This page describes how rule developers can make use of the n-gram data to detect errors. As an introduction and for set-up, please see Finding errors using n-gram data first.
To add word pairs, simply add them to the
resource/<languageCode>/confusion_sets.txt
file of your language. If
such a file doesn't exist for your language, ask the
developers to add one. However, you
need massive amounts of ngram data which is easily available only for
English, German, French, Spanish, Italian, Russian, Chinese, and
Hebrew. Here's an example line from that file:
word1; word2; 10 # comment
In this example, word1
and word2
are two similar words that can
easily be confused. The order of these words doesn't matter to LT, but
we recommend sorting them alphabetically. 10
is a factor that
determines how much the word used in the text is preferred. If you use
1
here, the other word will be suggested as a correction even when
it's only a little bit more probable than the word from the text. In
many cases, this would lead to false alarms. To avoid those, use a
larger factor, often something from 10
to 10000000
. To decide
what's a good compromise between finding many errors and not creating
too many false alarms, use ConfusionRuleEvaluator.java
:
Run ConfusionRuleEvaluator.java
from your IDE. It will print the
options you need to provide: the token pair, the language code, the
ngram directory, and an evaluation file with example sentences that
contain the words in their correct context. It will show output like
this:
Factor: 10 - 5 false positives, 11 false negatives
Summary: p=0.997, r=0.994, 2000, 3grams, 2015-09-16
Factor: 100 - 2 false positives, 22 false negatives
Summary: p=0.999, r=0.989, 2000, 3grams, 2015-09-16
Factor: 1000 - 2 false positives, 42 false negatives
Summary: p=0.999, r=0.979, 2000, 3grams, 2015-09-16
p=
is the precision value, i.e. the probability that a correct use of
a word is detected as correct. This value needs to be close to 1
, as
the difference between 1
and the value indicates the amount of false
alarms the rule will throw. r=
is the recall, i.e. the probability
that an incorrect use of a word is detected as incorrect. The closer
this is to 1
, the better, but a high precision should be preferred.
You can now choose the factor value that has a good compromise between
precision and recall and add it as the third value for this pair of
words in confusion_sets.txt
.
There's also the chance that even a very high factor doesn't cause a
good precision. As a rule of thumb, the precision should be at least
0.995
for very common words and 0.99
for other words. If that's not
possible with any factor, or if the recall is very low (< 0.5
), this
pair of words might simply not be a good fit for ngram-based error
detection.
Of course the result of evaluation depends on what you use as input:
the input itself may contain errors. That's why
ConfusionRuleEvaluator.java
will also print the cases that are
probably false alarms. If they aren't, you should probably clean up the
input. For English, we use a combination of Wikipedia and Tatoeba as
input. The fewer of the affected words appear in Wikipedia and Tatoeba,
the less meaningful the evaluation output will become. For English, we
try to use at least 1000 example sentences for each word.
- Currently, only errors can be detected where both words are one token.
For example, a confusion between
your
andyou're
cannot be detected, asyou're
is internally split into more than one token. - This works only for pairs - but if you have a set of three similar words, just arrange them into pairs: word1/word2, word1/word3, word2/word3