Sentiment analysis involves evaluating a piece of writing and determining whether it’s positive, negative or neutral. Sentiment analysis is devised to gauge how we feel about a certain topic. For instance, you can determine what users in Mumbai on Twitter feel about the weather in the city. You can also dive deeper into the data by extracting exact words or phrases to understand what they liked or disliked about the weather in the city.
For a brand, this data is valuable as it allows them to take corrective actions and they can even use feedback to develop better products and services.
Sentiment analysis: Understanding context and tonality
Sentiment analysis can be tricky considering humans are fairly intuitive when it comes to understanding the tone of writing while algorithms may not be able to spot errors.
Consider this sentence for example:
My flight has been delayed. Brilliant!
Researchers believe that most sentiment prediction systems fail because they look at words in isolation and assign positive points to certain words and negative to others and then summing the total of this points. Let us clarify this point by showcasing an example.
The movie was neighther funny, nor super witty.
So while the words funny and witty are positive. The above sentence is still negative overall. Researchers at Stanford are trying to solve this problem by building a new sentiment prediction framework which used Recursive Neural Network. The advantage of using Recursive Neural Network is that it sets out weights recursively over a structure, to produce a structured prediction over variable-size input structures. To enhance the model further Stanford team is building a sentiment treebank (dataset) on which the model will be trained. The initial research has shown significant improvement in the overall accuracy of labeling sentences.
The team at Brandwatch believes the current labeling system of rating opinions as positive, negative and neutral might not be very effective in future. Real emotions could also be that of skepticism, hope, anxiety, excitement and anger. Brands and publishers, therefore, have to accept that sentiment analysis is not an accurate measure to gauge what people think about your brand or topics.
Google recently launched Syntaxnet which has been a significant milestone when it comes to sentiment analysis. Syntaxnet has a library of tools that uses Google Natural Language Processing capabilities to understand a word’s syntactic function and the relationship between words used in a sentence.
The process of understanding the words in a sentence and their relationship to each other is termed Parsing. Tools built to complete this task are called Parsers.
Sentiment analysis always had issues when it came to analyzing sentences that compared two or more objects. Take the following sentence.
Indigo is better than Spicejet.
While if you are from Indigo, you will label the sentence as positive due to the positive sentiment shifter (better). If you work for SpiceJet, then you would mark the sentence as negative.
Social listening tools have difficulty in dealing with sentences with multiple comparisons. Take the following sentence.
Indigo is better than Spicejet but they’re both worse compared to Vistara.
Most listening tools would label first part of the sentence as positive (better) and the second part of the sentence as negative (worse) and mark the overall sentence as neutral. Whenever the algorithm can make out the exact sentiment of a sentence, it labels it as neutral.
A parser will enhance sentiment analysis by identifying nouns, verbs, and sentence shifters. Parsers will improve the overall accuracy of sentiment analysis and reduce the number of neutrals. Ben Donkor has written a more detailed write-up about the use of parsers which should be helpful in understanding this technology.
How can marketers make use of sentiment analysis?
Most brands today use social listening tools to conduct sentiment analysis. Sentiment analysis enables you to track and respond to conversations. It’s important to note that no listening tool available in the market today tracks every property on the web. Also, it is not practically possible to scan every conversation around your brand.
Beyond social listening tools, you can also integrate sentiment analysis into the application or tools you are building brands using APIs. Google provides access to its Natural Language API to analyze sentiments. Similarly, Lexalytics also provides an API that integrates with your applications and tools.
With all the innovations happening in the space social analysts are certainly going to have a great time in the days to come.