23 /342
臺大管理論叢
第
27
卷第
2S
期
23
chance of success for a social media listening project. Second, in addition to opinion
sentence identification, many essential opinion mining tasks remain (e.g., opinion orientation
determination and opinion searching). The proposed semi-supervised learning concept can
be easily extended to support other opinion mining tasks.
Nonetheless, several limitations of this study must be acknowledged. First, only one
data set is employed to evaluate the performance of the R-OSI technique. To increase the
reliability of the evaluation results, it is essential to collect additional data sets from diverse
sources. Second, the traditional association rule mining algorithm with minimum support and
minimum confidence as interest measures is adopted to learn the opinion sentence
identification rules. Some novel interest measures or extended association rule mining
algorithms are proposed in the literature. Extending the R-OSI technique to novel designs
represents a compelling future research direction. Finally, only the task of opinion sentence
identification was addressed in this study. The concept of the R-OSI technique must be
applied to other vital opinion mining tasks such as opinion orientation determination.
5. Originality/Contribution
Most relevant studies have adopted either a supervised learning or an unsupervised
learning approach for the target opinion sentence identification task. A supervised learning
approach generally demonstrates higher performance but requires manual preparation of
training data. By contrast, an unsupervised learning approach requires no time- or labor-
consuming process for preparing training data at the cost of lower identification
effectiveness. Therefore, this study proposes a semi-supervised learning approach. The
R-OSI technique can achieve comparable or even higher performance than that from using a
supervised learning approach without the need for preparing training data. Only some (one to
three in this study) keywords are provided by the user. The cognitive effort of keyword
preparation should be lower than that of training data preparation. This represents the major
contribution and novelty of this study to the literature on opinion mining. Furthermore, the
R-OSI technique has additional advantages. First, it can identify opinion sentences with
implicit product features. Second, the R-OSI technique can group opinion sentences with
different product feature synonyms into the same category.