Memory-based learning: Predicting gender of French nouns

I'm trying to predict the gender of French nouns based on their suffixes. I have a corpus of 10k nouns. For each noun, I split the root from the suffix. I create five instances but with varying suffix lengths: 1, 2, 3, 4 and 5.

The class name I'm assigning to each instance is composed of the gender and the length of the suffix, e.g. "f2" for a feminine noun with a suffix of the length 2.

To give you an idea of how the data looks, here's a small extract: http://pastebin.com/tn6DFJAy

Do you think this is a reasonable approach? Is it sensible to separate the classes based on the suffix length? And is it safe to generate five instances for each noun?

For my project, I'm using TiMBL. The results for the data are: http://pastebin.com/5wai65i1

-

I'm a little unclear on your motivation for this (if it's for an actual application, wouldn't a dictionary be easier?), but it could be a neat computational linguistics project if you're trying to "rediscover" the rules or find patterns in the irregular ones.

Although you're getting reasonable performance (mean AUC=0.7), I think you should rethink how you generate both the features and the class labels.

If you look at the confusion matrix at the end of your results, a lot of the 'f5' examples are misclassified as 'f4', 'f3', 'f2', and 'f1'. It seems weird to count these as incorrect, since it's getting the gender correct (good!), but the suffix length wrong (do you actually care?). If you don't, maybe just recode them as 'f' and 'm', which should be easy enough with find and replace.

As for the features, it seems like it'd be better to have multiple features per instance. If you are testing the hypothesis that gender is encoded in the suffix, I would be tempted to try making each of the last five letters a feature. For example, instead of doing:

esthé, t, f5,
esthét i f4
esthéti q f3
esthétiq u f2
esthétiqu e f1


represent the same data as

t, i, q, u, e, F #esthétique


You might need an $\emptyset$ symbol for shorter words.

Finally, it might be neat to run this through an algorithm that gives you rules. It would cool if your learnt rules match the ones taught in French class (e.g., [don't care],e,u,s,e-->female).

-
Thanks for the reply. After applying your suggestions, the results are even better. Now the AUC climbed up to 89%. This surprises me because now the five characters are considered separately. In my approach I chose to regard the suffix as a whole since I thought this might yield a better precision (which doesn't seem to be the case). – tdzk Jan 30 '12 at 23:26
So if I split up the suffix "age" into the three features "a", "g" and "e", does this imply that the features are aligned by column? In this particular case, an instance with the features "g", "a" and "e" wouldn't match because every single column needs to be equal. Correct? – tdzk Jan 30 '12 at 23:26
Yup. You can think about it as feature1='a', feature2='g', feature3='e', which is clearly different from feature1='g', feature2='a', etc. – Matt Krause Jan 31 '12 at 2:31
Ah, the suffix as a whole idea is interesting. Have you thought about doing a sort of "cumulative suffix" representation? Here, each feature would be the last N characters, so you'd have something like: e, ue, que, ique, tique, F for esthetique. My gut feeling is this more closely matches how French grammar actually works. – Matt Krause Jan 31 '12 at 2:40
Ahhh. This really should have gone in my original answer, but is there any chance you could get a French phonetic ("pronouncing") dictionary? Morphological stuff like this usually acts on the phonetic version of the word and not its spelling. It might be interesting to try that, using the last N sounds. – Matt Krause Jan 31 '12 at 2:44

Have a look to my website on this topic, especially the page on suffixation and gender - http://genderfrenchnouns.yolasite.com/derogation-1.php

-

We're looking for long answers that provide some explanation and context. Don't just give a one-line answer: please explain why you're recommending it as a solution. Answers that don't explain anything will be deleted. See Good Subjective, Bad Subjective for more information.