Question Answering System Based on Ontology and Semantic Web

wafflebazaarInternet and Web Development

Oct 21, 2013 (3 years and 9 months ago)

71 views

Question Answering System Based on Ontology and
Semantic Web













ABSTRACT

Semantic web technologies bring new benefits to knowledge-
based Question Answering system. Especially, Ontology is
becoming the pivotal methodology to represent domain-specific
conceptual knowledge in order to promote the semantic capability
of a QA system. In this paper we present a QA system in which
the domain knowledge is represented by means of Ontology. In
addition, personalized services are enabled through modeling
users' profiles in the form of Ontolog, and a Chinese Natural
Language human-machine interface is implemented mainly
through a NL parser in this system. An initial evaluation result
shows the feasibility to build such a semantic QA system based
on Ontology, the effectivity of personalized semantic QA, the
extensibility of ontology and knowledge base, and the possibility
of self-produced knowledge based on semantic relations in the
ontology.

Categories and Subject Descriptors

I.2.7 [Natural Language Processing]: Language Parsing and
Understanding – www, ontology, semantic web, question
answering.
General Terms

Management, Documentation, Languages, Theory
Keywords

www, ontology, semantic web, question answering.
1. INTRODUCTION
Semantic web technologies bring new benefits to knowledge-
based Question Answering system. Especially, Ontology is
becoming the pivotal methodology to represent domain-specific
conceptual knowledge in order to promote the semantic capability
of a QA system.
Specific research in the areas of QA has been advanced in the past
couple of years particularly by the QA track of the TREC-QA
competitions [1]. The TREC-QA competitions focus on open-
domain systems, i.e. systems that can potentially answer any
generic question. Therefore, these competitions are based on large
volumes of unstructured text, which makes deep text analysis
become resource-consuming tasks. In contrast, a QA system
working on a specific technical domain can make use of the
specific domain-dependent terminology to recognize the true
meaning included in a segment of natural language text. So we
realize that the terminology plays a pivotal role in a technical
domain such as Java programming. A great deal of work has been
done representing domain-specific concepts and the terminology
by means of Ontology, i.e. UMLS [2]. Recent research
advancements on Knowledge Representation with Semantic Web
and Ontologe have proved that this methodology is able to
promote the semantic capability of a QA system.
The Semantic Web is a Web that includes documents, or portions
of documents, describing explicit relationships between things
and containing semantic information intended for automated
processing by our machines. It operates on the principle of shared
data. When you define what a particular type of data is, you can
link it to other bits of data and say "that's the same", or some other
relation. For example, "zip" in my QA system based on Semantic
Web is the same as "zip" in my friends. Although it gets more
complicated than this, that is bascially what the Semantic Web is
all about, sharing data through ontologies, and processing it
logically. Trust is also important, as the trust of a certain source is
fully in the hands of the user. This is a fully decentralized system:
"you can't make something be the center of all knoweldge".
Although the Semantic Web is a Web of data, it is intended
primarily for humans; it would use machine processing and
databases to take away some of the burdens we currently face so
that we can concentrate on the more important things that we can
use the Web for.
For example, recent research in information processing has
focused on health care consumers [3]. These users often
experience frustration while seeking online information, due to
their lack of understanding of medical concepts and unfamiliarity
with effective search strategies. We are exploring the use of
semantic relationships as a way of addressing these issues.
Semantic information can guide the lay health consumer by
suggesting concepts not overtly expressed in an initial query. For
example, imagine that a user submits a full question to a search
system in the health care domain to find out whether exercise
helps prevent osteoporosis. The semantic relationship prevents in
the proposition representing the question, namely ”exercise
prevents osteoporosis”, can support this effort; prevents might be
used with osteoporosis to determine additional ways of avoiding
this disorder. We present an analysis of semantic relationships
that were manually extracted from questions asked by health
consumers as well as answers provided by physicians. Our work
concentrates on samples from Ask-the-Doctor Web sites. The
Semantic Network from the Unified Medical Language System
(UMLS) [4,5] served as a source for semantic relationship types
and this inventory was modified as we gained experience with
relationship types identified in the health consumer texts.
A semantic relationship associates two (or more) concepts
expressed in text and conveys a meaning connecting those
concepts. A large variety of such relationships have been
identified in several disciplines, including linguistics, philosophy,
computer science, and information science. Some researchers
have organized hierarchies of semantic relationships into
meaningful but not formal structures. Others examine specific
relationships in depth, for instance, subsumption, temporality, and
meronymy. In addition, ontologies contain semantic relationships
that are elements of the overall system. WordNet, for example,
contains these primary relationships between concepts:
hypernymy, antonymy, entailment, and meronymy (part—whole).
A number of projects have involved the study of semantic
relationships specifically within the domain of medicine. Work on
the GALEN Common Reference Model examined part—whole
relationships and other aspects of “tangled” taxonomies. Other
ontology projects, such as the foundational model of anatomy
(FMA), are central to the delineation of relationships for use in
specific types of applications, in this case representation of
anatomical structures.
2. SEMANTIC WEB AND AGENT-BASED
SEMANTIC WEB SERVICES QUERY
Making the Web more meaningful and open to manipulation by
software applications is the objective of the Semantic Web
initiative. Knowledge representation and logical inference
techniques form the backbone. Annotations expressing meaning
help software agents to obtain semantic information about
documents [6]. For annotations to be meaningful for both creator
and user of annotations, a shared understanding of precisely
defined annotations is required. Ontologies – the key to a
semantic Web – express terminologies and semantical properties
