Vague Information

OpenEphyra goes beyond a pure keyword search. The system finds answers that are not written in the Wikipedia text. Broadly speaking, OpenEphyra analyzes and classifies both the questions and the data corpus available in textual form. Consequently, for the question "Who wrote Hitchhiker's Guide?," it is ready with the right answer: "Adams." If you are interested in the precise configuration, you will find a diagram of the underlying architecture online [12].

The test team fired off a total of 12 questions to Sirius from various fields of knowledge. For comparison, we questioned two of the proprietary competitors, Google Now and Apple's Siri. The results, summarized in Table 2, unfortunately relegate the question-answering system in Sirius to the status of nice gimmick without major practical uses. The open source assistant lands a few serious hits that do command respect, given the background of a data corpus developed without semantic or ontological input. For a reliable helper in everyday use, however, the number of hits is certainly not sufficient.

During the comparison with Google Now on Android and Siri on the iPhone, you should bear in mind that many answers have clearly been prepared on them with paid editorial work. Who should know better than Google what questions users frequently ask? Even the often emphasized "personality" of Apple's assistant hardly comes from machine intelligence, but from careful human development. Since OpenEphyra is very wide of the mark on a few questions, and the software is clearly difficult to compile on modern Linux distributions, the developers are already considering alternatives for Lucida [13].

Look Closely

Image recognition needs a database that includes known images before it can accept your service. Users can shift to this in the sirius-application/image-matching file and call up the make-db.py script. This accepts the name of the database (landmarks) and the directory of images (matching/landmarks/db) as parameters:

./make-db.py landmarks matching/landmarks/db

The name of the database is hard-coded in the start script start-imm-server.py. If you choose an identifier different from landmarks, you need to adjust the script accordingly.

When the ./start-imm-server.sh command from the run-scripts directory is called up, the image recognition server starts up. Afterward, sirius-imm-test.sh accepts an image file as a parameter. The name of the recognized object appears in the terminal output.

The test team added a few of their own images to the database for the comparison. For this, we used photos from the Freeimages service [14] and chose depictions of buildings that resembled Sirius' test images. With a few of the photos, the perspective changed somewhat, and it quickly became clear that the image matching only worked as long as the shooting location was almost the same. Stepping up against Sirius and the underlying OpenCV are Google Goggles [15] on Android and the CamFind app [16] on iPhone. Figure 4 shows the result.

Figure 4: If you submit images from different angles for Sirius image recognition, you find little joy. Only Google Goggles was almost flawless.

Until now, only the individual Sirius components could be seen in action. This means that the software is still a long way from a personal assistant à la Google Now, Siri, or Cortana. If you are looking for a handy, complete product, you can naturally combine the services using shell script.

Form a Team

Another possibility is to use the web server enclosed in the sirius-web folder. The README file unfortunately still contains dated information about the dependencies; the necessary software was copied onto the disk with pip install in the get-dependencies.sh script used earlier.

The web server started up successfully on Ubuntu 14.04, but it looked meager on the current system, and although the display worked in the browser (Figure 5), interaction with the controls was not possible.

Figure 5: The web front end included with Sirius in the sirius-web directory was unfortunately not functional in the test.