请问这个log-likelihood结果该怎么解读

[majorlv511]

这个是我比较的一个词在两个语料库 （corpus1, corpus 2）中的使用情况，

比如附件中的这个 LL值为 -135.92， 该怎么解读呀？

谢谢大家了！

[majorlv511]

Log-likelihood calculator results
Key:
O1 is observed frequency in Corpus 1
O2 is observed frequency in Corpus 2
%1 and %2 values show relative frequencies in the texts.
+ indicates overuse in O1 relative to O2,
- indicates underuse in O1 relative to O2

[xiaoz]

Under use of your search word in corpus 1. The difference in the frequencies between corpus 1 and corpus 2 is highly significant in statistical terms p<0.001.

[majorlv511]

Under use of your search word in corpus 1. The difference in the frequencies between corpus 1 and corpus 2 is highly significant in statistical terms p<0.001.


but what's the concrete use of the number?

can i say "the LL value (-135.92) means under-use of the search word in corpus 1 is 135.92? or 135.92%?"

[xiaoz]

the number is called loglikelihood (LL) score. in your example (a 2 x 2 contigency table, i.e. 1 degree of freedom), the critical value for an observed difference to be statistically significant (at p<0.05, i.e. you can be 95% confident that the observed difference is real but not due to chance) is 3.84
(6.64 for p<0.01 or 99% significance level, 10.83 for p<0.001 or 99.9%, 15.13 for p<0.0001 or 99.99%).

So in your case, you can be over 99.99% confident that the search item is under used in your corpus 1 in relation to corpus 2.

[armstrong]

thans a lot for your explaination,Dr.Xiao.you are always helpful.

[majorlv511]

many thanx, Dr.xiao.

but where can i get the critical value of log-likelihood score? the same as Chi-square test score?

[xiaoz]

Yes, you are right. The distribution table for chi-square tests also applies to LL tests.

You will also find the following links of use if you do not have a distribution table at hand:

http://stattrek.com/Tables/ChiSquare.aspx (http://stattrek.com/Tables/ChiSquare.aspx)

http://www.fon.hum.uva.nl/Service/Statistics/ChiSquare_distribution.html (http://www.fon.hum.uva.nl/Service/Statistics/ChiSquare_distribution.html)

[armstrong]

可以从卡方分布表中查得，如果使用李绍山的《语言研究中的统计学》一书，可以在221页中获得，其中概率是0.05 是值为3.84;概率是0.01 时是6.64;概率是0.001时值是10.83.
你计算出的值是大于100，肯定超过99.999%的可信。
如果你使用的是韩宝成的《外语教学科研中的统计方法》一书可在188页中查到。

[xujiajin]

是的，韩宝成书的188－189页是X2表。