and create shared understanding. Ontologies consist of
hierarchical definitions of important concepts in a domain and
descriptions of the properties of each concept, supported by
special logics for knowledge representation and reasoning. Web
ontologies can be defined in DAML+OIL – an ontology language
based on XML and RDF/RDF Schema.Some effort has already
been made to exploit Semantic Web and ontology technology for
the software engineering domain [7]. DAML-S [8] is a
DAML+OIL ontology for describing properties and capabilities
of Web services, which shows the potential of this technology for
software engineering. Formality in the Semantic Web framework
facilitates machine understanding and automated reasoning.
DAML+OIL is equivalent to a very expressive description logic
[9]. This fruitful connection provides well-defined semantics and
reasoning systems. Description logic is particularly interesting for
the software engineering context due to a correspondence
between description logic and dynamic logic (a modal logic of
programs). We propose to define a semantic interface definition
language IDL and a reasoning technique for component matching
in form of ontology. The connection between description logic
and modal logics allows us to introduce reasoning about
component and service matching within a Semantic Web
framework.
In the conventional Web Services approach exemplifiedby WSDL
or even by DAML Services, the communicative intent of a
message is not separated from the application domain. This is at
odds with the convention from the multi-agent systems world,
where there is a clear separation between the intent of a message,
which is expressed using an agent communication language, and
the application domain of the message, which is expressed in the
content of the message by means of domain-specific ontologies.
This separation between intent and domain is beneficial because it
reduces the brittleness of a system. If the characterisation of the
application domain (the ontology) changes, then only that
component which deals with the domain-specific information
need change; the agent communication language component
remains unchanged. In addition, the domain-neutral performatives
may be combined to form common patterns of interaction such as
contract nets, markets or auctions that enable the behaviour of a
system to be considered in more abstract terms.
When the service in the QA example is invoked, the value of the
input parameter should be an instance of the class restriction that
is given as the input parameter types in both the profile and the
process descriptions. For the various query performatives (query-
if, query-ref and subscribe), this input parameter contains the
query expression that would be contained in the message content
in a conventional agent-based system. However, there is as yet no
standard query language for RDF, DAML + OIL or OWL,
although there are several under development, including DAML
Rules [10] (which builds on DAML + OIL and expresses queries
as Horn clause-like structures), the DAML Query Language
As an example, the domain ontology that we have designed for
this application is centred on events and reports of events. We
have taken the approach that communication in the system will be
about these events and reports (rather than about any persistent
world state which the reports might suggest), so the queries can be
expressed using the anonymous resource technique by specifying
the properties that the report (and the event it contains) must
possess. It should be noted, however, that we did not specifically
design the ontology in this report to circumvent the expressive
limitations of our chosen query language, but rather that the query
language was chosen because it was appropriate for use with the
domain ontology that we had already designed.
An additional limitation of this approach to query construction,
which is unfortunately also shared with several of the other query
languages currently under development, is that it is not possible to
specify literal ranges in queries. An example of such a literal
range might be a query of movement reports about entities that
were north of a particular point, or to put it a different way, whose
latitude was greater than a certain value. This limitation arises
because the RDF and DAML + OIL models have no notion of
how different literal datatypes behave (particularly with respect to
ordering and inequalities). The most likely solution to this
problem is based on the use of ontology, entities which have
specific knowledge of the behaviour of different literal types
(integers, latitude/longitude pairs, dates, etc.) and which may be
used by inference and query engines to evaluate tests based on
those types.
3. ONTOLOGIES AND SEMANTIC DATA
INTEGRATION
A truly semantic representation of knowledge that is inherently
more scalable and flexible, and better able to support multiple
existing and future business applications has recently become
available at an appropriate scale. This representation is ontology,
which is a discipline founded in philosophy and computer science
that is at the heart of the new wave of semantic technologies such
as Semantic Web. Ontology contains a representation of all of the
concepts present in a domain and all of the relationships between
them. These associations between concepts are captured in the
form of assertions that relate two concepts by a given relationship.
These triplets (variously described as concept–relationship–
concept, subject–predicate–object or assertions) are the building
blocks of most ontology formats, including the open Semantic
Web standards of RDF and OWL (http://www.w3.org/TR/owl-
ref/). In addition to commercial suites, several open-source tools
have been developed to support the application of such standards
[11,12]. These assertions can, in their simplest form, use IS-A
relationships that, when aggregated, build into a hierarchy or
taxonomy. The taxonomies of concepts (and relationships) can be
extremely useful in their own right, especially when the concepts
are annotated with properties such as synonyms. This enables
users to specify high-level ‘family’ concepts when performing a
search or selecting data for analysis. True ontologies, however,
have a broad range of relationships among concepts. These
relationships also include all known synonyms. This enables, for
example, all of the many English variant forms of the BINDS-TO
relationship between proteins and compounds to be used to build
a complete and detailed picture of the interactions around a
protein or protein family.
The map of concepts and relationships in ontology provides a
crucial enabling resource for true semantic data integration.
Ontology enables information from one resource to be mapped
accurately at an extremely granular level to information from
another source. Multiple instances of a concept (or its synonyms)
