This time we have focused on whether those services can provide not only polarity for a given sentence, but also the topic which the opinion applies to. This level of detail is needed to provide granularity in sentiment analysis and it is essential to be able to extract useful information from your Sentiment services.
Current example is
“this Android is awful, my iPhone was so sleek”.
This example includes two opinions within a single sentence: a negative opinion about Android, and a positive opinion about iPhone. The key in this example is the fact that the qualifying adjectives are not next to the nouns they modify. In other words, syntactic analysis is required in order to ascertain that “awful” is modifying “Android”, whereas “sleek” is modifying “iPhone”. In addition, there is an adverb, “so”, which intensifies the meaning of “sleek”.
What is the result of the analysis for this sentence in the different API Demos?
The main trends we have detected:
- Most APIs don’t seem to carry out a syntactic analysis and, therefore, they miss the identification of the opinion topics.
- Because of that, some APIs even make a mistake of assigning wrong polarities to “Android” or “iPhone”
In short, if an API is not able to detect the topic of every opinion in the sentence, it will end up mixing the sentiment for all topics present in the sentence, thus misleading the user about the information he should receive.