计算n_grammes困惑后的结果很奇怪

如何解决计算n_grammes困惑后的结果很奇怪

希望您能帮助我解决这个问题； 所以我训练了3个n_grammes模型(n=1,2,3)，但在计算出与现实不符的困惑（unigram模型的困惑低于bigram和trigram ...）之后，我发现了奇怪的结果。 这是代码

'''

proverbs_fn = "./data/proverbes.txt"
test1_fn = "./data/test_proverbes1.txt"
test_reponses_fn = "./data/reponses_du_test.txt"

n_grammes1=[]
n_grammes2=[]
n_grammes3=[]
model1 = Laplace(1)
model2 = Laplace(2)
model3 = Laplace(3)



def load_proverbs(filename):
    with open(filename,'r',encoding='utf-8') as f:
        raw_lines = f.readlines()
    return [x.strip() for x in raw_lines]


def load_tests(filename):
    with open(filename,encoding='utf-8') as fp:
        test_data = json.load(fp)
    return test_data

def load_tests_reponses(filename):
    with open(filename,encoding='utf-8') as f:
        raw_lines = f.readlines()
    return [x.strip() for x in raw_lines]


def train_models(filename):
    proverbs = load_proverbs(filename)



    for i in proverbs:
        tokens = word_tokenize(i)
        token_sentence=list(ngrams(pad_both_ends(tokens,n=1),n=1))
        for j in token_sentence:
            n_grammes1.append(j)

    for k in proverbs:
        tokens = word_tokenize(k)
        token_sentence=list(ngrams(pad_both_ends(tokens,n=2),n=2))
        for l in token_sentence:
            n_grammes2.append(l)

    for m in proverbs:
        tokens = word_tokenize(m)
        token_sentence=list(ngrams(pad_both_ends(tokens,n=3),n=3))
        for n in token_sentence:
            n_grammes3.append(n)


    vocabulary = build_vocabulary(proverbs)   


    model1.fit([n_grammes1],vocabulary_text=vocabulary)

    model2.fit([n_grammes2],vocabulary_text=vocabulary)

    model3.fit([n_grammes3],vocabulary_text=vocabulary)



    print("\n probabilites des n-grammes (n= 1,3) :\n")
    print("\t",model1.score("mentir",context=None))
    print("\t",model2.score("mentir",["beau"]))
    print("\t",model3.score("mentir",("a","beau")))

    print("\n perplexites des n-grammes (n= 1,3) : \n")
    print("\t",model1.perplexity([("aide-toi",),(",",("le",("ciel",("t",)]))
    print("\t",model2.perplexity([("aide-toi","),"le"),"ciel"),"t")]))
    print("\t",model3.perplexity([("aide-toi","ciel","t")]))

'''

结果是：

n个语法概率（n = 1,3）：

 0.00015090922809929828
 0.00044375416019525185
 0.00044672771945499217

n个语法的困惑（n = 1,3）：

 600.345333247329
 551.5106526623035
 2237.4997765363005