in different structured or unstructured data sources can be mapped
to a specific ontology concept and, therefore, the data in those
original sources can be integrated semantically. The ontology
provides the common vocabulary for the data integration –
showing the preferred names for a concept, and the synonyms and
properties associated with it. This enables forward-looking
integration by collecting data using names that are already well
understood rather than ones that might not be shared widely
throughout the organization. This makes the assimilation of new
data easier and quicker, and facilitates communication between
groups [13]. Organizing data integration around the ontology
provides the middle layer that makes data integration more
efficient – reducing the cost, maintenance and risk of the project.
Furthermore, because the ontology can be grown over time as
new data become available, new links are continually being made
and new knowledge assimilated in the ontology.
Ontologies provide a highly dynamic and flexible map of the
information contained in the data sources within a domain [14].
Because ontologies enable true semantic integration across the
data sources that they represent, it is possible not only to draw
wider conclusions from the data but also to look at the data from
several distinct perspectives relevant to the specific job being
undertaken. The generation of ontologies to represent data from
several underlying data sources is a precursor to and an important
enabler of semantic data integration. Ontologies make data
integration more efficient and more detailed, and reduce the risk
associated with the continual redevelopment of project-specific
integration strategies. Ontologies form an atlas of all of the
knowledge of an organization as embodied in its databases,
licensed data sources and personal observations of its scientists.
The ontology grows as new data sources are added to it,
becoming a core corporate asset rather than another new super-
silo for data. The ontology can underpin a range of applications
that delivers the right information at the right time to make better-
informed decisions throughout the lifecycle of drug discovery,
development, marketing, sales and post-market surveillance. In
terms of safety, identifying new patterns of toxic response,
understanding the molecular basis of adverse events and
formulating a response to events in the marketplace such as the
withdrawal depends on the delivery of just such a set of consistent
and systematic information. With this transparent view of the
whole landscape of knowledge available to the organization, a
pharmaceutical company can begin to redefine its best practice for
handling information: making informed scientific and business
decisions, and communicating with its key constituents– the
regulators, analysts and customers.
4. BUILDINGWEB SERVICE DOMAIN
ONTOLOGIES
Several Web service tasks can be automated by using semantic
descriptions. For example, service offers and requests can be
matched automatically [15]. This matchmaking is flexible
because it allows retrieving services that only partially match a
request but are still potentially interesting. For example, the hotel
booking service will be considered a match for a request for
Accommodation booking services, if the used domain ontology
specifies that Hotel is a kind of Accommodation. This
matchmaking is superior to the keyword search.
A basic requirement for being able to perform complex reasoning
with multiple semantic Web service descriptions is that (many of)
these descriptions should use concepts of the same (or a small
number of) domain ontology. If eachWeb service description uses
different domain ontology then a mapping between these
ontologies has to be performed before any reasoning task can take
place. However, ontology mapping itself is a difficult and largely
unsolved problem in the Semantic Web. Therefore, a quality
domain ontology will reflect a high percentage of the domain’s
terminology so that manyWeb services in that domain can be
described with its concepts. This requirement makes the building
of the domain ontologies difficult, as it is evident in the next
section where we present an analysis of the ontology building
process in two concrete research projects.
The creation of semanticWeb service descriptions is a time
consuming and complex task whose automation is desirable, as
signaled by many researchers in this field. This task can be
broken down in two smaller tasks. First, acquiring a suitable Web
service domain ontology is a prerequisite when creating semantic
Web service descriptions. This is required to create a shared
terminology for the semantic service descriptions. Second, the
actual process of annotating Web services with the concepts
provided by the domain ontology has to be performed. We
concentrate on theWeb service annotation task. Then we annotate
the elements of the ontology file with the concepts of the selected
ontology. In the rest of this section we describe the ontology
building process as it took place in the context of the actual
research projects. These projects offered realistic requirements,
data sets and evaluation standards for our work. In both cases we
detail the kind of data sources used for ontology building and the
resulting manually built domain ontologies. These manually built
ontologies serve as standards when evaluating the automatically
learned ontologies. We also highlight the difficulties encountered
during building these ontologies.
The domain ontology is composed of nodes and arcs among nodes.
Commonly, nodes are used to show physical entity, concept or
state, while arcs showing relations among them. For instance,
Figure 1. is a semantic network for describes “MY-CHAIR”.
In Fig.1, the parts above the node “MY-CHAIR” mean “MY
CHAIR IS A CHAIR”,“A CHAIR IS A KIND OF FURNITURE”,
“ SEAT IS A PART OF A CHAIR”; The left parts of “ MY-
CHAIR” mean “MY CHAIR BELONG TO ME”, “I AM A
PERSON”; The right parts of “MY CHAIR” mean “MY CHAIR
IS PALM BROWN”, “PALM BROWN IS A SORT OF
BROWN”; the parts below “MY-CHAIR” mean “MY CHAIR IS
COVERD WITH LEATHER”. “ISA” and “AKO” are normal
relations in semantic network. “ISA” which is read “is…an
example” means that a certain individual is one element in some
set, such as “MY CHAIR IS A CHAIR”; “AKO” is abbreviation
of A-KIND-OF, it denoting a set is a subset of another set, an
good example is “A CHAIR IS A KIND OF FURNITURE”. The
relations in Fig.2 namely “ISPART”, “OWNER”, “COLOR” and
“COVERING” denote attributes of node object. We can see from
Fig.2 that semantic network can describe and express knowledge-
information distinctly
Domain ontology is good for AKO, but both FURNITURE and
multi ones can be denoted by AKO. For instance, FURNITURE
(CHAIR) may be denoted by AKO (CHAIR, FURNITURE).
Multi ones may be changed into FURNITURE and AKO. For
example, the predication of “John gives a book to Mary” is
denoted logically:
(

