Sentiment analysis (also referred to as social listening and opinion mining) is used in media monitoring and allows brands to get a real handle on what people are saying about their product or service on the world wide web.
It’s been used to a great extent to achieve the following:
- Get an overview of how people feel about the brand, product, or service
- Gather feedback
- Perform crisis management
- Contact social media users directly to answer questions and address issues
For PR professionals, specifically, sentiment analysis tools really help to lighten the load by gathering important intel on the perception of a brand. This allows them to re-target public relations campaigns in order to have a greater impact where it matters.
An objective view
As great as sentiment analysis tools are, they’re not without their limits. While the technology is advancing in leaps and bounds, there are a few areas in which these tools fall down:
- Emojis: Many sentiment analysis tools are unable to recognize and to analyze emojis as yet
- Sarcasm: Known as the lowest form of wit, sarcasm is a major area in which sentiment analysis struggles
- Insufficient data: Often, there simply isn’t enough data out there to form an accurate overview of how people feel about a brand or product.
As well as the above, one of the major issues with sentiment analysis speaks to its ability (or the lack thereof) to figure out whether a comment is based on fact or, is simply the poster’s personal opinion. And that’s precisely what we’re going to explore in this article.
I run UpperKey, a real estate business and, although sentiment analysis is really useful for me, I find that I have to study the results quite closely. For some reason, real estate agents have a certain reputation and, because of this, I often find that social media comments shown by the tool are actually sarcastic, rather than positive. – Johan Hajji, CEO & Founder @ Uppperkey.com
They say that opinions are like backsides—everybody has one! While this has always been the case, the advent of social media has given individuals a platform from which to voice their opinions on everything from politics to product quality.
This is essentially what social media is for (apart from sharing photographs of cats) and is often a good thing as it offers a voice to people who would not otherwise be able to make their feelings widely known.
From a PR point of view, this has its ups and its downs. On one hand, it’s a great way of collecting data on their brand but, on the other hand, we’re dealing in opinion rather than fact. Sentiment analysis doesn’t judge, it just presents its findings.
Example of an opinion with a positive sentiment. Source: Twitter
Great as that is, it does pose a problem when it comes to sentiment analysis. Why? By its very nature, sentiment analysis models generally work by removing factual and neutral comments in order to only gather those missives which can be said to contain an emotion or feeling. This is necessary for the tool to be able to identify an overall ‘sentiment’ for a specific search.
The problem with this is that the sentiments which are collected are often subject to personal opinions about a product or service and can be subject to bias. For example, while playing one of our solitaire games, we had a user who didn’t like the ad that was played next to the game. The user posted negative comments on social media about our brand which of course can skew the results of the sentiment analysis unfairly. – Neal Taparia, Founder @ solitaired.com
The second opinion-based factor in sentiment analysis is that of socio-economic groups and subgroups. Social media is used by people from a diverse range of backgrounds, including culture, education, income, and religion. This means that people from different socio-economic groups are likely to feel differently about a product or service to other groups.
I consider that social media doesn’t provide an alternative way for discussion for those who can still remain silent to state their issues and opinions. I mean if they think that their followers and friends have different views with them in social media, they will be less likely to share their opinions and thoughts in other contexts, for example, gatherings of co-workers, or neighbors. This looks like a spiral of silence that might spill over from online contexts to in-person ones. – Irina Weber from SE Ranking.
An example of this would be a technology product or gadget being sold for a brand. A social media user from a particular background may post a negative comment about the product being too expensive whereas, somebody from a different background might consider it to be a good value for money. A tablet that costs $1000 may be perfectly affordable for one user but well beyond the means of another. For the PR professional, this presents the challenge of sorting through these factors in order to gain a true perspective.
Example of an opinion affected by their economic situation. Source: Twitter
A third factor when it comes to opinion is down to knowledge—particularly in the case of technology. There are tons of topic-specific groups and forums on the Internet which encourage opinion on particular products. This can cause issues for sentiment analysis due to the fact that different group members may have different levels of knowledge and ability.
An example might be a Facebook group dedicated to professional cameras and equipment. Within this group, there may be a thread regarding a specific camera which the members are discussing.
A professional photographer in the group might offer an opinion that the camera is fantastic and really easy to use whilst a less experienced photographer in the group might say that the camera is too complicated and not user friendly. This, of course, would result in a combination of positive and negative comments which may not fairly reflect the quality of the product.
