Question answering system plays a vital role in search engine optimization model. Natural language processing methods are typically applied in QA system for inquiring user’s question and numerous steps are also followed for alteration of questions to query form for receiving a precise answer. This presentation analyzes diverse question answering systems that are based on semantic web technologies and ontologies with different formats of queries.It ends by addressing various reasoning alternatives.
2. Overview
Problem Description
What is Question Answering
Question Answering Challenges
Linked Data Challenges
Systems Analysis
Data Linking
Question Answering
Question Answering Evaluation Metrics
Reasoning Analysis
Reasoning Challenges
Conclusion
3. What is Question Answering?
3
QA
System
Knowledge
Bases
Question: From which
university did the wife of
Barack Obama graduate?
Answer: Harvard Law
School
4. Querying Data over the Web
(a) natural language query over two search engines;
(b) corresponding SPARQL representation;
(c) semantic gap between the user’s information needs and data
6. Structured Queries Vs. Q&A
Structured Queries:
A priori user effort in
understanding the schemas
Effort in mastering the syntax of
a query language
Satisfying information needs may
depend on multiple querying
operations
Input: Structured query
Output: data records,
aggregations etc.
6
Q&A:
Delegates more ‘interpretation
effort’ to the machines
Input: natural language query
Output: direct answer
7. Ability to query datasets by referencing elements in data model structure,
as well as to operate over the data (aggregate results, express conditional
statements, etc.)
Easy-to-operate, intuitive, and task-efficient query interface
Matches entities expressed in the query to semantically equivalent dataset
entities
Ability to answer queries not supported by explicit dataset statements
e.g. “Is Natalie Portman an Actress?” can be supported by the statement “Natalie Portman starred Star
Wars,” instead of an explicit statement “Natalie Portman occupation Actress,” which might not be present in
dataset
Ability to semantically match user query terms to dataset vocabulary-level
terms
Question Answering Challenges
Query
Expressivity
Usability
Entity
Reconciliation
Semantic
Tractability
Vocabulary-
level Semantic
Matching
8. Linked Data Challenges
Linked Data,
Is huge
Is not “pure”
Is inconsistent
Is evolving
Needs more than RDFS and OWL
What is needed from the Semantic Web Pie?
What parts of RDFS and OWL do people use?
9. Data Linking Systems
Normalize information to common
vocabularies
Structured/Semi-Structured Information to
Linked Data
Automated/Semi-Automated Mechanisms
Solve the Semantic Tractability, the
Vocabulary-level Semantic Matching, and the
inconsistency problems.
10. Data Level Knowledge
Level
System Internal External
LN2R String matching - Word Net synonyms
dictionary
COREF String matching - -
OKKAM String matching Translation service Entity names vocabulary
LDMAPPER String lookup search - -
SILK String matching numerical similarity - -
LIMES String matching on metric spaces - -
Knofuss String matching - -
RDF-AL String matching Translation service Word Net taxonomic
distance
Zhishi.links String matching - Abbreviations list
Data Linking Systems Analysis
(1/2)
11. Data Level Knowledge
Level
System Internal External
Serimi String matching - -
Knofus String matching on numerical similarity - -
Limes String matching Translation service Word Net
Querix String matching on metric spaces Translation service Word Net
Quacid String matching - -
SWSE String matching - -
SMART String matching - -
ORAKEL String matching - -
AQALOG String matching - -
Data Linking Systems Analysis
(2/2)
16. Reasoning Challenges on Q&A
Abrupt and strict answers ( Not statistical
approach)
Inefficient
Not Scaling
Modeling Complexity
17. Reasoning Approaches for LD
Context-Dependent Reasoning
Authoritative Reasoning
LiDaQ
Link Traversal Based Query Execution(LTBQE)
extension with Reasoning
18. Conclusion
Open Questions
Graph Inference Algorithms
Scaling Reasoning on the Web
Context Dependent Reasoning
Are there better standards for Linked Data?
Future Work
A question sentence is a sequence of tokens, The input question is fed into the following pipeline of six steps:
Phrase detection. Phrases are detected that potentially correspond to semantic items such as ‘Who’, ‘played in’, ‘movie’ and ‘Casablanca’.
Phrase mapping to semantic items. This includes finding that the phrase ‘played in’ can either refer to the semantic relation acted In or to played For Team and that the phrase ‘Casablanca’ can potentially refer to Casablanca (film) or Casablanca, Morocco . This step merely constructs a candidate space for the mapping.
Q-unit generation. Intuitively, a q-unit is a triple composed of phrases.
Joint disambiguation, where the ambiguities in the phrase-to-semantic-item mapping are resolved. This entails resolving the ambiguity in phrase borders, and above all, choosing the best fitting candidates from the semantic space of entities, classes, and relations.
Semantic items grouping to form semantic triples. For example, we determine that the relation married to connect person referred to by ‘Who’ and writer to form the semantic triple person married to writer. This is done via q-units.
Query generation. For SPARQL queries, semantic triples such as person married to writer have to be mapped to suitable triple patterns with appropriate join conditions expressed through common variables: ?x type person , ?x marriedTo ?w, and ?w type writer for the example.