x)[GIVE(JOHN, x, MARY) BOOK(∧ x)]
May be denoted by AKO:
ISA (G1, GIVING-EVENTS)
GIVER (G1, JOHN) RECIP (G1, MARY)
OBJECT (G1, BOOK1) ISA (BOOK1, BOOK)
Of course, more information may be added. For instance:
HUMAN (JOHN) HUMAN (MARY)
Fig.2 denotes corresponding semantic network.
Ontology building is difficult and should be automated. Our case
studies agree on a set of problematic factors that hampered the
ontology building activity. First, the ontology curators had to
analyze a high number of textual documents to ensure the quality
of their ontologies. The number of analyzed documents is likely
to increase as many domains witness a rapid increase in the
number of available Web services to several hundreds. A second
impediment was the lack of guidelines on what knowledge such
ontologies should contain and what design principles they should
follow. This resulted in different groups building different
ontologies to describe Web services in the same domain, as
reportedly happened in bioinformatics.
High quality domain ontologies are essential for successful
employment of semanticWeb services. However, their acquisition
is difficult and costly, thus hampering the development of this
field. The first stage of ontology building aims to develop (semi-)
automatic ontology learning tools in the context of Web services
that can support domain experts in the ontology-building task.
The goal of this first stagewas to get a better understanding of the
problem at hand and to determine which techniques might be
feasible to use. To this end, we developed a framework for (semi-)
automatic ontology learning from textual sources attached to Web
services. The framework exploits the fact that these sources are
expressed in a specific sublanguage, making them amenable to
automatic analysis. We implement two methods in this framework,
which differ in the complexity of the employed linguistic analysis.
We evaluate the methods in two different domains, verifying the
quality of the extracted ontologies against high quality hand-built
ontologies of these domains. Our evaluation lead to a set of
valuable conclusions on which further work can be based. First, it
appears that our method, while tailored for the Web services
context, might be applicable across different domains. Second, we
concluded that deeper linguistic analysis is likely to lead to better
results. Finally, the evaluation metrics indicate that good results
can be achieved using only relatively simple, off the shelf
techniques. Indeed, the novelty of our work is not in the used
natural language processing methods but rather in the way they
are put together in a generic framework specialized for the
context of Web services.
5. THE STOCHASTIC SYNTAX-PARSE
MODEL NAMED LSF OF KNOWLEDGE-
INFORMATION IN QAS
Local environment information is regarded as an important means
to WSD in sentence structure all along [16]. But in some lingual
models, which are assigned by probability on the basis of rules
traditionally, the probability of grammar-producing model is only
decided by non-terminal, while is independent of glossarial
example in analyzing tree. This quality of non-vocabulary makes
lingual phenomena description inadequate for probability model.
Therefore, QAS adopts the stochastic syntax-parse model named
LSF.
5.1 Model describing
Here, we describe a sort of basic probability depending model. It
is named lexical semantic frame (LSF for short [17]) in order to
be put easily. LSF is supposed as a result of character string
FURNITURE
AKO
CHAIR
ISPART
SEAT
ISA
MY-CHAIR
COLOR
TAN
OWNER
ME
PERSON
ISA
COVERING
AKO
LEATHER
BROWN
Figure 1. The semantic network of “chair”
GIVING-EVENTS
G1
ISA
BOOK1
OBJECT
GIVER
JOHN
MARY
RECIP
BOOK
ISA
HUMAN
ISA
ISA
Figure 2.The semantic network of the sentence
“John gives a book to Mary”
s=w
B
i
B
…w
B
j
B
, SR(R, h, w
B
i
B
) denotes that wi among LSF relies on the
word h through semantic relation, thus we can write down the
function SR(i)=SR(R, h, w
B
i)
B
. Analyzing semantic probability
p(SR(i)|h, w
B
i
B
)among words is on the basis of this model. The
model supposes that there exists high conjunction between
depending relation R and Hyponym node, the contradiction of
data sparsely is less. So we can give LSF the analyzing
probability from w
B
i
B
…w
B
j
B
:
P(LSF|w
B
i
B
…w
B
j
B
)=

