The team & my task

Two frontend developers and one backend developer worked on the application. I was involved as UX designer and involved the Director of Design to discuss ideas and solutions.
In the background numerous people from the fields of research, software development and quality assurance participated.

My first task was to talk to the developers to understand the technical implications, possibilities and consequences.
Discussions with the stakeholders of the project gave me an impression of what customers would expect from it. I was responsible for creating the concepts for the analysis tools and the visualization app. I accompanied the implementation phase and then worked closely with the QA team to test the functionality and appearance.

Process

Phase 1

• Kick off meeting
• discussions with research and engineering
• creating first concept drafts for the folder-based analysis tool
• first concepts for the visualization application
• accompanying the implementation of phase

Phase 2

• collection of feedback
• creation of user stories
• reworking the concept (include extraction-based text classification)
• accompanying the implementation and QA testing

^{Phase 1}

Discovery

The outcome of the first phase was a concept and design for the folder-based analysis tool and the visualization app. Since the conception of the analysis tool has a high proportion of technical requirements, I will only describe the visualization app in the following.

In the first step, the task was to classify only whole documents.
Discussions with potential users have shown that both a general overview of classification results and detailed classifications are of interest.
General overview: how many documents in which class
Detailed classifications: which document has been assigned to which class

^{Example: The personal inbox}
It could be divided into the following classes: Advertising, bills, postcards, letters, important documents, etc.
All stacks can then be used further, e.g. by opening the letters and reading individual ones.

^{Phase 1}

Sketches

We divided the app into two parts. An overview page with the number of available documents, classes and documents per class.
A detail page, where for each document is exactly indicated: title, other information (sender / author / date / etc.), classification.

The overview part explains to the user what the classification has produced.
For an overview about the distribution of classes, I decided to present the results in a diagram as well.
The table shows the detailed numbers. The columns of the table consist of the class names, the number of classified documents and a guideline on how confident the system is with this classification.

The second part is a detailed list of all individual classifications. Initially, it is independent of the classifier and class. This view is intended to view the results based on the document level.

^{Phase 1}

Screens

Together with the other project participants, I agreed on the necessary UI frameworks High Charts for diagrams and AG Grid for tables. The visual design of the concepts in Sketch had a significant influence on their further development.

Selection of screens: First screen prototypes of previously scribbled ideas.

Since the application mainly consisted of the named frameworks and the largely existing component library, the implementation of the first phase could be completed relatively quickly.


Both of these initial results were achieved, and we were able to use the time until the second phase of the project to gather feedback.

Phase 2

During a period of three months, the results of phase 1 were analysed, new use cases identified and the underlying classification system was extended.

^{Phase 2}

Interviews & Insights

After some discussions with users of the text classification feature, we gained the following insights:

• we have to distinguish between document classification and page classification
• overview of the classification does not allow to compare different classification models
• app doesn't show what could not be classified
• no indication of how confident the system is about a particular classification

^{Phase 2}

Use Cases

In addition, there were a number of new use cases, some of which had already been addressed and some which had to be considered. Here are some examples:
As a user of the visualization application,

• I want to see a confidence value for each class of my classifier, so that I can judge the accuracy of the results.
• I want to know how many documents a class has been trained on, so that I can gain a deeper understanding of the confidence value.
• I want to see which and how many pages have been classified for a class, so that I can see the distribution.
• I want to see in how many of my documents contain which page classifications, so that I can get an overview of the contents of the documents.
  Example: How many of my documents contain a bill to pay?
• I want to see how many documents / pages have been processed but could not be classified.
• I want to see how sure the system is with each classification.

^{Phase 2}

Key things to redesign

Extension of the summary page
It now no longer contains only one classifier but all those selected in the current project.

Reviewing the summary page
By extending the overview page, we have chosen a new format. The cake diagram has been replaced by bar charts that have been incorporated into the table. This makes it possible to see different classifiers and compare them at the same time.

Document- and now also page classification
The new method of extraction-based text classification also required the ability to classify individual pages of a document.


Extension of detail table with more columns
The main focus remains on the individual classifications. This view has only been extended to include information about the classification confidence of the system.


Revision of columns grouping
We revised the grouping options to make the details of the individual classifications more accessible and to create an overview according to current requirements.
This feature was particularly interesting because grouping by document title can indicate that a document may have been classified into different classes.
In addition, the individual classified pages were again presented as a coherent group and a detailed insight into the content of the document was given.

Challenges & problems

A major challenge has been that text classification has to extend across an entire existing platform. In addition to new applications, old ones had to be adapted, which was not part of the project scope.
In addition, we set tables for the AG Grid framework. This framework has many possibilities, but also many limitations and cannot always be adapted without great effort.