Mining Text Conversations - Methods for Mining and Summarizing Text Conversations - page 56

￿ Online Sentiment Classification Companies, political parties and other institutions are often
very interested in what people think of their “product.” One way of soliciting feedback and
opinions is to carry out a small-scale focus group using specially selected individuals. An
attractive alternative is to mine the web, analyzing the opinions of thousands or millions of
people who are discussing the products online. These discussions could be reviewer comments
on a site such as Amazon.com, or more informal conversations on blogs and in discussion fora.
The needs of the institution will determine the level of sentiment granularity required, but
often the result will simply be an indication of whether opinions are favourable, unfavorable
or mixed- whether the polarity is positive or negative—and how strong those opinions are.

polarity

￿ Document Sentiment Classification At times, we want to know whether a single document,
such as a product or movie review, is positive or negative. That is, we want to know the polarity
of the document. This will also often involve a rating of the polarity strength. Did the reviewer
hate the movie, or simply find it boring? The former would be a stronger negative polarity
than the latter. This polarity rating will necessarily be a simplification, and sometimes a very
crude one. For example, if the reviewer loved the movie script but hated the acting, should we
give a “neutral” polarity rating? In any case, deriving a polarity rating will entail losing some
information, but note that reviewers themselves often assign a star-rating to the review they
authored. Such simplification can be useful and is a key part of sentiment analysis 6 .

￿ Sentence Sentiment Classification We may want to dig deeper and identify all the points
where somebody was expressing an opinion. This could involve classifying every sentence in
a document as subjective or non-subjective. Given the transcript of a meeting where people
were discussing the product they were designing, we could run a sentiment classifier over the
data and quickly see all the opinions the people involved had about the product, favourable
or not. Similarly, we could run a polarity classifier over the data that would allow us to extract
just the positive or just the negative opinions.

￿ Facet Sentiment Classification Knowing that a person likes or dislikes something is often
not enough. We want to determine the reasons behind the opinions they hold. Given a product
or some other entity, we can identify facets of the product and correlate opinions to each facets

individual facet. For example, given the same meeting transcript described above, we could
identify facets of the product they are describing such as the size, weight, color and interface of
a remote control they are designing, and then describe what people think about each of these
specific facets.

There are several dimensions we can use to describe work on subjectivity and sentiment
detection. Like most research in text mining and natural language processing, there are supervised
and unsupervised approaches. Most of the tasks listed above lend themselves well to supervised supervised

vs. unsu-
pervised

6 Indeed, all tasks described in this topic involve simplification of some kind.