≠=
j
hwik
k
,
P(SR(k)|h, w
B
k
B
) (1)
If we input S=w
B
1
B
…w
B
n
B
, the task of stochastic analyzer lies in
finding the best analysis T*:
T*=arg max
p
n
wParseT )(
1

(T|
n
w
1
) (2)

Among those, Parse (
n
w
1
) should be all possible structure-parses
for the input sentence, p (T,
n
w
1
) is defined as the product of all
used LSF probability in analysis. We make out that the
probability-parse model is based on Bi-vocabulary depending
relation.
5.2 Exercise the model parameters
Unlike rules probability model, the probability model parameter
based on vocabulary association is usually gained from
supervised training as well as using tagged corpus. In fact, The
reasons that we use both the words in corpus and their Hyponym
POS information to estimate P(LSF |w
B
i
B
…w
B
j
B
) are:
①vocabulary information plays a vital WSD role on
matching of depending frame.
②considering the limit to corpus scale, words repetition has
little probability in sentence analysis, we must deal with statistic
result smoothly [18]. Vocabulary information is needed to
“magnify” to reduce the degree of data sparseness with the help of
Hyponym part of speech. But the close word class such as
preposition or adverb uses statistic information of words.
Let W to be the set of whole vocabulary in training corpus,
SUBCAT to be the set of all Hyponym part of speech, CORPUS
to be training set, RS to be set of semantic depending relation,
LSF to be the semantic frame of vocabulary, we can define the
following functions:
0, SR(R, h, w
B
i
B
)

LSF
1, SR(R, h, w
B
i
B
)

LSF
(3)
Above CSR can denote the counting condition of depending
relation in corpus training.
Counting function C(<w
B
h
B
, sc
B
h
B
>, <w
B
c
B
, sc
B
c
B
>)denotes the occurring
times of upright depending relation (<w
B
h
B
, sc
B
h
B
>, <w
B
c
B
, sc
B
c
B
>) in
training corpus, and w
B
h
B
, w
B
c
B

W, sc
B
h
B
, sc
B
c
B

SUBCAT in that
function.

C(<w
B
h
B
, sc
B
h
B
>, <w
B
c
B
, sc
B
c
B
>)=



RSR
CORPUSLSF
CSR(SR(R, w
B
h
B
, w
B
c
B
),LSF)
(4)
Counting function C(R, <w
B
h
B
, sc
B
h
B
>, <w
B
c
B
, sc
B
c
B
>) tells the common
occurring times of depending relation R, headword w
B
h
BB

B
and
adjacent sub-node w
B
h
BB

B
in corpus.
C(R, <w
B
h
B
, sc
B
h
B
>, <w
B
c
B
, sc
B
c
B
>)=

∈CORPUSLSF
CSR(SR(R, w
B
h
B
, w
B
c
B
),LSF)
(5)
Function F(R|<w
B
h
B
, sc
B
h
B
>, <w
B
c
B
, sc
B
c
B
>) denotes the common
probability of depending relation R, headword w
B
h
B
and adjacent
sub-node w
B
c,
B

F
being its most likely estimate, then

F
(R|<w
B
h
B
, sc
B
h
B
>, <w
B
c
B
, sc
B
c
B
>)=
(6)


Finally, we can conclude:
P(LSF|w
B
i
B
…w
B
j
B
)=

≠=
j
hwik
k
,
P(SR(k)|h, w
B
k
B
)=

≠=
j
hwik
k
,

F
(R|<w
B
h
B
, sc
B
h
B
>, <w
B
c
B
, sc
B
c
B
>)
(7)
In above formula, we may use parameter smoothing technology
that is the Back-off –based method [12]. In analysis course,
dynamic scheming pruning process and probability computing
process are similar to rules probability model. If the analysis of
the two parts in one cell case having the same attribute structure,
then the analysis result of the part which has lower probability
will be cast aside and will not participate in the following
analyzing-combining process.
Supposing that we inputting a sentence in QAS: “She eats pizza
without anchovies”, now we have:
P(T
B
1
B
)=

)(
i
LSFP
=P(AGT|eat, she)P(OBJ|eat, pizza)P(MOD
| pizza, anchovies) (8)
P(T
B
2
B
)=

)(
i
LSFP
= P(AGT|eat, she)P(OBJ|eat,
pizza)P(MOD|eat, anchovies) (9)
Supposing that we can gain the correlative model parameter
through corpus statistics such as Table 1, then:
P(T
B
1
B
)=0.0025×0.002×0.003=1.5×10
P
-6
P

P(T
B
2
B
)=0.0025×0.002×0.0001=5×10
P
-8
P

T
B
1
B
may be chosen to be the right result according to this. If we
convert “anchovies” to “hesitation”, then P(T
B
1
B
)=5×10
P
-8
P
, P(T
B
2
B
)=4
×10
P
-7
P
. We find that language model may also help us to choose
sound analysis result with the change of words in sentence. This
is just about its merit.

