Sentiment Analysis & Social Data — Picking the Right Tool for the Job

Chris Moody of Gnip

We were recently at the Sentiment Analysis Symposium with many of the leaders in the sentiment space. Conference presenters spanned a spectrum: academics, engineers and executives involved in research and developing sentiment analysis tools. The event was full of great content, and we wanted to share some takeaways.

It was impressive to see the variety of use cases and solutions that were presented — both signs of a rapidly maturing market.  Because of that variety, it was also clear that a one-size-fits-all solution for sentiment analysis of social data is not on the immediate horizon.  An overarching message from the event was that the different use cases for sentiment analysis and the types of social content being analyzed demand tailored sentiment tools purpose-built for the job.

I was also struck by just how central social data was to the conversation and how most of the speakers talked about analyzing social data in some form.  Sentiment analysis has clearly become important to social data analysis, and vice versa, social data is really attractive to companies and researchers seeking ways to understand the moods and opinions of broad populations of consumers and citizens.

Here are some of the highlights of the business cases that were presented at the conference:

  • Mood Analysis Informs Market Decisions at ThomsonReuters — Scanning a broad swath of social content and news allows ThomsonReuters to build their MarketPysch Indices, including their “Gloom Index,” which they’ve demonstrated provides a leading indicator for market drops. ThomsonReuters’ Aleksander Sobczyk made an interesting point for our readers: the social content they focus on is long form only (i.e. blogs), and not short form content like Twitter.  Developing the depth of insight and certainty essential to their current solution requires a bigger text sample than Twitter provides.
  • Guiding Product Development at Dell — Through its Social Media Listening Command Center, Dell picks up on customer sentiment and uses it to drill in and learn more. For instance, they picked up negative sentiment about a new Alienware laptop right after a release: units were overheating. The negative sentiment tipped them off, and their product team took action to learn more from users. Through engaging their customers online, they were able to discover an idiosyncrasy in laptop use among power users. When plugged into big auxiliary monitors, users were keeping the laptop screen nearly closed, blocking the exhaust fans. Dell fixed the fan placement and eliminated the problem for future customers. Knowing very specifically what the negative sentiment was about was critical to catching a problem spot among loyal fans and power users.
  • Understanding Weibo Emotions (了解微博的情感) — Detecting sentiment in other languages with different linguistic structures and use habits can be difficult, which is why Soshio has tried to normalize around universal emotions to help Western brands make sense of what Chinese weibo (microblog) users are saying about their brands.  According to Ken Hu, Soshio’s founder, they’re actually having to translate written language “into sound” to address sentiment challenges like written puns that only make sense when pronounced audibly.

These are just a sample of the stories we heard from the companies at the event, and what we hear from our customers building sentiment tools.  Some serve customers who need to know about general sentiment (i.e. positive or negative), while are others are interested in specific moods or opinions. Some support tailoring to a specific cultural context or industry vertical, while others help evaluate conversation in a broad, general context.  A few are working hard to analyze sentiment in a global context — including native natural language processing for different languages, each with their own norms related to sarcasm, puns, irony, etc. Content format is a major focus as well. Evaluating 140 characters rich with hashtags and URLs on Twitter requires a much different approach than evaluating a WordPress blog post. These are all valid use cases for sentiment analysis, and all come with unique nuances and benefit from different techniques.

Sentiment analysis is a critical category of tool for broader social data analysis, but at the same time, it’s not a monolithic category. If you’re building a product to analyze social data and want to evaluate sentiment, the good news is that you can pick from a wide range of approaches or tools (or maybe even choose more than one?) for sentiment analysis that is suited to the business problem you’re trying to address, and the content you’re trying to analyze.

If you’re interested to learn more about the event and some other interesting developments and business cases in the sentiment analysis world, check out our recent interview with Seth Grimes, the lead organizer for the Sentiment Analysis Symposium.

Data Story: Interview with Seth Grimes of the Sentiment Symposium

Seth Grimes is an industry analyst and consultant who specializes in data analysis technologies. He also organizes the Sentiment Analysis Symposium, which Gnip is pleased to sponsor. You can meet Gnip, including President & COO Chris Moody, at the next symposium, October 30 in San Francisco. If you register, you can use the code GNIP for a $200 discount!

