先看一下語言模型的輸出格式

1. \data\
2. ngram 1=64000
3. ngram 2=522530
4. ngram 3=173445
6. \1-grams:
7. -5.24036 ‘cause -0.2084827
8. -4.675221 ‘em -0.221857
9. -4.989297 ‘n -0.05809768
10. -5.365303 ‘til -0.1855581
11. -2.111539 </s> 0.0
12. -99 <s> -0.7736475
13. -1.128404 <unk> -0.8049794
14. -2.271447 a -0.6163939
15. -5.174762 a‘s -0.03869072
16. -3.384722 a. -0.1877073
17. -5.789208 a.‘s 0.0
18. -6.000091 aachen 0.0
19. -4.707208 aaron -0.2046838
20. -5.580914 aaron‘s -0.06230035
21. -5.789208 aarons -0.07077657
22. -5.881973 aaronson -0.2173971

（註：上面的值都是以10為底的對數值）

上面是一個語言模型的一部分，三元語言模型的綜合格式如下：

1. \data
2. ngram 1=nr # 一元語言模型
3. ngram 2=nr # 二元語言模型
4. ngram 3=nr # 三元語言模型
6. \1-grams:
7. pro_1 word1 back_pro1
9. \2-grams:
10. pro_2 word1 word2 back_pro2
12. \3-grams:
13. pro_3 word1 word2 word3
15. \end\

第一項表示ngram的條件概率，就是P(wordN | word1，word2，。。。，wordN-1）。

第二項表示ngram的詞。

最後一項是回退的權重。

舉例來說，對於三個連續的詞來說，我們計算三個詞一起出現的概率：

1. P(word3|word1,word2)

表示word1和word2出現的情況下word3出現的概率，比如P(平|習，進）的意思是已經出現了“習進”兩個字，後面是平的概率，這個概率這麽計算：

1. if(存在(word1,word2,word3)的三元模型){
3. return pro_3(word1,word2,word3) ;
5. }else if(存在(word1,word2)二元模型){
7. return back_pro2(word1,word2)*P(word3|word2) ;
9. }else{
11. return P(word3 | word2);
13. }

上面的計算又集中在計算P(word3 | word2)的概率上，就是如果不存在習進平的三元模型，此時不管何種路徑，都要計算P(word3 | word2) 的概率，計算如下：

1. if(存在(word1,word2)的二元模型){
3. return pro_2(word1,word2);
5. }else{
7. return back_pro2(word1)*pro_1(word2) ;
9. }

SRILM語言模型格式解讀