What's the Use of Sentiment Analysis at the Document Level? Not Much, without Topic-Level Analysis

Quite often, documents express opinions, like web reviews or open-ended questions in market surveys. And most of the times documents don’t just represent a single point of view, a single opinion. Documents typically consist of multiple opinions representing several closely related but nuanced positions. By reducing a whole document to a single opinion score, many sentiment analysis services hide the best insights, the useful ones, and prevent clients from drilling down to extract the useful information.

In contrast, at Bitext we detect every opinion in every sentence for every document, and we analyze it separately to ensure that the value of the opinion is maximized. If a person writes

“this Android I've got is awful, my iPhone was so sleek, I loved it!”

our Deep Linguistic Analysis engine detects 3 opinions being expressed and we will return a separate score for each one. And most important: it identifies for each opinion:

  • the WHAT, the sentiment topic
  • the WHY, the sentiment text

 

Check the JSON output of our API:

"text":"this Android I've got is awful, my iPhone was so sleek,
I loved it",
"global_value":2.000000,
"details": [
{
"valuables": "Android",
"valuers": "awful",
"value":-4.000000
},
{
"valuables": "iPhone",
"valuers": "sleek",
"value":3.000000
},
{
"valuables": "iPhone",
"valuers": "loved",
"value":3.000000

}

#opinion 1
#sentiment topic 1
#sentiment text 1

#opinion 2
#sentiment topic 2
#sentiment text 2

#opinion 3
#sentiment topic 3
#sentiment text 3

And another example where the Sentiment Topic is a concept rather than a brand:

"text":"The room was kind of dirty but it was
really peaceful and quiet.",
"global_value":5.000000,
"details": [
{
"valuables": "room",
"valuers": "dirty",
"value":-2.000000
},
{
"valuables": "room",
"valuers": "really,peaceful,quiet",
"value":7.000000
}

#opinion 1
#sentiment topic 1
#sentiment text 1

#opinion 2
#sentiment topic 2
#sentiment text 2

Now, once we have analyzed low-level data, at the opinion level, we can move up and compute a score for each sentence, paragraph of for the full document. If this high-level score is based on low-leveldata, then it will be helpful: we will be able to decompose it, categorize the different opinions in clusters...

 Try our API now

Subscribe Here!