Opinion mining:
Goal of opinion mining is to identify the textual parts that express
emotions. In other words it is Sentiment analysis. Application of
opinion mining comes under the decision making process. It converts people's voice into a statistic table so that it will be useful for entrepreneur.
Nowadays the area of sentiment analysis is flourishing
with lots
of research activities.
Relevance
As
per survey 81% of Internet users are surfing for their product
research (for restaurants, hotels, and various services) at least
once. Among those between 73% and 87% report that, reviews had a
great role on their purchase. Consumers are ready to pay more for a
higher-rated item than a lower-rated. Online ratings systems provide
32% reviews and 30% of them have posted as online comments or
reviews.
Due
to the diversity of sources it is not easy to get all the review
contexts from Web. In some cases it requires authentication, some
other cases opinions are hidden along with forum posts and blogs. It
is very difficult for a human reader to go through relevant sources,
collect contexts, extract pertinent sentences, analyze, summarize,
classify and organize them into a usable form. In this situation an
automated opinion mining tool will be a help desk for a customer.
Early
work
How
do people think about..? Researchers are trying to address this
question using opinion mining.
Identifies the
polarity of opinion words and document-level positive or negative
sentiment classification are some of the initial work has done in
this area. In fact this is not exactly needed for a feature based
opinion mining process. For example let’s take a review on a phone.
A customer might like its screen but dislike its battery. Then
researchers started working on feature based opinion mining which
mined opinions on different product features. This task is known as
feature-level opinion mining.
Challenges
Let’s
see some challenges that we have faced during our OM process.
- Figure out the proper linking between emotions and its topic is really a thought-provoking task.
For
example: “I'm looking for a good twitter app for my apple ipad”.
Here
there are 2 possible heads (twitter app, apple ipad) and a single
adjective (good). Proper Linking should be between “good” and
“twitter app”. Like “good twitter app”.
- Find out the emotions from sarcastic sentences is not easy, cases like, if the sentence has some sarcastic meaning or else if it needs an external knowledge to define the emotion.
Sometimes
people may make comments sarcastically, either by putting some
sarcastic smileys or by having some sarcastic meaning.
For
example: “I like their product verrrrrrry much ….. ;) ;)”. It
may be a sarcastic review. To determine this we need some external
knowledge regarding this person or his/her previous comments.
- In some other case topic might be in the previous sentence and referring that using some pronouns such as “it, he, and they” etc...
Look
at this example,
“I
showed it to Tom and Mary. He also liked”
Here
“He also liked” is the opinion part, normally the head taken as
“He”. But here the actual head is Tom. Pronouns are not proper
heads.
- Nowadays people using shorthand such as “U” instead of “You”, smileys etc...
“I
lve ma ipad ”
People are widely using shorthand to make comments. It is very
difficult to resolve these shorthand words, like
“lve”
= “love”, “ma” = “my” and “2moro” = “tomorrow”
etc…
- In some scenarios combination of some words can create some emotions,
For
example: damn beauty
Here
“damn” is a negative emotion and “beauty” conveys a positive
emotion and “damn beauty” is a positive emotion. Other example is
“deep shit”
- How to get the data? We can pull out only 20-30% of user reviews from World Wide Web using some connectors to the social media websites such as Twitter, Facebook, YouTube and DIGG etc…
What
are the available methods?
Basically
there are 2 methods, Supervised and Unsupervised. We get more
accurate results by using the Machine learning approach (Supervised),
but the challenge is to get the training data and also its scope is
always limited. Languages and its usages are very flexible. Even if
we made some training sets it will be outdated soon. There are lot of
tools are available for Topic Extraction and Sentiment analysis. Some
of them are listed below.
Tools
There
are some tools available for this purpose, like KEA, MAHOUT, MALLET,
MAUI, WEKA, SmILE, SentiWordNet
and RapidMiner.
Almost
all the topic extraction tools are based on machine learning. It is
useful for document level extractions and classifications. This is
not what we exactly needed for feature level opinion mining.
SentiWordNet is a good one for finding the emotion of a word.
Development frameworks
To develop such kind of applications there are some development frameworks like GATE, UIMA, and NLTK etc… According to the use and development criteria we can choose any one of these. These all are open source tools. It allows different types of plugins that are useful for this type of tasks.
Mining
process (Unsupervised)
Every
opinion has at least two parts a Head (Topic) and a Sentiword (the
word describes emotion).
For
the proper identification of Opinion parts (Head and Sentiword), an
excellent POS Tagger and Gazetteers (list of commonly used nouns,
phrases, sentiwords and smileys) are needed. Topic can be a Person,
an Object or a Term and Sentiwords are basically categorized into
Positive and Negative. Linking
of a sentiword to a proper head is based on some constraints that we
have given. Opinion text should be understandable and meaningful. A
feature based classification of opinions is an added task for opinion
mining.
Accuracy
How
can we measure the accuracy of an Opinion mining application? While
doing Opinion mining process, agreement between humans is around 85%
only, using some sort of training we can make it above 90%. But
agreement between human and system is pretty much lesser than this.
Measurement of this can be done by precision and recall.
Correlation
can give the closeness towards the predicted value.
Benchmarking tools are available for such type of measurements.
Uses
Today
it has a wide range of applications like Brand Monitoring, Buzz
Monitoring, Online Anthropology and Online Consumer Intelligence. In
other words, say social media monitoring. Opinion mining helps us in
decision making process. It is useful for individual as well as
organization. Summarization of opinions makes consumer to take
informed as well as valid decisions. Opinion
mining applications are becoming as the essential part of businesses
and organizations. For example, it is always critical information for
a product manufacturer “how consumers accept their products” and
those of its competitors. This information is not only useful for
marketing and product but also useful for product design and product
developments.
External
References
This is a challenging area definitely..There can be a matrix which leads to a final decision about the emotion in case of sarcastic inputs. Matrix for smileys, matrix for sarcastic words and expressions . Hopefully the sentence can then be processed using each matrix and normalised.Again a normailised value against emotion matrix is to be formed and then later if there is very minute variation it can be used for manual decision. This manual decision then can be used in future to automatically map such behaviours to the automatic emotions.But again the varying input terminologies and the usages will be again a matrix which needs to be updated everytime..Just some thoughts on initial understanding.
ReplyDeleteHi Vineeth, Yes you are right, sarcasm can redirect to manual decision. But identifying sarcasm is the bottle neck here. I am not saying it is not possible. But it will always has exceptions, because the sarcastic element directly links to the behavior (or thought or perception or experience) of the writer and the target (time is also one dimension). To understand this we have to track a set of comments posted from his/her side and comments about the target. We need to scratch this portion a bit more for the real matrix. Currently some techniques has been proposed. University of sheffield recommended some method to identify this like they are getting the tags #sarcasm from twitter and treating them as sarcastic comments. :)
Delete