@@ -178,7 +178,7 @@ import qualified Text.Pandoc.UTF8 as UTF8 (putStrLn)
178178import Text.Parsec hiding (token)
179179import Text.Parsec.Pos (newPos)
180180import Data.Char ( toLower, toUpper, ord, chr, isAscii, isAlphaNum,
181− isHexDigit, isSpace )
181+ isHexDigit, isSpace, isPunctuation )
182182import Data.List ( intercalate, transpose, isSuffixOf )
183183import Text.Pandoc.Shared
184184import qualified Data.Map as M
@@ -405,9 +405,18 @@ emailAddress = try $ toResult <$> mailbox <*> (char '@' *> domain)
405405 domain = intercalate "." <$> (subdomain `sepby1` dot)
406406 dot = char '.'
407407 subdomain = many1 $ alphaNum <|> innerPunct
408− innerPunct = try (satisfy (\c -> isEmailPunct c || c == '@') <*
409− notFollowedBy space)
410− emailWord = many1 $ satisfy isEmailChar
408+ -- this excludes some valid email addresses, since an
409+ -- email could contain e.g. '__', but gives better results
410+ -- for our purposes, when combined with markdown parsing:
411+ innerPunct = try (satisfy (\c -> isEmailPunct c || c == '@')
412+ <* notFollowedBy space
413+ <* notFollowedBy (satisfy isPunctuation))
414+ -- technically an email address could begin with a symbol,
415+ -- but allowing this creates too many problems.
416+ -- See e.g. https://github.com/jgm/pandoc/issues/2940
417+ emailWord = do x <- satisfy isAlphaNum
418+ xs <- many (satisfy isEmailChar)
419+ return (x:xs)
411420 isEmailChar c = isAlphaNum c || isEmailPunct c
412421 isEmailPunct c = c `elem` "!\"#$%&'*+-/=?^_{|}~;"
413422 -- note: sepBy1 from parsec consumes input when sep