Locowise looks at the sentiment of comments left by fans on your content. For Facebook, our model also takes into account Reactions and negative engagements (hide post, unlike post etc.).
Locowise extracts sentiment from comments using the VADER method. It analyzes them using a modified version of VADER to determine where they fall on the spectrum of negative, neutral, and positive.
For a detailed explanation of how this works, please refer to the following resource: http://comp.social.gatech.edu/papers/icwsm14.vader.hutto.pdf
What is VADER?
VADER is a method of sentiment analysis which is specifically attuned to social media. The challenge with social media is that you are often dealing with short snippets of text as well as slang and abbreviated language – this was given special consideration during the development of the method.
How accurate is it?
In the study, VADER matched or outperformed 11 other highly respected sentiment analysis methods.
The VADER lexicon performs exceptionally well in the social media domain. The correlation coefficient shows that VADER (r = 0.881) performs as well as individual human raters (r = 0.888).
Essentially, this means that the model has an accuracy of 88%. By making our own enhancements, we have managed to push this into the 90% and up range. Our system has also been modified to incorporate emojis.
How does VADER work?
First you start with the VADER lexicon. Typical lexicons consist of a large library of words which are classed as negative, neutral or positive. What makes VADER so unique is that it also includes acronyms (LOL, WTF etc.), slang (meh, nah etc.) and emoticons, which make it particularly effective in a social media context.
There is also an ‘intensity’ assigned to each object in the lexicon. For example, “excellent” is given a higher positivity rating than “good”. The classifications and intensities were assigned by a group of individual human raters, with the final rating being the average of their individual scores.
On top of ratings assigned to individual words, the developers then added an extra layer of depth by determining a set of grammatical rules that affect the overall sentiment of a string of words. These rules drastically improve the accuracy when analyzing snippets of text, causing the model’s ratings to fit more closely with the scores assigned by the human raters. The rules include:
1. Punctuation, namely the exclamation point, increases the intensity. For example, “The food here is good!!!” is more intense than “The food here is good.”
2. Capitalization increases the intensity. For example, “The food here is GREAT!” conveys more intensity than “The food here is great!”
3. Degree modifiers (also called intensifiers, booster words, or degree adverbs) impact sentiment intensity by either increasing or decreasing the intensity. For example, “The service here is extremely good” is more intense than “The service here is good”, whereas “The service here is marginally good” reduces the intensity.
4. The word “but” signals a shift in sentiment, with the sentiment of the second part of the sentence being dominant. “The food here is great, but the service is horrible” has mixed sentiment, with the latter half dictating the overall rating.
5. The model catches nearly 90% of cases where negation flips the sentiment of the text. A negated sentence would be “The food here isn’t really all that great”.
What languages are supported?
Currently, the model supports the following languages (with more being added):