diff --git a/geval.cabal b/geval.cabal
index bd6bbb5..b77f8b8 100644
--- a/geval.cabal
+++ b/geval.cabal
@@ -31,6 +31,7 @@ library
                      , Data.Conduit.AutoDecompress
                      , Data.Conduit.SmartSource
                      , Data.Conduit.Rank
+                     , GEval.FeatureExtractor
                      , Paths_geval
   build-depends:       base >= 4.7 && < 5
                      , cond
@@ -65,6 +66,7 @@ library
                      , Glob
                      , naturalcomp
                      , containers
+                     , statistics
   default-language:    Haskell2010
 
 executable geval
diff --git a/src/GEval/Core.hs b/src/GEval/Core.hs
index 4fa49c4..6f6bef2 100644
--- a/src/GEval/Core.hs
+++ b/src/GEval/Core.hs
@@ -213,7 +213,7 @@ getExpectedDirectory :: GEvalSpecification -> FilePath
 getExpectedDirectory spec = fromMaybe outDirectory $ gesExpectedDirectory spec
                             where outDirectory = gesOutDirectory spec
 
-data GEvalSpecialCommand = Init | LineByLine | Diff FilePath | PrintVersion
+data GEvalSpecialCommand = Init | LineByLine | WorstFeatures | Diff FilePath | PrintVersion
 
 data ResultOrdering = KeepTheOriginalOrder | FirstTheWorst | FirstTheBest
 
diff --git a/src/GEval/FeatureExtractor.hs b/src/GEval/FeatureExtractor.hs
new file mode 100644
index 0000000..e0f5f5b
--- /dev/null
+++ b/src/GEval/FeatureExtractor.hs
@@ -0,0 +1,25 @@
+{-# LANGUAGE OverloadedStrings #-}
+
+
+module GEval.FeatureExtractor
+  (extractUnigramFeatures,
+   extractUnigramFeaturesFromTabbed)
+  where
+
+import Data.Text
+import Data.List
+import Data.Monoid ((<>))
+
+extractUnigramFeatures :: Text -> Text -> [Text]
+extractUnigramFeatures namespace record = Prelude.map (prefix <>) $ nub $ tokenize record
+  where prefix = namespace <> ":"
+
+tokenize :: Text -> [Text]
+tokenize t = Data.List.filter (not . Data.Text.null) $ split splitPred t
+   where splitPred c = c == ' ' || c == '\t' || c == ':'
+
+extractUnigramFeaturesFromTabbed :: Text -> Text -> [Text]
+extractUnigramFeaturesFromTabbed namespace record =
+  Data.List.concat
+  $ Prelude.map (\(n, t) -> extractUnigramFeatures (namespace <> "<" <> (pack $ show n) <> ">") t)
+  $ Prelude.zip [1..] (splitOn "\t" record)
diff --git a/src/GEval/LineByLine.hs b/src/GEval/LineByLine.hs
index 3ce8697..80141ee 100644
--- a/src/GEval/LineByLine.hs
+++ b/src/GEval/LineByLine.hs
@@ -9,6 +9,7 @@
 
 module GEval.LineByLine
        (runLineByLine,
+        runWorstFeatures,
         runLineByLineGeneralized,
         runDiff,
         runDiffGeneralized,
@@ -25,12 +26,17 @@ import qualified Data.Conduit.List as CL
 import qualified Data.Conduit.Combinators as CC
 import Data.Text
 import Data.Text.Encoding
+import Data.Conduit.Rank
 
-import Data.List (sortBy, sort)
+import Data.List (sortBy, sort, concat)
 
 import Control.Monad.IO.Class
 import Control.Monad.Trans.Resource
 
+import Data.Monoid ((<>))
+
+import GEval.FeatureExtractor
+
 import Data.Word
 
 import Text.Printf
@@ -39,6 +45,11 @@ import Data.Conduit.SmartSource
 
 import System.FilePath
 
+import Statistics.Distribution (cumulative)
+import Statistics.Distribution.Normal (normalDistr)
+
+import qualified Data.Map.Strict as M
+
 data LineRecord = LineRecord Text Text Text Word32 MetricValue
                   deriving (Eq, Show)
 
@@ -54,6 +65,65 @@ runLineByLine ordering spec = runLineByLineGeneralized ordering spec consum
          formatScore :: MetricValue -> Text
          formatScore = Data.Text.pack . printf "%f"
 
