A distributional structured semantic space for querying rdf graph data

Digital Enterprise Research Institute www.deri.ie

A Distributional Structured Semantic
Space for Querying RDF Graph Data
André Freitas, Edward Curry, João G. Oliveira,
Seán O’Riain

 Copyright 2009 Digital Enterprise Research Institute. All rights reserved.

Outline

 Motivation & Problem Space
 Description of the Proposed Approach
 Evaluation
 Conclusions

Motivation & Problem Space

 Linked Data: Heterogeneous and distributed data
environment.
 Traditional Database and Information Retrieval
approaches for querying/searching Linked Data
provide limited solutions for casual users.
 Querying and searching Linked Data remains a
fundamental challenge.

?

Query/Search Spectrum

Adapted from Kauffman et al (2009)

Semantic Gap

From which university did the wife of
Barack Obama graduate?
Data model independent (DMI) queries
Semantic Matching

Semantic Matching Problem

From which university did the wife of Barack Obama graduate?

Semantic Matching Problem

From which university did the wife of Barack Obama graduate?

Semantic matching

Semantic Matching Requirements

 Expressivity:
 Supporting expressive natural language queries.
 Path, conjunctions, disjunctions, aggregations, conditions.
 Flexibility & Accuracy:
 Accurate and flexible semantic matching approach.
 Maintainability:
 Easily transportable across datasets from different domains.
 Can support Open domain (e.g. Wikipedia) and doman-specific
(e.g. Financial) datasets without customization.
 Performance:
 Suitable for realtime querying (low query execution time).
 Scalability:
 Scalable to a large number of datasets (Organization-scale, Web-
scale).

Research Questions

Q1: How to provide a query mechanism which allows
users to expressively query linked datasets without a
previous understanding of the vocabularies behind
the data?

Q2: Which semantic model could support this query
mechanism?


Proposed Approach

Proposed Approach (Outline)

 Treo:
 Query approach based on dynamic Linked Data navigation.
 Performance limitations.
 Treo T-Space:
 Existing IR approaches: traditional Vector Space Models
(VSMs) were not able to: (i) capture the structure of graphs
and (ii) support a precise semantic matching behavior.
 A VSM supporting these two requirements was formulated:
T-Space.
 Ranking function based on semantic relatedness.

Proposed Approach (Outline)

First Approach (Treo)

 Natural language queries.
 Ranked query results.
 Two phase search process combining entity search
with spreading activation search.
 Use of Wikipedia-based semantic relatedness to
semantically match query terms to dataset terms.
 Limited query execution performance.

Requirements: Expressivity, Flexibility & Accuracy.

Semantic Relatedness

 Computation of a measure of “semantic proximity”
between two terms.
 Allows a semantic approximate matching between
query terms and dataset terms.
 Most existing approaches use WordNet-based
solutions for approximate semantic matching
(limited solution).
 Use of Wikipedia-based semantic relatedness
measures (addresses the limitations of WordNet-
based measures).

Query Approach Rationale (Treo)


RDF


Wikipedia Link
Measure


Final Query- Data Matching: Currently limited to path queries

Querying Linked Data using Semantic Relatedness: A Vocabulary Independent
Approach. In Proceedings of the 16th International Conference on Applications of
Natural Language to Information Systems (NLDB) 2011.


Treo (Irish): Direction, path

From which university did the wife of Barack Obama
graduate?

Second Approach: Treo T- Space

 Motivation:
 Performance problem of the original approach.
 Transportability across different domains.
 Rationale:
 Rephrasing of the proposed approach as a vector
space model.
 Addressing the limitations of the vector space
model (BoW: lack of structure) to represent the
structure present in the data.
 Keeping the semantic matching behavior.
Requirements: Performance & Scalability, Maintainability.

T- Space

 A Vector Space Model for building RDF indexes
that preserve the graph structure and which
allows a semantic matching/search behavior.

