(Un)constrained submission

[Jakub Macháček]

Hello,

I have some lists of words. I extracted them from the training set and manually filtered out some misleading words. Do I get it right that regardless its source (training set), those lists would still count as additional resources and thus could not be used in constrained submission? Thank you in advance.

Best regards,

Jakub Macháček

[Dimitris Galanis]

Hello,

Did you use any resources (e.g. polarity lexica) to extract the words?

DG

--
You received this message because you are subscribed to the Google Groups "SemEval-ABSA" group.
To unsubscribe from this group and stop receiving emails from it, send an email to semeval-absa...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Jakub Macháček]

No, I used no lexicons to extract the data. But I used my knowledge of English language to decide which data should be in which list.

Dne středa 20. ledna 2016 15:37:00 UTC+1 Dimitris Galanis napsal(a):

[Dimitris Galanis]

Hello again,

This is a borderline case between constrained and unconstrained.

Since the words are extracted from *only* the training dataset without 

using any external resources (e.g. lexica, additional reviews), we think that 

-even though you used your knowledge of English language- your submission should be marked as **constrained**. So, please mark it as constrained and submit it when you are ready.

In the overview paper we will mark it either as constrained or as special case of constrained (e.g constrained*). We used a constrained* mark for similar submissions in the past (ABSA-14, ABSA-15).

Please in your submission ("Short submission description") and in your paper describe in detail 

your method and the resources you used.

DG

[Ayush]

Hi,

Does using tokenizer also puts into unconstrained submission?

Thanks

Ayush

[Dimitris Galanis]

Hi,

The use of generic text processing tools such as tokenizers, POS taggers, dependency parsers etc does not make a submission unconstrained. If you use a sentiment analyzer that is based on a sentiment lexicon then it is unconstrained.

DG

[Stefan Falk]

I am keeping coming back to this issue because in my opinion this should not be considered as "constrained". Imho this is absolutely equivalent as using an external resource. Could somebody explain to me why exactly this is not considered as unconstrained?

I would like to give an example: Assume among those 2000 sentences are very precise words like e.g. "savory". Now let such a word only appear a few times in the entire corpus. This means I will find it hard to conclude that "savory" (rare) is actually similar as "delicious" simply because there are not enough samples. However, using pre-trained GoogleNews word embeddings I could determine that those words do in fact show a certain similarity. Thus, using external resources, I am absolutely able to improve my results because GoogleNews word embeddings are going to deliver knowledge form outside of the provided training data.

How is that different as if I were just using my own brain?