CSR(SR,LSF)=
C(R, <w
B
h
B
, sc
B
h
B
>, <w
B
c
B
, sc
B
c
B
>)
C(<w
B
h
B
, sc
B
h
B
>, <w
B
c
B
, sc
B
c
B
>)
Table 1. Interrelated model parameters
P(AGT|eat, she) 0.0025
P(OBJ|eat, pizza) 0.002
P(MOD|pizza,
anchovies)
0.003
P(MOD|eat, anchovies) 0.0001
P(MOD|pizza, hesitation) 0.0001
P(MOD|eat, hesitation) 0.0008

6. SEMANTIC ANALYSIS IN QAS
Semantic analysis is the basis of QAS. We propose and construct
a frame-based Chinese Semantic Web Ontology. The hierarchy is
composed of semantic units and semantic rules. The semantic unit
is composed of semantic primitives, senses and semantic chunks.
The semantic primitive is the basic unit of the hierarchy and
describes all kinds of semantic features. The sense expresses
concepts and is described by primitives. The semantic chunk
expresses compound concepts and is decribed by nested frames.
In semantic frames, semantic relations of each component are
expressed by semantic roles. Semantic rules are the abstract of
semantic knowledge and are composed of well-formed semantic
links and generated chunk templets. Well-formed semantic links
describe the semantic restriction among semantic units. Generated
chunk templets describe the semantic structure that is generated
by well-formed semantic links. Based on the semantic hierarchy,
we design and implement an algorithm of semantic rule acquiring,
which acquires well-formed semantic links from the combination
instance corpus automatically and acquires generated chunk
templets semi-automatically.
Segmentation and tagging is the first step of Semantic analysis.
We design and implement an automatic semantic tagging
algorithm to tag sense for the result of segmentation. The
semantic tagging deals with monoseme, polyseme and unknown
word respectively and uses syntactic and semantic knowledge to
determine the sense tagging set for each word. The result can be
non-single and sense disambiguation can be done in the semantic
analysis. If the sense tagging set contains the correct sense,
semantic tagging hits the target. If the sense tagging set only
contains the correct sense, semantic tagging is correct. The
experiment is done and proves the algorithm can get a very high
hit rate and a higher correctness rate. Usually, the size of hit sets
will become rather small. The relation between Syntax and
semantic meaning is very close. Syntax is the form and semantic
meaning is the content. We combine syntactic parsing and
semantic analysis, use the extended context free grammar to
analyze Chinese. Each syntactic production has a precondition
judge function. When the analyzer needs to use a production to do
reducing, the precondition judge function corresponding to the
production is activated and in it the corresponding semantic
analysis is performed. The reduction can be done only if the
semantic analysis is reasonable, otherwise the current analysis
tree should be terminated. When reducing, syntactic structures
and semantic structures are generated together. So, semantic
analysis leads to the actions of the analyzer of the system.
Ambiguation is one main feature of natural languages. We
propose a semantic based disambiguation strategy to do word
sense disambiguation and syntactic structure disambiguation. At
the semantic tagging stage, word sense disambiguation is done
using well-formed semantic links. At the semantic analysis stage,
semantic rule matching and the determinant of semantic chunk
generating do word sense disambiguation and syntactic structure
disambiguation. The method can get satisfactory results.
Semantic Neural Network is a new methodology for Natural
Language Understanding that combines symbolism and
connectionism. It is considered that the Natural Language
Understanding is the process that the linguistic labels activate the
corresponding neuron in the human brain, and set up or activate
the Semantic Neural Network.But this research is just at the
beginning, researchers only provide the idea and its models. In the
other hand, Chinese understanding is always difficult. No matter
regular method or statistic method, none of them systematically
proposes an analytical method according to the characteristics of
Chinese. We find that the Semantic Neural Network relies on
semantic analysis mainly, which just accords with the
characteristics of Chinese as the language of analytical type.
Meanwhile, the neuron is connected according to semantic
relation between the concept, which is very good to solve the
difficult problems that the fuzzy on semantic and the
differentiation of morphology. The paper is come to simulate and
realize tentatively the model of Semantic Neural Network, and
combine with the semantic characteristic of Chinese to realize the
understanding of the surface semantics of Chinese. Semantic
Neural Network breaks the traditional linear understanding model
and simulates a mechanism for language processor of the human
brain; it is considered that the Natural Language Understanding is
the process that the linguistic labels activate the corresponding
neuron in the human brain, and set up or activate the Semantic
Neural Network. Deep-seated Semantic Computing of Semantic
Neural Network is that continue to use Deep-seated Semantic
knowledge embed in neuron, and through each other's
communication and inside computing to finish this processing.
The result of Deep-seated Semantic Computing in whole
Semantic Neural Network is a result of nature language
understanding.
7. EXPLAINING ANSWERS FROM THE
SEMANTIC WEB
Semantic Web aims to enable applications to generate portable
and distributed justifications for any answer they produce. Users
(humans and computer agents) need to decide when to trust
answers before they can use those answers with confidence. We
believe that the key to trust is understanding. Explanations of
knowledge provenance and derivation history can be used to
provide that understanding [19]. In one simple case, users retrieve
information from individual or multiple sources and they may
need knowledge provenance (e.g., source identification, source
recency, authoritativeness, etc.) before they decide to trust an
answer. Users may also obtain information from systems that
manipulate data and derive information that was implicit rather
than explicit. Users may need to inspect information contained in
the deductive proof trace that was used to derive implicit
information before they trust the system answer. Many times
proof traces are long and complex so users may need the proof
transformed (or abstracted) into something more understandable
that we call an explanation. Some users will decide to trust the
deductions if they know what reasoner was used to deduce
