Measuring Phenomena in Sentiment

I’m currently reading Logan Dillard’s Masters thesis on sentiment mining titled: “I Can’t Recommend This Paper Highly Enough”: Valence-Shifted Sentences in Sentiment Classification. It’s an interesting read, approaching the space in exactly the right way: looking for types, measuring their distribution and then using that information to improve existing methods. One of the key observations is captured in the table below. 15% of the sentences in the corpus studied contains valence shifters. Of those, 49.4% were valence shifters affecting verb phrases.

VS Type	Corpus	VS Sentences
negated verb phrase	7.4%	49.4%
negated noun phrase	3.9%	25.9%
negated adj phrase	2.9%	19.3%
modals	1.2%	8.1%
total	15.0%	100.%

In addition, it was observed that valence shifters are more often used in negative expressions than positive.

Note: valence shifters are lexical items which later (generally invert) the appraisal orientation of an expression; for example ‘not’ in ‘not good’.

Comments

[Hannes Carl Meyer]

Maybe a bit offtopic, but I had a talk with a researcher from Fraunhofer IAIS yesterday on this topic. In their sentiment mining project which grabs data from social communities, they also look at emoticons (similies etc.) to boost a sentiment. I found it a great idea!