Word Clouds for Management Presentations: A Workflow with R & Quanteda
In this post, we’ll take a look at a basic text visualization technique we’ve seen elsewhere on this blog: word clouds. There are lots of great text analytics tools in R for this, and the process of making a basic word cloud is very straightforward. However, part of my job is providing data analysis and visualizations for senior management, and the basic word cloud approaches one finds straight out-of-the box don’t easily accommodate this use case. Below, I’ll outline my workflow for making customizable word clouds that you won’t be afraid to show to anyone in your organization!
Data Visualization for Management Presentations
It’s not easy to bridge the gap between analytics and management, and to ensure that data analysis is properly communicated to business stakeholders. This communication is an essential part of the job, and without it, the chances that your data analysis will have any impact are very small.
A critical aspect of communicating any type of data analysis is story telling, e.g. telling a narrative that relies on data analysis and suggests actionable conclusions. When presenting to management stakeholders, it’s very important to tell a story that flows logically and unimpeded towards its conclusion. Such presentations can easily get derailed when they are not focused, are structured based on methodological considerations (e.g. telling the story of the data analysis, not of the business problem + relevant insight), or present visuals that are unclear or invite excess scrutiny. Essentially, when presenting the results of a data analysis to management, you don’t want anything to distract from the data-driven conclusions you’re trying to convey.
Problems with Out-of-the-Box Word Clouds
The need to present clear and intuitive data visualizations is therefore of paramount importance. However, when using out-of-the-box word cloud routines in R, I’ve noticed two primary issues that make it difficult to make compelling visualizations for management stakeholders.
- Stemming (removing the end of a word to harmonize different forms, e.g. argue, argued, argues, arguing are all truncated to argu) is essential to getting good word counts, but stemmed words in word clouds look strange and are therefore distracting (e.g. they can easily derail a presentation into methodological discussions about text processing, rather than the implications of your data analysis for decision-making).
- However, to the best of my knowledge, none of the the packages for text analysis or word clouds in R make it easy to “un-stem” a word. From a strictly data analytic point of view, you would never really want to do this. For the current use case, though, it’s essential to do so.
- Sometimes you want to remove a word from a word cloud, even if it occurs very frequently. For example, the name of your company could occur very frequently (in internal documents or open-ended responses on surveys), but it doesn’t tell you anything insightful about the topic under study. In such cases, these words are often prominent in the word cloud (because they are used quite often), but have no added-value, don’t help you advance your story, and at worst can distract from other important topics that appear in the word cloud.
- While it is possible in every text analysis package to remove words from a document, corpus, or document-term frequency matrix, this typically occurs far upstream from the code used to make the word cloud. As such, it is not always straightforward to do so, particularly if you have many different columns in your dataset that should be turned into word clouds.
My Workflow for Word Clouds for Management Presentations
In this post, we’ll go through a work flow that I use in order to remedy the two above-mentioned problems with existing word cloud packages. The main workhorse of this process is the Quanteda package (which we’ve seen in a previous post). There’s lots of great things about this package, but something I really appreciate is that the package developers have thought a lot about making common text analytic procedures (e.g. stemming, term weighting, n-gram selection, removing numbers, etc.) very robust and easy-to-use.
The work flow uses a number of custom-built functions, which we’ll go over below. There are separate functions for all of the different steps that we need in the “Quanteda way” of analyzing text data. We first turn the text field in our dataframe into a corpus, from which we extract and clean text tokens (e.g. terms or words). We then convert the tokens into a document-feature matrix, and pass this along to the Quanteda wordcloud routine to create word clouds. Built into this workflow, I’ve created ways to specify words that should be replaced and their replacements (effectively “un-stemming” stemmed words) and to specify words which should be removed from the word cloud. Once the functions have been defined, it’s very easy to make a basic out-of-the box word cloud, examine it to see what needs to be changed, and to then re-make the word cloud with these changes taken into account.
Step 1: Data Frame and Text Field to Corpus
The first function takes a data frame with a text field and creates a corpus object. The corpus is the most basic element in the Quanteda text process flow, and is essentially a “library” of original documents which are stored along with meta-data at the corpus level and at the document-level.
Step 2: Corpus to Cleaned Tokens Object
The second function takes the corpus object we created with the first function, performs a number of text cleaning operations, and returns a tokens object (a list of tokens in the form of character vectors, where each element of the list corresponds to an input document).