answers and what data sources were used in the proof. Other users
may need additional information including how an answer was
deduced before they will decide to trust the answer. Users may
also obtain information from hybrid and distributed systems and
they may need help integrating answers and solutions. As web
usage grows, a broader and more distributed array of information
services becomes available for use and the needs for explanations
that are portable, sharable, and reusable grows. Inference Web
addresses the issues of knowledge provenance with its registry
infrastructure called Semantic Web Ontology. It also addresses
the issues concerned with inspecting proofs and explanations with
its browser. It addresses the issues of explanations (proofs
transformed by rewrite rules for understandability) with its
language axioms and rewrite rules.
Every query-answering environment is a potential new context for
the Inference Web. We provide two motivating scenarios and use
the second scenario for our examples throughout the article.
Consider the situation where someone has analyzed a situation
previously and wants to retrieve this analysis. In order to present
the findings, the analyst may need to defend the conclusions by
exposing the reasoning path used along with the source of the
information. In order for the analyst to reuse the previous work,
s/he will also need to decide if the source information and
assumptions used previously are still valid.
InferenceWeb includes a new explanation dialogue component
that was motivated by usage observations. The goal is to present a
simple format that is a typical abstraction of useful information
supporting a conclusion. The current instantiation provides a
presentation of the question and answer, the ground facts on
which the answer depended, and an abstraction of the metal
information about those facts. There is also a follow-up action
option that allows users to browse the proof or explanation, obtain
the assumptions that were used, get more information about the
sources; provide input to the system, etc. Additionally all
information presented on any of the screens is “hot” and thus if
someone clicked on any explanation element, they could obtain
information about that element including its description and
information. This interface is expected to be the interface with
which the average human user of inference web interacts.
8. REALIZATION OF QAS
Our Automatic Question Answer System includes three models:
question’s semantic comprehension model based on Ontology and
Semantic Web, FAQ-based question similarity match model
(FAQ: frequently-asked question), document warehouse-base
automatic answer fetching model. The question’s semantic
comprehension model combines many natural language
processing techniques, including Ontology and Semantic Web,
Segmentation and Part-Of-Speech Tagging, the confirmation of
the question type, the extarction of keywords and extending, the
confirmation of the knowledge unit, Through these works, the
intention of the user is holded, which greatly helped the last work
of this system. The FAQ-based question similarity match model is
implemented by sematic sentence similarity computation, which
is improved by our system, this model can answer frequently-
asked question fastly and concisely. The document warehouse-
base automatic answer fetching model fistly deal with the
document warehouse beforehand and construct inversed index,
then use high efficient information retrieval model to search in the
base and return some relevant documents, lastly, we use answer
extraction technique to get the answer from these relevant
documents and present it to users. For the question that cannot be
answered by FAQ base, this model can automaticly return exact
answer fastly.
The document repository pre-processing module including Web
pages crawlering, HTML format filtering, segmentation and
Tagging etc. we receive a term-document matrix by computer the
word frequency. This matrix is then analyzed to derive our
particular latent semantic structure model for later document
retrieval and passage retrieval. Question analysis module is
important to QA system. Given a question, the system generates a
number of weighted rewrite strings. And then, transform the
query into a vector by those weighted rewrite strings. In this
module, lay emphasis on question classification. Systems
classifies a query into the predefined classes based on the type of
answer it is looking for, and then use the question types to
identify a candidate answer within the retrieved sentences.
Answer extraction module including: document retrieval, passage
retrieval and answer matching. System provides a varying method
to calculate weight and sort theanswer by the weight. Finally, the
answer been restricted within 50 words long and returned to user.
The QAS focuses on the key techniques of pattern knowledge
based question answering. We design and implement the question
answering system and take part in the evaluation of Text Retrieval
Conference. We also apply the pattern matching technique to a
new related research area Reading Comprehension, and a satisfied
result is acquired. The key task to implement the pattern matching
technique is to construct a perfect pattern knowledge base. We put
forward a novel question classifcation hierarchy that is based on
answer type and question pattern. It retains the semantic and
structured information of questions. We make use of the questions
on FAQ base as our training and test data. The answer patterns to
different question types are studied and evaluated automatically.
We have implemented pattern learning to questions with complex
structure. It is more effective and reliable to extract the correct
answer with answer patterns containing multiple question terms.
This cannot be covered by simple answer patterns. For higher
precision, we give semantic restriction to candidate answers that
are extracted by answer patterns. We adopt generalization
strategy to answer patterns using named entity information. It
makes the answer patterns have better extending ability; the
constituent elements of answer pattern contain both
morphological and semantic information with better robustness.
We evaluate all the answer patterns by the concept of Confidence
and Support, which are borrowed from data mining. Answer
patterns with higher confidence lead to choose the answer with
greater reliability.
Table 2 is the experimental results of QAS.
Table 2. Experimental results of QAS
number of
questions
answer
correctly
answer
mistakenly
no
responsion
accuracy /
%
recall /
%
2000 1641 198 161 82.05 91.95