Generalization (Treo T- Space)

 Approach:
 Ranked results.
 Natural language queries.
 Two phase search process combining entity search
with spreading activation search.
 Use of distributional semantics to semantically
match query terms to dataset terms.
 Semantic manifold (‘structured vector space’)
model: T-Space.

Distributional Semantic Model

 Assumption: the context surrounding a given word
in a text provides important information about its
meaning.
 Simplified semantic model.
 Explicit Semantic Analysis (ESA) is the primary
distributional model used in this work.

A wife is a female partner in a marriage. The term "wife"
seems to be a close term to bride, the latter is a female
participant in a wedding ceremony, while a wife is a married
woman during her marriage.
...

T- Space

Embedding the structure of a graph into a
semantic vector space
Distributional
relations
semantic model:
Semantic statistical
knowledge extracted
from large Web
corpora
instances

classes

Querying the T- Space

Querying the T- Space

Querying Heterogeneous Datasets on the Data Model Independent Queries over RDF
A Multidimensional Semantic Space for Linked Data Web: Challenges, Approaches
and Trends.Proceedings of the 5th Special Issue on Internet-Scale Data, Semantic
Data. In IEEE Internet Computing, IEEE International Conference on 2012.
Computing (ICSC), 2011.

Ranking/Filtering Results

 How to filter non-related results.
 Analysis of semantic relatedness as a ranking
function.
 Creation of a methodology for the determination of
a semantic relatedness threshold.
 Semantic differential analysis.

Requirements: Flexibility & Accuracy.

A Distributional Approach for Terminological Semantic Search on the
Linked Data Web. 27th ACM Applied Computing Symposium,
Semantic Web and Its Applications Track, 2012.

VSM & Formalization

 Based on the Generalized Vector Space Model.
 Simplification and clarification of T-Space
properties.
 Tensors are used to support domain specific
meaning.
 Tensors can be used to connect concepts across
datasets and T-Spaces

VSM & Formalization

1. The distributional
space can be
transformed to the
keyword space.

Requirements: Flexibility
& Accuracy

VSM & Formalization

2. The vector field defines
an additional structure
over the vector space
model.

Requirements: Expressivity

Customized Distributional Model

Distributional models customizable for different datasets.

Customized Distributional Model

Transformation from different distributional models

Wikipedia Financial
Corpus
DBPedia Financial Dataset

A Distributional Structured Semantic Space for Querying RDF Graph
Requirements: Maintainability
Data. International Journal of Semantic Computing (IJSC), 2012.


Evaluation
(Treo T- Space)

Quality of Results

 Q1: How does the approach work for addressing the
vocabulary problem?
 Q2: How does the approach work for addressing
natural language queries?

 Precision, recall, mean reciprocal rank, % of
answered queries.

Quality of Results

 QALD DBPedia Training Set.
 50 natural language queries.
 DBpedia 3.6.
Full DBPedia QuerySet (50 queries)
Avg. Precision Avg. Recall MRR % of queries
answered
0.482 0.491 0.516 58%

Partial DBPedia QuerySet (38 queries)
Avg. Precision Avg. Recall MRR % of queries
answered
0.634 0. 645 0.679 76%

Error Distribution

 Q4: What is the error associated with each
component of the search approach?

 Error distribution measurements.

Error Distribution

 Conclusion: Robustness of the distributional
semantic model for open domain queries.
 Conclusion: Major need for improvement of the
pre/post processing phases.

Evaluating Terminology- level Semantic
Matching

 Q1: How does the approach work for addressing the
vocabulary problem?
 Q5: Does the approach improve the semantic
matching of the queries?

 Quantitative evaluation: P@5, P@10, MRR, % of the
queries answered, comparative evaluation using
string matching and WordNet-based query
expansion.

Evaluating Terminology- level Semantic
Matching

Avg. Avg. MRR % of queries
Precision@5 Precision@10 answered
0.732 0.646 0.646 92.25%

Approach % of queries answered

ESA 92.25%
String matching 45.77%
String matching + WordNet QE 52.48%

Comparative Analysis

 Q3: What is the best distributional semantic model
for the vocabulary problem?

