The State of Sentiment Analysis

??????????????????????????????????????????????Sentiment analysis is under deep scrutiny. After the initial enthusiasm about the concept and the flood of many tools trying to measure sentiment, many in the web analytics space are disappointed and left searching for better ways to measure and apply it to business problems.

In the recent Social Media & Web Analytics conference organized by Innovation Enterprise in San Francisco, with attendees from big tech companies such as LinkedIn, Twitter, and Yahoo, there was a clear consensus that Sentiment Analysis is problematic and inaccurate when left to machines counting positive and negatives comments. Human language is too complex to reduce it solely to algorithms.

Emerging terms, negation, sarcasm, word roots (stemming), sentence completeness, term frequency, semantic ambiguity, lack of context, and differences in topic categories, among others, are issues undermining the accuracy of the sentiment metrics currently used, so humans are brought to verify and correct the models used to support machine learning.

Some users of sentiment analysis are trying to go beyond simple counts of positive, negative, and neutral comments. They want to measure how strong the sentiment is using rating scales, while others try to identify human emotions expressed in those comments.

Due to the large amounts of unstructured data that is being generated by social media and customer feedback systems, among other sources, algorithms are still needed, and the data scientist community is still searching for the best ones. More text mining tools (e.g., OdinText, Bitext, MonkeyLearn, Luminoso, etc.) are coming to the text analytics space racing to offer a solution, but some companies have opted for developing their own proprietary tools as they don’t find any that satisfy all their needs or are too expensive.

Aware of the limitations of this metric, some companies are trying to use sentiment as one more variable to consider together with others in the context of the business problems they face.  Many examples were presented showing that this metric could be used to:

  • Increase customer retention
  • Resolve customer experience pain points
  • Identify what customers like
  • Optimize customer service by matching customer service representatives with customer issues they are best prepared to handle
  • Optimize pricing
  • Measure social media ROI

My hope is that the frustration with the current state of sentiment analysis keeps driving the development of better tools that would allow extracting insights from unstructured data faster and in a more cost-effective way. They can become powerful tools in the hands of market researchers when faced with analysis of qualitative data and open-ended survey questions.

Michaela Mora is the president of Relevant Insights, LLC, seasoned in custom market research with more than 20 years of experience in industries such online subscription services, software, entertainment, offline and online retailing, automotive, travel, hospitality, consumer packaged goods, non-profit, insurance, and beverage among others.


Speak Your Mind