Suggesting topics

The most frequent challenge when analyzing a corpus of textual answers is to cluster them into consistent groups of similar answers

The answer to this question is pretty straight-forward: There is no definition of a 'good' clustering, as it depends on what you are trying to achieve.

The Topic function is a way to group answers along a logical axis of your choice, on top of any sort of filter you may have added during import or distinct from the rating value attached to the answer.

The 'right' topic structure depends on the objectives of your analysis.

SimpleX can not (yet?) read your mind and guess what kind of groups you aiming at. So, it is worth understanding what kind of mathematical computations are behind the scenes.

SimpleX is able to compute the semantic similarity between two different text answers. If the two answers roughly bear the same meaning, even if formulated with different words or even different languages, the semantic distance between the two will be very short. If they are about two very different subjects, th distance will be very long.

NB: two answers saying opposite things about the same topic ("I love eating vegetables" vs "I hate eating vegetables") will be much closer than two answers about completely different topics  ("I love eating vegetables" and "Ford is a car brand"). The topics mentioned in the answer are therefore more important than say other relevant aspects such as sentiment, emotion or writing register.

When prompted, SimpleX computes the semantic distance between every couple of answers. 

Then, our algorithm delineates consistent groups of answers in such a way that answers from the same group will be close of from another and distant from answers of another group. Once groups are formed, another algorithm search for the keywords or that are representative of the answers from each given group.

By design, SimpleX suggest groups of answers mentioning the same topics. 

To obtain topic suggestions, select in the console the answers you are trying to cluster. Then select the grouped action Suggest topic for the selection and click Apply.

Topics will be created and be displayed in the left sidebar.

Usually, topic labels consist in one keyword (Services) or expression (Heavy workload).

As topic grows larger, SimpleX will use a combination of keywords (Manager / Reward / Recognition) to strive to reflect overall topic

Topic and topic label are created along two distinct algorithms. Both are not perfect and might deviate from expectation for different reasons:

• Topics: SimpleX clustering will provide a group of answers who share something in common. Sometimes, this 'something' is not relevant to the expected classification angle. In this case, the topic will not be judged as relevant, even though in 99 % of cases, there is indeed 'something else' tying together these answers. In fact, looking at the topic suggested by SimpleX oftne leads at uncovering overlook commonalities or clusters.
• Topic labels: SimpleX extracts one or several words that is close in meaning to the different answers from the topic. In this process, (too) generic words might stand out. Human brain is still much better than machines at picking one specific, yet transvers word to describe a topic.  

Having this limitations in mind, it is a good practice to always look at the contet of a cluster, regardless of the validity of the suggested label.

Topics and topic labels are imperfect. But editing them to better suit your needs is easy as 1-2-3.

Once automatic suggestions have been made, you are able to:

• edit the topic label
• move one or several answers from one topic to another
• merge two or more topics move one or several answers from one topic to another
• delete one topic

See also Topic essentials