Five Top Tips for Successful Text Analytics
Successfully unearthing text analytics insights does not need to be complicated. With our five top tips, avoid the pitfalls and obtain actionable insights from unstructured text.
1. Know your purpose Vague questions get vague answers. If you are planning text analytics of survey data and can control the questionnaire, ask a focused question. This will help orientate the respondents’ thinking so they provide richer comment, as well as the text analytics set up. For example, changing ‘any other comments’ to ‘tell us how to improve customer service’ will produce much more focused, actionable results.
In social media, identifying the question is no less important – but here it is more about deciding what you would like to learn. For example, needing to know what is being said about a particular product on a particular forum requires a different approach to that required for analysis of customer service comments relating to a given retailer.
2. Manage your organisation’s expectations Text analytics is not a perfect solution. In 2014, Ipsos carried out a comprehensive review of the major text analytics providers in the market and found that for all tools a balance is needed between precision (correct categorisation of comments) and recall (ensuring all relevant comments are included in a category).
Recognising that this compromise is necessary, and finding this balance, can be a challenge in text analytics. This often translates into ensuring that your business understands that the goal is to run analytics that are good enough on which to base business decisions, rather than perfect. To optimise this, it is well worth committing to investment in a thorough set up – and accepting that there is no miracle press-button solution in order to justify this.
3. Place the analyst at the heart of the process Don’t kill the analyst just yet: much of text analytics’ success relies on the analyst leveraging the text analytics tool’s strengths to deliver insight. This means trusting that the analyst has understood your business requirements, and can translate those requirements into a categorisation and/or text analytics outputs that respond to those needs. It may also mean acknowledging that the tool has some weaknesses – and finding a way to move past those. Better trust the analyst who recognizes difficulties and discusses them, rather than denying any are there.
4. Choose the right text analytics tool Text analytics tools fall into two main groups: linguistics/rules based and machine learning. Linguistics/rules based approaches tend to take longer to set up, but offer transferable resources and transparency over how comments are categorised.
In contrast, machine learning may be quicker to set up, but it can be harder to correct if it ‘learns’ the wrong approach and can feel “black box” as it is not possible to see how it is “thinking” about the categorisation.
By the same token, not all tools provide the same outputs: different tools provide sentiment at different levels of granularity, some provide webs that link up themes emerging in the data, some output results into SPSS or excel files for additional interrogation post text analytics, some provide online graphics and so on.
It is therefore worth thinking through your requirements before picking your provider: how much transparency over the categorisation process does your business need? How do you want to share the results? What do you want to focus on (sentiment, content, both)? Does your business need results online? etc.
5. Think big in terms of data volumes Scalability is at the heart of text analytics’ benefits versus more manual approaches: set up can be time consuming, but a high volume of comments ensures that text analytics remains time/cost effective versus more manual approaches. The definition of ‘high’ will depend on the tool you choose, but as a rule of thumb several thousand or more comments are required.
There is no need for all of the comments to come from the same source though, making text analytics an ideal approach for data integration (e.g. for survey and web data). Similarly, the comments are not needed all at once: text analytics can also be run over time (e.g. month on month over survey results). A continuous approach like this can give the volume of comments required over time for a return on the investment of initial set up.