Specifically, we remove punctuation, numbers, and symbols. We then convert all the letters to lower case, and stem the words (e.g. removing the end of the word to harmonize different variations on the same root). Finally, we remove any remaining words that are less than 3 characters and select unigrams and bigrams (e.g. 1 and 2 word combinations). Finally, the function returns the cleaned tokens object.
Step 3: Cleaned Tokens to DFM (Document-Feature Matrix)
The third function takes the cleaned tokens and generates a DFM (document-feature matrix), which is a matrix associating values for certain features with each document, with the documents in the rows and “features” in the columns.
Step 4: Remove Words We Don’t Want in the Word Clouds
The fourth function removes the words that we do not want to see in the word clouds. For example, when mining internal documents, the name of one’s company might occur very frequently. But this is rarely interesting or informative in the larger context of the data analysis, and provides no useful insight upon which to make a decision. Such words can only distract from the main point of the presentation, and so it’s a good idea to remove them.
In the function below, we remove the words from the tokens object created in Step 2 above.
Step 5: “Un-stem” the Stemmed Words
The fifth function allows us to specify a list of “to-be-replaced” words (e.g. “busi”) and the “replacement” words (e.g. “business”). Using this method, we can ensure that we don’t have any truncated words in our word cloud.
Step 6: Master Cleaning Function
The sixth and final function puts together all of the component functions we have defined above. The function takes a data frame and a text field, along with some optional parameters (e.g. words-to-replace, words to remove), and returns a dfm that is cleaned according to our specifications. We can pass this dfm directly to the Quanteda word cloud plot method to make our word cloud.
A Worked Example: Word Clouds for Management with Wine Data
Let’s go through the entire process with some sample data we’ve seen before on this blog. The dataset contains Winemaker’s Notes (a short text describing the qualities of a wine) for 2000 wines (1000 red and 1000 white). The data and code for the below analysis are available on Github here.
As an illustrative example, the first text in the data set looks like this:
In what follows, we will assume you have already defined the above functions in your R session.
Part 1: “Out-of-the-Box” Word Clouds
We first make an “out-of-the-box” word cloud, displaying the words “as is”, given the cleaning functions used in text processing. This visualization will allow us to see which specific words in our corpus should be changed or removed. (The changes we often want to make are unique to each context and data set, so it’s not possible to automate them.)
In the code below, we first define the color palette we will use in the word cloud. We then pass the data frame and the text field to our master function, without specifying any changes. In this example, we will work with the term frequencies (e.g. sum of word usage across all the documents). The code on Github also has an example using document frequencies.
Which returns the following plot:
Part 2: Specifying Changes and Making the Final Word Cloud
This already looks very nice! There are, however, a couple of changes needed to make a “management-ready” word cloud.
First, the most prominent word is “wine.” This makes perfect sense - these texts describe wine. However, we already knew that, and having the term dominate the plot does not add any insight or value. Let’s remove this word!
Second, we notice some stemmed words in the plot. For example, “palate” is truncated to “palat.” In the code below, I specify each of the stemmed words (old words in the function below) and indicate which words should serve as replacements (new words in the function below).
We define our modifications and pass everything to our master function like so:
Which returns the following plot:
This looks perfect! We have removed the term “wine” which didn’t add anything, and we’ve “un-stemmed” all of the stemmed words. There’s nothing here that will be an obvious distraction if we present this visualization to management!
Summary and Conclusion
In this post, we focused on text analysis and data visualization. We outlined a process to create customizable word clouds for use in management presentations, removing frequent words that did not add any insight and ensuring that all words in the visualization were complete, even after stemming.
This process allows us to produce “management-ready” visuals for presentation to decision-makers. I recently had to prepare a number of word clouds for presentation to senior management, and in every case, it was necessary to make these types of tweaks to the word clouds. The workflow described in this post made the task very straightforward and saved a tremendous amount of time when conducting the analyses. Based on how the presentations were received, I would also say that effort spent to customize the word clouds was definitely worth it!
Coming Up Next
In the next post, we will analyze text data from the lyrics of rap albums reviewed by Pitchfork, and use transfer learning and network analysis to identify influential albums across an 18 year period.