Spam filtration

The problem that has caused comments from regular readers to be treated as spam should now be resolved.

For the information of other MT and Akismet users, it relates to how MT handles multiple spam tests. MT-Akismet assigns a score of +/-6 to each comment, then the e-mail and URL fields are checked against past comments in order to establish trust. You used to get a score of +1 for previous publication of your e-mail and +1 for a URL.

The problem arose at the next step, where the total score is averaged across the three tests. MT doesn’t include zero scores in this, so the following bug results: if you get one point for each of the trust checks, and six from Akismet, your average score is +2.67, which is below the trigger level. If, however, you had never commented before, rather than divide the 6 points from Akismet across the three tests, MT disregards the zero scores, so you get a final score of +6 and instant publication.

The workaround is to increase the scores for links and e-mail to 2.

4 thoughts on “Spam filtration

Comments are closed.

Spam filtration

Regular commenters may have noticed that a disturbingly large percentage of their comments have been held as spam. This issue should now be resolved. For the information of other Akismet/MT users, the problem was that our spam filter assigned a score of +1 if one’s URL had been previously published, and likewise if one’s required e-mail address had been. With the introduction of Akismet, which we can heartily recommend, a problem developed.

Specifically, the spam filter averages the score across all tests, so a genuine comment might have the +6 from Akismet and +1 from each of the other tests. Hence, an average of +2.67 – unfortunately, the threshold value is +3. This would not have been so much of a problem, had it not been that the filter disregards negative tests before averaging. Therefore, commenters with no track record who passed Akismet would get the full +6 points as their final score, but regulars, although getting a total of +8, would be averaged to +2.67

I have now increased the score for previous publishing, and we haven’t yet had a false positive.