 Preliminary comparative analysis between
different distributional semantic models.

Comparative Analysis (Treo vs Treo T- Space)

 Wikipedia Link Measure (WLM) vs Explicit Semantic Analysis (ESA)

ESA (Full Query Set)
Avg. Precision Avg. Recall MRR % of queries answered
0.482 0.491 0.516 58%

WLM (Full Query Set)
Avg. Precision Avg. Recall MRR % of queries answered
0.395 0. 451 0.489 56%

Improvement
% Avg. Precision % Avg. Recall % MRR % of queries answered

18% 8.2% 5.2% 3.5%

Comparative Analysis

 Conclusion: From the two tested approaches, ESA
provides a better semantic model.

Conclusions

 The T-Space model provides a principled way to build data
model intependent queries over RDF graphs.
 The distributional semantic model supports a flexible
matching between query terms and dataset terms in a
semantic best-effort scenario.
 The ESA semantic model provides a better distributional model
compared to WLM.
 Improvements are needed on the pre/post processing phase of
the approach.

Associated Article

André Freitas, Edward Curry, João Gabriel Oliveira, Sean O'Riain, A
Distributional Structured Semantic Space for Querying RDF Graph
Data. International Journal of Semantic Computing (IJSC),2012.

http://www.worldscinet.com/ijsc/05/0504/S1793351X1100133X.html

http://andrefreitas.org/papers/preprint_distributional_structured_space.pdf

Related Publications

André Freitas, Edward Curry, João Gabriel Oliveira, Sean O'Riain, Querying Heterogeneous Datasets on the Linked Data Web:
Challenges, Approaches and Trends. IEEE Internet Computing, Special Issue on Internet-Scale Data, 2012 (Article).

André Freitas, Edward Curry, João Gabriel Oliveira, Sean O'Riain, A Distributional Structured Semantic Space for Querying
RDF Graph Data. International Journal of Semantic Computing (IJSC), 2012 (Article).

André Freitas, Sean O'Riain, Edward Curry, A Distributional Approach for Terminological Semantic Search on the Linked Data
Web. 27th ACM Applied Computing Symposium, Semantic Web and Its Applications Track, 2012 (Conference Full Paper).

André Freitas, João Gabriel Oliveira, Edward Curry, Sean O'Riain, A Multidimensional Semantic Space for Data Model
Independent Queries over RDF Data. In Proceedings of the 5th International Conference on Semantic Computing (ICSC),
2011. (Conference Full Paper).

André Freitas, João Gabriel Oliveira, Sean O'Riain, Edward Curry, João Carlos Pereira da Silva, Querying Linked Data using
Semantic Relatedness: A Vocabulary Independent Approach. In Proceedings of the 16th International Conference on
Applications of Natural Language to Information Systems (NLDB) 2011. (Conference Full Paper).

André Freitas, João Gabriel Oliveira, Sean O'Riain, Edward Curry, João Carlos Pereira da Silva, Treo: Combining Entity-Search,
Spreading Activation and Semantic Relatedness for Querying Linked Data, In 1st Workshop on Question Answering over
Linked Data (QALD-1) Workshop at 8th Extended Semantic Web Conference (ESWC), 2011 (Workshop Full Paper) .

André Freitas, João Gabriel Oliveira, Sean O'Riain, Edward Curry, João Carlos Pereira da Silva, Treo: Best-Effort Natural
Language Queries over Linked Data, In Proceedings of the 16th International Conference on Applications of Natural
Language to Information Systems (NLDB), 2011 (Poster in Proceedings).


http://andrefreitas.org

A distributional structured semantic space for querying rdf graph data

More Related Content

What's hot (20)

Similar to A distributional structured semantic space for querying rdf graph data (20)

More from Andre Freitas (20)

Recently uploaded (20)

A distributional structured semantic space for querying rdf graph data