+runWorstFeatures :: ResultOrdering -> GEvalSpecification -> IO ()
+runWorstFeatures ordering spec = runLineByLineGeneralized ordering spec consum
+   where consum :: ConduitT LineRecord Void (ResourceT IO) ()
+         consum = (rank (lessByMetric $ gesMainMetric spec)
+                   .| featureExtractor
+                   .| uScoresCounter
+                   .| CL.map (encodeUtf8 . formatFeatureWithZScore)
+                   .| CC.unlinesAscii
+                   .| CC.stdout)
+         formatOutput (LineRecord inp exp out _ score) = Data.Text.intercalate "\t" [
+           formatScore score,
+           escapeTabs inp,
+           escapeTabs exp,
+           escapeTabs out]
+         formatScore :: MetricValue -> Text
+         formatScore = Data.Text.pack . printf "%f"
+
+data RankedFeature = RankedFeature Text Double
+                     deriving (Show)
+
+data FeatureWithZScore = FeatureWithZScore Text Double Int
+                     deriving (Show)
+
+formatFeatureWithZScore :: FeatureWithZScore -> Text
+formatFeatureWithZScore (FeatureWithZScore f z c) =
+  f <> " " <> (pack $ show c)  <>  " " <> (pack $ printf "%0.20f" z)
+
+featureExtractor :: Monad m => ConduitT (Double, LineRecord) RankedFeature m ()
+featureExtractor = CC.map extract .| CC.concat
+  where extract (rank, LineRecord inLine expLine outLine _ _) =
+          Prelude.map (\f -> RankedFeature f rank)
+          $ Data.List.concat [
+              extractUnigramFeatures "exp" expLine,
+              extractUnigramFeaturesFromTabbed "in" inLine,
+              extractUnigramFeatures "out" outLine]
+
+uScoresCounter :: Monad m => ConduitT RankedFeature FeatureWithZScore m ()
+uScoresCounter = CC.map (\(RankedFeature feature r) -> (feature, (r, 1)))
+               .| gobbleAndDo countUScores
+               .| CC.map (\(f, (r, c)) -> FeatureWithZScore f (zscore (r - minusR c) c (2942 - c)) c)
+  where countUScores l =
+           M.toList
+           $ M.fromListWith (\(r1, c1) (r2, c2) -> ((r1 + r2), (c1 + c2))) l
+        minusR c = (c' * (c' + 1)) / 2.0
+              where c' = fromIntegral c
+        zscore u n1 n2 = let n1' = fromIntegral n1
+                             n2' = fromIntegral n2
+                             mean = n1' * n2' / 2
+                             sigma = sqrt $ n1' * n2' * (n1' + n2' + 1) / 12
+                             z = (u - mean) / sigma
+                         in cumulative (normalDistr 0.0 1.0) z
+
+lessByMetric :: Metric -> (LineRecord -> LineRecord -> Bool)
+lessByMetric metric = lessByMetric' (getMetricOrdering metric)
+  where lessByMetric' TheHigherTheBetter = (\(LineRecord _ _ _ _ scoreA) (LineRecord _ _ _ _ scoreB) ->
+                                             scoreA < scoreB)
+        lessByMetric' TheLowerTheBetter = (\(LineRecord _ _ _ _ scoreA) (LineRecord _ _ _ _ scoreB) ->
+                                             scoreA > scoreB)
+
 runLineByLineGeneralized :: ResultOrdering -> GEvalSpecification -> ConduitT LineRecord Void (ResourceT IO) a -> IO a
 runLineByLineGeneralized ordering spec consum = do
   (inputFilePath, expectedFilePath, outFilePath) <- checkAndGetFilesSingleOut True spec
diff --git a/src/GEval/OptionsParser.hs b/src/GEval/OptionsParser.hs
index 6647612..2ec66c0 100644
--- a/src/GEval/OptionsParser.hs
+++ b/src/GEval/OptionsParser.hs
@@ -46,6 +46,11 @@ optionsParser = GEvalOptions
                    <> short 'l'
                    <> help "Give scores for each line rather than the whole test set" ))
                  <|>
+                 (flag' WorstFeatures
+                 ( long "worst-features"
+                   <> short 'w'
+                   <> help "Print a ranking of worst features, i.e. features that worsen the score significantly" ))
+                 <|>
                  (Diff <$> strOption
                     ( long "diff"
                       <> short 'd'
@@ -194,6 +199,9 @@ runGEval''' (Just PrintVersion) _ _ = do
 runGEval''' (Just LineByLine) ordering spec = do
   runLineByLine ordering spec
   return Nothing
+runGEval''' (Just WorstFeatures) ordering spec = do
+  runWorstFeatures ordering spec
+  return Nothing
 runGEval''' (Just (Diff otherOut)) ordering spec = do
   runDiff ordering otherOut spec
   return Nothing