I have provided a solution for humor detection in binary data.

Features extracted:
-> remove word with length less than 3
-> text cleaning (bring the words to their 'complete' form; that means removing the words with aportrofe (like haven't, it will become have not)
-> stop word (but just a few of them, I don't want to extract all the stop words because some of them can give the text a negative sentiment like 'not' or 'no' and if I remove them, then the sentence will automatically become a positive one, which is wrong.
-> punctuation
-> stem words
-> lemm words
-> n-grams (2, 3 words)

Algorithms used:
-> Logistic Regression
-> Naive Bayes (Bernoulli, Multinomial)
-> SVM (LinearSVC, SVC (with kernel = linear or rbf), NuSVC (with kernel = linear or rbf))
-> DecisionTree
-> RandomForest
-> KNeighbors

The best values of C (LogisticRegression), random_state (DecisionTree, RandomForest), nu (NuSVC), n_neighbors (KNeighbors) were chosen with GridSearchCV.

                  clean  sw   punct  stem  lemm   ngrams  all
C                 1000   100  1000   100   1000   100     0.001 
nu*(1)            0.3    0.3  0.3    0.3   0.3    0.7     0.7 
nu*(2)            0.3    0.3  0.3    0.3   0.3    0.3     0.3 
n_neighbors       11     21   17     19    13     21      21 
random_state(1)*  1      12   13     18    15     12      12 
random_state(2)*  15     10   9      11    7      8       15

nu*(1) - NuSVC rbf
nu*(2) - NuSVC linear
random_state(1)* - RandomForest
random_state(2)* - DecisionTree