PCFGs [e.g., Chi & Geman 1998]
Summary
Conclusion
• A model for a probabilistic XML generators
• Unconstrained case
– Generation and learning optimal generators can be done efficiently – Termination is guaranteed
• Constrained case
– Restart generator
• # of restarts is unbounded
– Continuation-test generators
• Generation and learning optimal generators are expensive
• Termination is not guaranteed