9. EVALUATION AND CONCLUSION
An initial evaluation is performed on our QA system, focusing on
4 aspects: the feasibility to build such a semantic QA system
based on not traditional natural language text but Ontology, the
effectivity of personalized semantic QA, the extensibility of
ontology and knowledge base, and the possibility of self-produced
knowledge based on semantic relations in the ontology. The test
set includes 100 questions sampled from a set of questions asked
by the students in a one-semester programming lesson, excluding
the questions about reading a segment of program, writing a small
program to finish a function and so on, which is beyond the
ability of a QA system. At the same time, all these 100 questions
are ensured within the covering scale. For the scale of the initial
evaluation, we don't distinguish the situations between no answer
and a false answer. These two situations are regarded as the same
- no answer.
The initial evaluation result shows the feasibility of building a
semantic QA system based on Ontology and Semantic Web. The
personalized answering based on a user model benefits to
focusing the user's more attentions on fresh learning material. A
user can get the direct answers about some questions based on
semantic QA, which shows the effectivity of the system. In no
answere situation, the system takes a big proportion, which shows
the good extensibility, for the answer can be easily supplied into
the knowledge ontology without conflicting with the semantic
relations defined in the ontology. At last, the system takes a small
proportion, in which the ontology needs to be expanded and
ontology consistency must be ensured. How to prove the
possibility of self-produced knowledge based on semantic
relations in the ontology? A simple example is that the property
require is a transitive property, so if the fact that document A
requires document B and document B requires document C is
stated in the knowledge ontology, a new document relation,
document A requires document C, would be self-produced based
on the system inference. Afterwards, the update of inter-
dependency between documents would bring new answer for a
question. And experiments do prove that it is feasible to use the
method based on Ontology and Semantic Web to develop a
Question Answering System, which is valuable for further study
in more depth.
10. ACKNOWLEDGMENTS
We would like to acknowledge the support from the National
Natural Science Foundation of China (90412010, 70572090), the
National High Technology Research and Development Program
(863 Program in china: 2004AA1Z2450), HP Labs China under
“On line course organization”, NSCF Grant #60573166 and the
Ph.D Degree Teacher Foundation of North China Electric Power
University (H0585) in China.
11. REFERENCES
[1]
Voorhees, E.M. The TREC Question Answering Track.
Natural Language Engineering, 7, 4, (2006), 361-378.
[2]
UMLS. Unified Medical Language System. http:
www.nlm.nih.gov/research/umls/.
[3]
Laura, A., Slaughtera, Dagobert Soergelb, and Thomas, C.
Semantic representation of consumer questions and
physician answers. International Journal of Medical
Informatics (Winter 2006), 513—529.
[4]
Lindberg, D., Humphrey, B., and McCray, C. The Unified
Medical Language System, Meth. Inform. Med. 32 (2003)
281—289.
[5]
McCray, C., and Hole, W. The scope and structure of the
first version of the UMLS Semantic Network. Proc. Annu.
Symp.Comput. Appl. Med. Care (2006) 126—130.
[6]
W3C Semantic Web Activity. Semantic Web Activity
Statement, 2002. http://www.w3.org/sw.
[7]
Paolucci, M., Kawamura, K., and Sycara, K. Semantic
Matching of Web Services Capabilities. In I. Horrocks and J.
Hendler,ed itors, Proc.First International Semantic Web
Conference ISWC 2002,L NCS 2342, 279–291. Springer-
Verlag, (2006).
[8]
DAML-S Coalition. DAML-S: Web Services Description for
the Semantic Web.In I. Horrocks and J. Hendler,e ditors,
Proc. First International Semantic Web Conference ISWC L
NCS 2342, 311–319. Springer-Verlag, (2006).
[9]
Baader, F., McGuiness, D., and Schneider, P.P.The
Description Logic Handbook (May 2005). Cambridge
University Press, 199-216.
[10]
Decker,S. DAML Rules—An RDF Query, Inference and
Transformation Language, draft available online at
http://wwwdb.stanford.edu/∼stefan/daml/2001/07/03
/rules/damlrules.ps,2007.
[11]
Noy, N.F. et al. Protege-2000: an opensource ontology-
development and knowledgeacquisition environment. AMIA
Annu. Symp. Proc. (2006), 953
[12]
Lambrix, P. et al. Evaluation of ontology development tools
for bioinformatics. Bioinformatics 19, (2003), 1564–1571
[13]
Eilbeck, K. et al. The sequence ontology: a tool for the
unification of genome annotations. Genome Biol. 6, (2005) ,
R44
[14]
Stephen, P. Ontologies and semantic data integration. Drug
discovery today.10, 14, (2005), 1001-1008
[15]
Paolucci, M., Kawamura, T., and. Payne, T.R Semantic matching of
Web services capabilities. In Proceedings of the First International
Semantic Web Conference. (May 2002), Yale University Press, 206-
211.

[16]
Collins, M.A. new Statistical Parser Based on Bigram
Lexical De-pendencies.In Proceedings of the 44th Annual
Meeting of the ACL. (Washington, August 2006), 184~191.
[17]
Terje, B. and John, A.Natural language analysis for semantic
document modeling. Data and Knowledge Engineering, 42, 1,
January 2005, 45-62.
[18]
Shichao, A., and Mnhammed, J.Mining Multiple Data
Sources: Local Pattern Analysis. Data Mining and
Knowledge Discovery. Lecture Notes in Computer Science,
(Berlin Heidelberg New York: Springer-Verlag, August
2006) , 12, 8, 121-125.
[19]
McGuinness, D.L. Trusting answers on the web. in:M.T.
Maybury (Ed.), New Directions in Question Answering
(Oxford, August 2004).AAAI/MIT Press, 98-102.
.