Growing Quantitative: Text Analysis on Surveys and Voice of the Customer. Introduction. Many surveys are conducted nowadays because it is so easy. Some surveys are not designed properly, and therefore, analysis post-evaluation becomes difficult and conclusions wake.

Case 1 is based on an analysis of a survey about cooperation across various teams in the department. In a Six Sigma project, conducting the voice of the customer, the VOC, is essential to be able to understand what would be value adding to the customer. However, if there are many customers, they probably have different opinions. How do you make the VOC operational? That's what Case 2 is about, based on 50 interviews with internal customers, stakeholders.

Text is unstructured data and can be treated in JMP Text Explorer to enable statistical analysis, to utilize the hidden information in the text, or to reduce the complexity enabling interpretation. Voice of the customer could, for instance, be a series of interviews. The minutes of meeting from these interviews is unstructured data and difficult to convert into measurable parameters in an objective way. However, using the features in Text Explorer, an objective procedure can be made.

Survey questions can either be close-ended, partially close-ended, or open-ended. An example of close-ended question could go like this: How would you rate Europe Discovery Summit 2024? Excellent, very excellent, world-class, very world-class, or of course, best in the universe. A partially close-ended question could go like this: What is your preferred dessert? And you may choose as many as you like. You can choose ice cream only, ice cream and cake, cake and fruit, ice cream fruit, candy, candy fruit, candy only, and so on. An open-ended question could go like this: How would you describe the perfect day?

In a Six Sigma project, the voice of the customer, the need, has to be translated into quality drivers and measurable critical to quality. It is about going from customer domain, a rather fluffy and unstructured place to be, to a technical domain of functional requirements and specifications.

Let's move into JMP for the demonstration of the two cases. Case 1 is the survey about cooperation between the health and safety teams when building the offshore wind farms. Case 2 is the voice of the customer introduced from a continuous improvement project about the design process of the offshore wind farms.

Case 1, the survey. What you see is a part of the original survey about collaboration in the department, where responses equal rows and questions equal columns. First, I would like to check the correlation between the responses from two close-ended questions.

Question one, how would you classify your job role? Back office, frontline, or other? Against question four, how often do you collaborate with a colleague from another department region outside your daily routine tasks? Weekly, every day, monthly, or less often?

Here, I strongly recommend to value order the options. Otherwise, it will be very confusing. How to do that? You select the column, right-click on your mouse, choose Column Properties, Value Order. You can see I have customized the order like this, less often, monthly, weekly, every day. When you finish with this, you click Okay.

In JMP, there are three ways to perform this analysis. Two of them will be explained. First method, we go under Analyze, Consumer Research, Categorical. We choose question one as X, the grouping category, and question four as the responses. Sorry about this. We click Okay. What you see here is a table containing cells. Inside the cells, we have the frequency, that's the observed count, the share in percentages, and the cell chi-square, P-value.

With the data given, we can see there's a significant low response in the frontline everyday cell. We click, select this, and then we go to the next method. That's fit Y by X platform. Similarly, we choose question one as X and question four as Y, the response. We click Okay. Here you see that we get a slightly different table because instead of the P-value, we get the expected value. We select that as well. Then we can combine the two windows to get the full report.

You can see very clearly that the reason why this has a significantly low P-value is because we only observe 4 counts in this cell, but we have expected 12.6, and because it's humans, it's actually 13. The difference between the observed and the expected value is so big that this P-value becomes significant. Perhaps there's an action needed in the department to improve on this.

This will close the close-ended question treatment. Let's go back to the data table and move to question 14. It goes like this: What helps you collaborate in the department? In fact, that is a partially close-ended question because the respondent has the possibility to choose multiple options like collaboration culture, face-to-face meetings, informal chats, networking time within company to know who is who, IT tools, all hands meetings, and so on.

To see how many answers we got, we can show in Graph Builder. We choose Graph Builder. Question 14, drag and drop it to the zones. Then we choose tree map. Click done and make it big. I think it's clear to everyone that we have about as many ways of answering the response as we have respondents, and hence complicates any analysis, if not making it impossible.

However, we can quantify the data if we go to Text Explorer. Back to the data table. Under Analyze, you find Text Explorer. We choose question 14 as the text column. We leave everything else as it is and click Okay. What you see here is the term and phrase list, and I will add all the phrases to the term by highlighting them, right-click and add phrase.

Now we can see that this list gives very variable information. We can see that the informal chat networking has the biggest counts, followed by face-to-face meetings, leader support, good project management, online meetings, and so on.

You can make a very nice report out of this. If you're anywhere in this table, right-click and click make it into data table. We got a new data table, and I choose to delete all the unimportant terms. Then we can go to Distribution and choose Term as the Y and Count as the frequency. We click Okay. Under the hotspot, you can stack the report. Under the Term hotspot, we can value order the counts by count descending. Perhaps we should change the histogram color a little bit. Now we got a nice report, something we can communicate and act upon.

Let's give the phrase informal chat networking a closer look. We go back to our data table and our text, Explorer, and choose informal chat networking term. If we right-click here, we can see that we can save an indicator. That will give us a one or a zero for whether the respondence has chosen this option or not. We do that.

