#HSLIDE

• Supervised Machine Learning needs annotated data
• Existing training data are
  • often outdated
  • in military context
  • too universal or too specific
• Need to generate own training data
• Existing solutions are time consuming and exhausting
• ... and as such expensive

#HSLIDE

• Inconvenient available annotation interfaces
• Mostly with linguistic focus
• Only whole document views
• Distribution to annotators done manually

#VSLIDE

#VSLIDE

#VSLIDE

#HSLIDE

• To solve these problems, we restructured the way annotations are done fundamentally.
• We designed an iterative workflow to automate as much as we can and to save annotator's attention.
• And created the OpenSource project Dalphi.

#HSLIDE

• Web application, runs everywhere
• Helps building and maintaining annotated data
• Key features:
  • Iterating active learning supported workflow framework
  • Human readable presentation
  • Server side to propose useful annotations
  • Parallel distribution to annotators
  • Problem-agnostic document handling

#HSLIDE

• Service
• Raw data
• Annotation document
• Statistic
• Interface
• Project

#VSLIDE

• any system capable of communicating over HTTP
• maintaining problem specific jobs
• three types:
  • Iterate
  • Merge
  • Machine Learning

#VSLIDE

• data that needs to be annotated

#VSLIDE

JSON

{
	"foo": "bar",
	"foobar": 1.23
}

HTML

<h1>Impressum</h1>
<p>3antworten UG (haftungsbeschränkt)<br>Karl-Kunger Straße 64<br>12435 Berlin</p>

JPG

#VSLIDE

• a subset of raw data
• document which is renderable and annotable

#VSLIDE

• key-value pair
• mostly chronological numeric values

#VSLIDE

• problem specific user interface
• renders an annotation document as a subset of raw data

#VSLIDE

#HSLIDE

#VSLIDE

#VSLIDE

#VSLIDE

#VSLIDE

#VSLIDE

#VSLIDE

#VSLIDE

#VSLIDE

#VSLIDE

#HSLIDE

#VSLIDE

#HSLIDE

#VSLIDE

#VSLIDE

#HSLIDE

#VSLIDE