Semi-automatic content moderation

Ever wanted to score incoming data based on easily defined criteria? Here it comes - content moderator, a set of basic (and some more advanced) matchers which form together a pipe where you put a data on one side and get a score on the other one.

What matchers do we have in our toolbelt?

• upercase-matcher : returns number of capitals letters
• content-size-matcher : returns count of all characters
• bad-words-matcher : lowercases and splits incoming data into set of words (stripped of common non-letter characters) and compares with :blacklist
• bad-email-matcher : lowercases incoming data, wraps it into one-element set and compares with :blacklist
• repeats-matcher : returns number of subsequently repeating characters
• bayes-matcher : returns bayes classification in binary form (0 if input was classified positively, 1 otherwise)

Naïve Bayes classification is based on judgr and 2 wrapper function have been exposed to classify input phrase:

• negative : trains classifier suggesting given phrase as negative
• positive : trains classifier suggesting given phrase as positive

Note that Polish extractor is used by default.

How matchers work

Each matcher returns either a number (integer or float) or a set (data structure). Now the magic happens - if it's a number a :min and :max parameters are checked if numeric value fits in between (:min <= value <= :max). If this condition is met a :penalty value is added to final score. Otherwise score is not modified. Oh, by default :min is set to 0 and :max to Integer/MAX_VALUE so there is no need to provide them both if only minimal or maximal value really matters.

In case when matcher returned a set instead of number a :blacklist set is checked if it contains at least one of elements returned by matcher. If so, again, :penalty is added to final score.

How to define custom matcher

Well, the easiest way is to use defmatcher macro. As said before, matcher is a function which returns either a number or a set and that's the only rule a matcher has to obey:

(defmatcher my-matcher
   (fn [input]
      ...process input and return number or a set... ))

Examples

(require '[mbuczko.moderator.matchers :as m])

(def blacklists {:content #{"incomplete", "bullshit"}
                 :emails #{"bad@boy.from.ru"}})

(def candidate {:title "Amazing brand new Alfa-Romeo with A FEEEW minor glitches"
                :contact {:phone-numbers ["1234", "55556"]}
                :description "Lorem ipsum dolor sit amet, consectetur adipiscing elit. Bullshit."
                :username "bad@boy.from.ru"})

(-> candidate
    (m/content-size-matcher :penalty 20 :field [:title] :min 71)
    (m/content-size-matcher :penalty 10 :field [:contact :phone-numbers] :max 0)
    (m/uppercase-matcher    :penalty 20 :field [:description] :min 36)
    (m/bad-words-matcher    :penalty 30 :field [:description] :blacklist (:content blacklists))
    (m/bad-email-matcher    :penalty 20 :field [:username] :blacklist (:emails blacklists))
    (m/repeats-matcher      :penalty 10 :field [:title] :min 2))

As a result a Candidate record will be returned with 3 relevant keys: :body with original data, :scores with vector of applied penalties in form of [penalty field matcher-name] and :final with sum of all applied penalties.

{:body {:title "Amazing brand new Alfa-Romeo with A FEEEW minor glitches",
        :contact {:phone-numbers ["1234" "55556"]},
        :description "Lorem ipsum dolor sit amet, consectetur adipiscing elit. Bullshit.",
        :username "bad@boy.from.ru"},
 :scores [[30 [:description] bad-words-matcher]
          [20 [:username] bad-email-matcher]],
          [10 [:title] repeats-matcher]],
 :final 60}

If you want to negatively classify given phrase:

(m/negative "Ala ma kota a kot ma Alę")

Now, let's pass it through bayes-matcher:

(m/bayes-matcher {:text "Ala lubi kota"} :penalty 99 :field [:text] :min 1)

Result:

{:body {:text "Ala lubi kota"},
 :scores [[99 [:text] bayes-matcher]],
 :final 99}

As expected, we got penalty score 99 because bayes-matcher classified our phrase as a negative one (and returned 1 in result).

LICENSE

Copyright © Michał Buczko

Licensed under the EPL.