The extraction of unstructured data from the Web and to analyse them in
order to determine useful information which can be used by customers and
manufacturers to make decisions about product is a challengeable task. There
are some existing techniques to evaluate products based on the ratings and
product reviews posted on the Web. However, all these techniques have some
inherent issues and limitations and therefore not able to fulfil the needs and
requirements of both customer and manufacturer. For instance, the existing
sentiment analysis methods (which classify the opinions in customer reviews
about a product as positive or negative) are not able to determine the context
of word in a sentence accurately. In addition, negation handling methods
adopted while determining the sentiment are not able to deal with all types
of negations and they also do not consider all exceptions where negations behave
differently. Similarly, the existing product reputation models are based
on single source, not robust to false and biased ratings, not able to reflect the
recent opinions, do not allow users to evaluate product on different criteria,
and also do not provide a good estimation accuracy. On the other hand, the
existing product reputation systems are centralized which have issues such as
single point of failure, easy to falsify evaluation information and not suitable
approach to solve a complex problem.
This thesis proposes methods and techniques for evaluating product reputation
based on data available on the Web and to provide valuable information
to customers and manufacturers for decision making. These methods
perform the following tasks: 1) extract product evaluation data from multiple
Web sources 2) analyse product reviews in order to determine that whether
opinions about product features in customer reviews are positive or negative,
3) computes different product reputation values while considering different
evaluation criteria, and 4) finally the results are provided to customers and
manufacturers in order to make decisions. This thesis contributes in three
main research areas i.e. 1) feature level sentiment analysis, 2) product reputation
model and 3) multiagent architecture. First, a word sense disambiguation
and negation handling methods are proposed in order to improve the
performance of feature level sentiment analysis. Second, a novel mathematical
model is proposed which computes several reputation values in order to
evaluate product based on different criteria. Finally, multiagent architecture
for review analysis and product evaluation is proposed.
Huge amount of the product evaluation data on the Web is in textual form
(i.e. product reviews). In order to analyse product reviews to evaluate product
we propose a feature level sentiment analysis method which determines
the opinions about different features of a product. A word sense disambiguation
method is introduced which identify the sense of words according to the context while determining the polarity. In addition, a negation handling method is proposed which determine the sequence of words affected by different types of negations. The results show that both word sense disambiguation
and negation handling methods improve the overall accuracy of feature level
sentiment analysis.
A multi-source product reputation model is proposed where informative,
robust and strategy proof aggregation methods are introduced to compute
different reputation values. Sources from which reviews are extracted may
not be creditable hence a source credibility measuring method is proposed in
order to avoid malicious web sources. In addition, suitable decay principles
for product reputation are also introduced in order to reflect the newest opinions
about product quickly. The model also considers several parameters such
as reviewer expertise, rating trustworthiness, time span of ratings, reviewer
age, sex and location in order to evaluate product in different ways. Different
types of ratings (i.e. textual and numeric ratings) are considered to compute
reputation values which increase the choices for customers and manufacturers
to make decisions. The results show that the proposed model is robust,
strategy proof, able to reflect recent opinions, and estimates true reputation
values even if some ratings are false.
A multiagent layered architecture is proposed for product reputation evaluation.
The main idea behind this layered architecture is to divide the complex
problem of the product evaluation which is handled by a single entity in
a centralized fashion into simpler and smaller problems handled by several
entities in a distributed fashion. The architecture addresses different aspects
of product evaluation such as taking inputs and displaying results to users, reviews
extraction, feature level sentiment analysis and computing reputation
values. In addition, the architecture also addresses issues concerned with
centralized approach and also offers additional benefits such as autonomy,
pro-activeness, openness and social ability.