GermaNet with P1EDA does not produce alignments

[gloriann...@gmail.com]

Dear EOP-developers,

I have tried to use P1EDA with the GermaNet aligner.
I have used the P1EDA example template as provided here:

https://github.com/hltfbk/Excitement-Open-Platform/blob/master/alignmentedas/src/main/java/eu/excitementproject/eop/alignmentedas/p1eda/instances/SimpleWordCoverageDE.java

According to the documentation in the code, this instance of P1EDA should achieve 64.375% accuracy with the identical lemma aligner + GermaNet aligner.

I ran this template on the German RTE-3 testset and obtained exactly this result.
However, I noticed in the visualization html files that I additionally generated that not a single GermaNet alignment was added on the test set.

Therefore, I ran the template again without the GermaNet aligner and I still obtained 64.375% accuracy.


I am using GermaNet version 8 as suggested.
GermaNet also seems to get loaded correctly when I start the RTE-program, at least I get some output in the beginning that several GermaNet files are loaded.

Now I'm wondering why GermaNet is not adding any alignements and why I still get the same accuracy (seemingly) without using GermaNet at all.

Many thanks in advance and best wishes,
Glorianna

[Tae-Gil Noh]

Hello Glorianna,

Good to hear that you were able to run P1EDA as it was supposed to run. :-)

And about the GermaNet.

> Now I'm wondering why GermaNet is not adding any alignements and why I still get the same accuracy (seemingly) without using GermaNet at all.

You are using the same setup we used when we tested P1EDA and related paper. On that setup, GermaNet is really minimally used. To check this out, please check the following file.

package eu.excitementproject.eop.core.component.alignment.lexicallink.wrapped; GermaNetDELinker.java

There you will find the (hidden) configuration that sets "which relations reported in the GermaNet to be used in alignment".
Here's the code
// Filepath, weights (causes, entails, hypernym, synonym, hyponym, and antonym)

// 0 weight will make that relation not added as alignment.link. GermaNetWrapper lex = new GermaNetWrapper(germaNetPath, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0);

As you see, only "causes" relation is being used for generating alignment.link. By giving other weights (e.g. 1.0 or any value you want) to other relations, you can make that aligner to add more alignment links.

I do not clearly remember how we ended up with "only using causes" relation. --- which is a very conservative setup.  I think that I inherited the values from TIE EDA, but not really sure. So the reason why you get the same result with/without GermaNet is that ... there are only very few causes words recorded in GermaNet that also appears in RTE3 German test set.
Play with other values (e.g. synonym, hypernym, etc) will indeed make more links to appear.

When we did P1EDA and its paper, we didn't optimized those values to optimize test results.

Hope that this helps :-)

Sincerely,
Gil