In the data table, we can see that now we have created a column with ones or zeros. I would like to recode this to yes or no. If we go up here and click Recode, I will write no instead of zero. We accept that we convert the column to characters, and yes instead of one. Like this, and we click Recode. Now we got a new column with the same information, but with yes or nos instead of one or zeros.

Now we can perform a categorical test of the correlation between, for instance, question number 5, how would you rate the collaboration in the department in general? Very poor, poor, fair, good, or very good? Against our informal chat networking indicator, yes or no.

As we saw before, we can under Analyze, find the consumer research platform, and we choose Categorical. Now we choose question 5 as the X, and our informal chat networking indicator as the response. We click Okay. As before, we got this list or table. I think we should forget about the very poor cells because we only got two responses in those cells.

However, we can see that in the poor and no informal chats networking cell, we got a significantly low P-value. Almost all that rated the collaboration in the department as poor do not use informal chats networking. I think that is valuable information and perhaps something the department should improve. That was the partially close-ended question treatment and also the end of case number 1.

Let's move to case number 2, the VOC interviews. Case 2 is based on minutes of meetings transformed into stickers. The reason was to be able to group them on on the board. Each row represents a sticker, and we have here in this data table 361 stickers or rows, and they are coming from 50 respondents.

As you may have noticed, I have hidden the affinity diagram input, that would be the text on the stickers for sensitivity reasons. Now we will run the text explorer and do the add phrase the terms as we did before. I've already done that as you see here. All the phrase are grayed out, meaning that they have been added to the term list.

We can see that we have a term called project and another one called projects. Down here, we have a term called process, another one processes. To get rid of these redundancies, we can go under the hotspot and choose term options, stemming, stem all terms.

For further cleaning, we have some terms we cannot really use because they don't give any valuable information, for instance, Use. If I choose Use and right-click and add it as a stop word, it will disappear from the list. We can do the same with within. Add a stop word.

Here we have a very nice cleaned table we can work on. Now I will move into JMP Pro version because we are going to do a Latent Semantic SVD to get the topic analysis rotated SVD. The SVD transforms text data into a fixed dimensional vector space, making it amendable to all kinds of clustering, classification, and regression techniques.

If you have JMP Pro, you can do this. Under Hotspot, choose Latent Semantic Analysis SVD. We click Okay. Under this hotspot, the SVD-Centered and Scaled TF-IDF hotspot, we choose Topic Analysis, Rotated SVD.

Here we have to choose how many topics we want, and in this case, I choose four. I click Okay. What we see here is the top loading by topic report, it consists of four tables, one for each topic. The terms in these topics are the ones with the largest loadings in absolute values. We can use this report to determine conceptual themes for our Six Sigma project.

If we go to the Hotspot again and click Rename Topics. If we look at the terms in topic 1, I choose to call this Theme for Cost Efficiency. Topic 2, I call it Info System. And topic 3 is about the project execution itself, so it could be Project Execution. Topic 4, I choose to call that Knowledge Sharing. Click Okay. This is very, very great because now we have actually produced four quality drivers for our Six Sigma project.

By using topics, we now have an objective aggregation of the VOC statements, and instead of a manual individual dependent on hence subjective aggregation. Furthermore, we can show the topics as word clouds. We go to Display Options, Word Clouds by topic. Down here, we can open up and see that we get the same information for now as a word cloud by topic or quality drivers, as I used to call it.

If we to the Hotspot, we click on it, and we can see that we can actually save document topic vectors. If we do that and turn into our data table, we can see that we have created four new columns, each for one of the quality drivers. This is going to be very valuable when we come to the improve phase in our Six Sigma project, because there we have to discuss solutions.

Here we can see which of the customer we have to pay special attention to when it comes to discussing solutions for cost efficiency, the info system, the project execution, or the knowledge sharing. In order to see who is who, we can go to the Graph Builder and choose, I'll move it a little bit, to choose Topics as the Y, X, and Respondent as the X. Then we can choose Bar Plot. We click Done.

Here we can see that when we want to discuss solutions with our customers, this person is the go-to person when it comes to cost efficiency, whereas this person is good to talk to about solutions for the infosystem. Here we have the knowledge sharing person and here the project execution person. On that note, I will finish the demonstration in JMP, and return to the presentation.

Using JMP Pro Text Explorer, we have opened up the borderline between the voice of the customer domain to the technical domain of functional requirements and specifications. We have used JMP Text Explorer to create four objectives with quality drivers for us. We can easily derive measurable critical to qualities from these drivers. Consequently, we have saved a lot of time and uncoupled the human factor.

Finally, conclusion. Partially close-ended question response as well as complex text from, for instance, voice of the customer interviews, can be treated in Text Explorer into terms and phrases. Terms and phrases can be transferred into topics, quality drivers, or indicators, enabling an objective interpretation of the information. JMP Text Explorer offers a way to go from purely qualitative to quantitative text analysis. With this new knowledge, we can increase the quality of our VOC interviews and our surveys, and get solid conclusions on tests. Thank you for your attention.