Seth Grimes of Sentiment Symposium

1. How do sentiment, and sentiment analysis, differ from industry to industry? Sentiment does differ from industry to industry! Thin sheets and walls in a hotel are bad, but a thin mobile phone or tablet is good. Humans understand the difference; after all, it’s people who generate the content we’re mining, the flood of reviews, tweets, status updates, and messages. Our task as business people and technologists is to automate understanding. We need to make social, online, and survey analyses systematic and reliable and to align analyses with business goals. You do need sentiment capabilties that account for differences from industry to industry. Solutions should understand differences between hospitality and consumer electronics, and also healthcare and politics. But we also need to get at cultural and demographic factors behind the data, in order to completely understand what people are saying, feeling, reposting, and planning.

2. You’ve talked about how “good, clean, comprehensive social data is central for any meaningful social-media analytics initiative.” What are other considerations people should have when employing social data for sentiment? “Good, clean, and comprehensive” is a start; continue the list with relevant and useful. It’s easy to get overwhelmed by the social-data flood, but much, even most, of what’s being said on-line and on-social has little business value. You have to filter out the noise, by which I mean, in this context, data you can’t use. For instance, one of my friends just tweeted, “Let’s Stop Demonizing [Do It Yourself]. If Home Depot didn’t put contractors out of business,… why are we so worried about DIY [market] research?” I’d bet that Home Depot picked up that tweet, but I sure hope they ignored it. But if you’re following market research thought leadership, you want to pay attention. Of course, good, relevant data isn’t enough. You need analytical tools that will help you make sense of it and communicate usable business insights. Sentiment analysis plays that role, as a key element in larger social and enterprise analytics programs, applied for customer service and support, market research, media analysis, clinical medicine, and a host of other applications.

3. What do you think the biggest changes in sentiment have been in the last few years, and how do you see it evolving in the future? There has been immense growth in awareness of the power and possibilities offered by automated sentiment technologies, accompanied by market emergence of a slew of new tools. Unfortunately, many of them are simplistic and over-promise, which has led some to question the accuracy and usefulness of automated methods. Fortunately the challenge — and the business opportunities — have spurred the development of new, better methods. One is “active learning,” which is essentially human-curated machine learning. Another is application of crowdsourcing to sentiment-rating tasks, both for business analyses and to create training data for machine learning processes. Again fortunately, the major part of the technical complexity is hidden behind the scenes. There are some great analysis tools out there, accessible for business users and designed to produce business-usable insights.

4. What are the most innovative uses of sentiment analysis you’ve seen? It’s really interesting seeing applications of sentiment analysis for politics and policy, for 2012 election analysis, by media organizations, the campaigns, and researchers. We’re devoting a segment of the up-coming Sentiment Analysis Symposium, October 30 in San Francisco, to this topic. I expect there to be significant lessons learned, from political analyses, that are applicable to work in business market research, competitive intelligence, customer experience, marketing, and public relations. I’m also impressed by efforts that have linked social- and survey-mined sentiment to behavior models, psychological profiles, demographic data, and transactional records. Multi-source, multi-method “triangulation” is a real advance in creating business insight.

5. At the Sentiment Symposium, the American Cancer Society is talking about how to assess and react to market situations using social data. What else do you think social data can tell us? There are really very few limits at this point. Anything a human might grasp by reading social and online postings, a machine can also grasp, not perfectly but with increasing precision. Machines — computer software — has the speed and power to exceed human abilities. Automated analyses can find subtle patterns over time, geographically, related to certain language usage, that a human would never detect. They can link these patterns to real-life profiles and behaviors in order to predict people’s preferences and plans. We’re not quite there yet, but were closing in on the point where social data can tell use I’m looking forward to the symposium talk by Liz Keck of the ACS, also especially to social psychologist Kate Niederhoffer’s keynote, Sentiment Driven Behaviors, Sentiment Driven Decisions, and the Dow Jones, Luminoso, and eBay talks. I shouldn’t play favorites, however; we have some great speakers, and I hope folks in the Gnip community will join us.

Continue reading