(...) If topics can be automatically classified in a relatively simple way, sentiment analysis (positive, negative or neutral) gets harder as it requires a global analysis of the monitored text. This may be particularly difficult in cases of coordination between different parts of a sentence, of an anaphora or of a coreference (the recovery of an argument later in this document).
Another difficulty of natural language for automatic analysis of sentiments are intentional contexts, for which the expression of opinion is not a true feeling. Thus, the phrase" I thought the trams were more modern" may be identified as positive when the intention is, however, negative. If the analysis of conversations are established by "word packets", the sentences "I like this model not only because of ..." and "I did not like this model only because of ... "a real problem. They are composed of the same packet of words but represent an opposite feeling. Lucas Dini et al. "Showed the relationship between syntactic and semantic structures of a sentence and the expression of the opinion that vehicle".
The algorithm must be able to differentiate a neutral comment (eg: this car announces a fuel consumption of 3.3 l/100km) from an opinion (eg: this car announces a fuel consumption of only 3.3 l/100km!). It should also detect spelling errors, simplified syntaxes, slang or smileys. Finally it may be that a comment has positive and negative elements. Are they equal in value or is it that the positives outweigh? The phrase "I think this car is the best in its segment ... but also the most expensive "is an example of such ambiguity. What did the author say ? It is normal that its price is more high because of its beauty or the contrary, the price is not worth the candle? Even a human would have difficulty assigning a tone here. (...)
Source : Master Thesis "Mobility sector in Belgium : social media monitoring to built an efficient marketing strategy" 2011, ULB.