Information Technology Reference
In-Depth Information
Table 1. Extraction accuracy (%)
Accuracy
1-best
2-best 1-best by bootstraping 1-best by dependency parsing
Precision
85.2
97.4
88.6
85.2
Recall
76.9
87.2
46.2
76.9
Amount Ratio
-
-
10.8
89.2
4 Question-Answering System for Agriculture
4.1 Problem and Approach
The basic operation of our question-answering system is extraction of a triple
such as
from a query sentence by using morphological
analysis and dependency parsing. Any question words (what, where, when, why,
etc. are then replaced with a variable and the LOD DB is searched. In other
words, the
subject, verb
,and
object
< subject, verb, object >
triples in the LOD DB are matched against
<
in the query.
SPARQL is based on graph pattern matching, and this method corresponds to a
basic graph pattern (one triple matching). At the data registration, if there is a
resource corresponding to the
?
, verb, object >, < subject,
?
,object>
,and
<subject,verb,
?
>
subject
and a property corresponding to the
verb
from the user statement, a triple that has the
object
from the user statement as
the value is added to the DB.
However, since the schema is open, mapping of query sentence to the schema
poses a problem. Although mapping between the verb in the query sentence
(in Japanese) and a Property in the LOD schema of the DB must be defined
in advance, both of them are unknown in this open schema scenario (in the
closed DB the schema is given), so the score according to the mapping degree
cannot be predefined. The open schema means that the schema is not regulated
by any organization, and there may be several properties of the same meaning
and a sudden addition of a new property. In addition, we assume searching
over multiple LOD sets made by the different authors. We therefore use a string
similarity and a semantic similarity technique using the WordNet thesaurus from
the field of ontology alignment to map verbs to Properties, and attempt to
improve the mapping based on user feedback. We first register a certain set of
mappings
as seeds in the Key-Value Store (KVS). If a verb is
unregistered, we then do the following:
{
verb, property
}
(1) Expand the
to its synonyms using Japanese WordNet ontologies, and
then calculate the LCS with the registered
verb
verbs
to use as the similarity.
(2) Translate the new
into English, and calculate the LCS of the English
with the registered properties.
(3) If we find a resource that corresponds to a
verb
in the query sentence
in the LOD, we then calculate the LCSs of the translated
subject
with all
the properties belonging to the resource, and create a ranking of possible
mappings according to the combination of the above LCS values.
verbs
 
Search WWH ::




Custom Search