How to "debug" StanfordNLP text classifier - stanford-nlp

Question

I'm using StanfordNLP to do text classification. I have a training set with two labels: YES and NO. Both labels have more or less the same datums per label (~= 120K).
The problem is that StanfordNLP is misclassifying some text, and I'm not able to identify why. How do I debug it?
My train file look like:
YES guarda-roupa/roupeiro 2 portas de correr
YES guarda-roupa/roupeiro 3 portas
YES guarda roupa , roupeiro 3 portas
YES guarda-roupa 4 portas
YES guarda roupa 6p mdf
YES guardaroupas 3 portas
YES jogo de quarto com guarda-roupa 2 portas + cômoda + berço
YES guarda roupa 4pts
NO base para guarda-sol
NO guarda-sol alumínio
NO guarda chuva transparente
NO coifa guarda po alavanca cambio
NO lancheira guarda do leao vermelha
NO hard boiled: queima roupa
NO roupa nova do imperador
NO suporte para passar roupa
The YES label identifies "guarda roupa" (wardrobe) and NO identifies things that aren't "guarda roupa" but have one or more commons words (such as "guarda chuva" -- umbrella, or "roupa" -- clothes).
I don't know why, but my model insists to classify "guarda roupa" (and its variations such as "guardaroupa", "guarda-roupas", etc) as NO...
How do I debug it? I already double checked my train file in order to see if I misclassified something, introducing an error, but I could not find it...
Any advice is welcome.
UPDATE 1
I'm using the following properties in order to control features creation:
useClassFeature=false
featureMinimumSupport=2
lowercase=true
1.useNGrams=false
1.usePrefixSuffixNGrams=false
1.splitWordsRegexp=\\s+
1.useSplitWordNGrams=true
1.minWordNGramLeng=2
1.maxWordNGramLeng=5
1.useAllSplitWordPairs=true
1.useAllSplitWordTriples=true
goldAnswerColumn=0
displayedColumn=1
intern=true
sigma=1
useQN=true
QNsize=10
tolerance=1e-4
UPDATE 2
Searching the API, I discovered that ColumnDataClassifier has a method getClassifier() that gives access to the underlying LinearClassifier, which has a dump() method. The dump produces an output that looks like bellow. From API: "Print all features in the classifier and the weight that they assign to each class."
YES NO
1-SW#-guarda-roupa-roupeiro-2portas 0,01 -0,01
1-ASWT-guarda-roupa-roupeiro 0,19 -0,19
1-SW#-guarda-roupa-roupeiro 0,19 -0,19
If I do a toString() into LinearClassifier it will print:
[-0.7, -0.7+0.1): 427.0 [(1-SW#-guarda-roupa-roupeiro-2portas,NO), ...]
[0.6, 0.6+0.1): 427.0 [(1-SW#-guarda-roupa-roupeiro-2portas,YES), ...]