Finally, another factor that can alter the results of sentiment analysis is that of the omnipresent troll.
There are, unfortunately, those Internet users who delight in using the internet to leave abusive and false comments on social media pages, groups, and websites. Although many brands work tirelessly to find and remove such comments, a lot of these will be missed and will, of course, be added to the mix when it comes to an overall sentiment analysis score.
Example of a fake opinion. Source: Twitter
When it comes to measuring subjectivity in sentiment analysis, this is broadly sorted into three groups which are as follows:
1. Hand-crafted. This kind of model uses a series of hand-crafted or human-centric rules for each classification tag. This means that a series of terms such as great, beautiful, interesting, ugly, complicated, rubbish are applied to the model. Flaws in this kind of model are largely centered around context; for example, the comment “The product stopped working after just one use; fantastic job”, might be classed as positive due to the use of the word ‘fantastic’, even though the comment is clearly negative in its context.
2. Automatic. These models harness the power of machine learning in order to create algorithms that are able to identify and predict sentiment. This works by training data by transforming it into vectors with pre-defined tags and thereby calculating the overall sentiment. The one major flaw in these models is that they rely very much on the data sets which are fed into them. This means that the model is only as good as the data and the people who are inputting the data. A lack of experience or knowledge of sentiment analysis and data sets can often lead to an ineffective sentiment analysis model.
3. Multi-Modal. The most advanced form of sentiment analysis, multi-modal takes the traditional text-based form of analysis and raises it with input from images, audio, and video, including YouTube videos. As well as being able to analyze a larger range of content, this kind of model uses a fusion technique to gain a much broader sense of the true intention behind a post or a comment.
The results of a sentiment analysis review can very much depend on which of the above models are used. These days, the common opinion is that the multi-modal model is by far the most accurate and effective for use by PR professionals and marketers.
Does it matter?
So, why is it so important to be able to separate fact from fiction in sentiment analysis? The answer to this lies in what you are hoping to achieve from the project.
For brands who are simply looking to get a rough gauge as to how users feel about their brand, sorting fact from opinion may not be hugely important. However, for brands that are using sentiment analysis for crisis management and for decision-making purposes, this can be hugely important.
For example, a mobile phone brand might use sentiment analysis to find out which features of its phone are popular and which features users would like to see in the future.
The first part of this is very much opinion based as some users might consider the photograph filtering feature to be essential whereas others will claim that it is pointless and something that they never use. For this reason, a multi-modal model with a huge and comprehensive dataset is absolutely essential in gaining accurate results.
Despite the limitations highlighted in this article, sentiment analysis is still the best bet for businesses who want to know what’s being said about them (unless they want to hire a staff member to painstakingly trawl the internet for mentions manually). Even in its simplest form, sentiment analysis can be incredibly useful in revealing issues and problems that a brand may not be aware of, as well as giving the brand a valuable troubleshooting tool.
We’re likely to see advances in these tools for PR managers which will enable them to sort results into many more segments which will, in turn, form the basis of multi-faceted PR campaigns.
Going forward, it’s likely that proactive sentiment analysis tools will be using a Bayesian network to represent subjective degrees of confidence. In plain English, these tools will be able to tap into a number of additional factors such as prior knowledge, evidence, likelihood calculations, and distribution of the belief network.
This will allow the tool to examine a larger number of parameters in order to gain a truer picture of the overall sentiment, rather than a sweeping generalization. This is going to be invaluable for large companies and corporations who rely on sentiment analysis when making major decisions on the future direction of their brand.
Example of a futuristic use of sentiment analysis
After reading about the limitations of sentiment analysis in this article, you might be forgiven for thinking that the technology is deeply flawed. Far from it, the truth is that sentiment analysis is a game-changer for a huge number of PR professionals and businesses worldwide and is used, on a regular basis, to improve services and products as well as for crisis management.
This is, quite simply, an incredibly valuable tool for businesses and will be used much more widely in the future. While sentiment analysis is a brilliant tool for gaining an overview, it’s a good idea for brands to constantly monitor the results and to regularly read a cross-section of the actual comments and posts to ensure that they’re gaining a human insight into the results rather than just trusting the tech.
Finally, while most sentiment analysis tools have their benefits, I would always recommend that brands invest in the most advanced tools that their budgets will allow as these are more likely to be multi-modal which means, in turn, that they will tend to be much more accurate and therefore present better value for money in the long run.
Cover photo by Celpax