diff --git a/appsrc/NERTrainer.hs b/appsrc/NERTrainer.hs new file mode 100644 --- /dev/null +++ b/appsrc/NERTrainer.hs @@ -0,0 +1,16 @@ +{-# LANGUAGE OverloadedStrings #-} +module NERTrainer where + +import System.Environment (getArgs) +import NLP.Chunk (saveChunker) +import qualified NLP.Corpora.WikiNer as W + +main :: IO () +main = do + args <- getArgs + let output = last args + corpora = init args + + chunker <- W.trainChunker corpora + saveChunker chunker output + diff --git a/changelog.md b/changelog.md --- a/changelog.md +++ b/changelog.md @@ -1,3 +1,11 @@ += 0.5.2.0 = + + - Unceremoniously forced the Chunking model to do basic Named + Entity Recognition, and added a simple NER model based on Nothman + et al's WikiNER data set -- which is provided under a very + permissive CC license. + (http://schwa.org/projects/resources/wiki/Wikiner) + = 0.5.1.0 = - Moved to Tasty from test-framework. I'm treating this as a diff --git a/chatter.cabal b/chatter.cabal --- a/chatter.cabal +++ b/chatter.cabal @@ -1,5 +1,5 @@ name: chatter -version: 0.5.1.0 +version: 0.5.2.0 synopsis: A library of simple NLP algorithms. description: chatter is a collection of simple Natural Language Processing algorithms. @@ -37,6 +37,7 @@ ./data/models/brown.pos.model.gz ./data/models/conll2000.pos.model.gz ./data/models/conll2000.chunk.model.gz + ./data/models/wikiner.ner.model.gz source-repository head type: git @@ -65,6 +66,7 @@ NLP.Corpora.Email NLP.Corpora.Brown NLP.Corpora.Conll + NLP.Corpora.WikiNer NLP.Similarity.VectorSim NLP.Extraction.Parsec NLP.Extraction.Examples.ParsecExamples @@ -141,9 +143,9 @@ ghc-options: -Wall -main-is ChunkTrainer -rtsopts -Executable eval +Executable trainNER default-language: Haskell2010 - Main-Is: Evaluate.hs + Main-Is: NERTrainer.hs hs-source-dirs: appsrc Build-depends: chatter, @@ -154,27 +156,44 @@ cereal >= 0.4.0.1, containers >= 0.5.0.0 - ghc-options: -Wall -main-is Evaluate -rtsopts + ghc-options: -Wall -main-is NERTrainer -rtsopts -benchmark bench - type: exitcode-stdio-1.0 - default-language: Haskell2010 - Main-Is: Bench.hs - hs-source-dirs: tests/src - Other-modules: NLP.Similarity.VectorSimBench - Corpora +Executable eval + default-language: Haskell2010 + Main-Is: Evaluate.hs + hs-source-dirs: appsrc Build-depends: chatter, - criterion >= 0.8.0.1, filepath >= 1.3.0.1, text >= 0.11.3.0, - base >= 4.6 && <= 6, - deepseq, - split >= 0.1.2.3, - tokenize >= 0.2.0 + base >= 4.6 && <= 6, + bytestring >= 0.10.0.0, + cereal >= 0.4.0.1, + containers >= 0.5.0.0 - ghc-options: -Wall -main-is Bench + ghc-options: -Wall -main-is Evaluate -rtsopts + +-- disabled because we aren't using it, and it's causing issues with `stack ghci` +-- benchmark bench +-- type: exitcode-stdio-1.0 +-- default-language: Haskell2010 +-- Main-Is: Bench.hs +-- hs-source-dirs: tests/src + +-- Other-modules: NLP.Similarity.VectorSimBench +-- Corpora + +-- Build-depends: chatter, +-- criterion >= 0.8.0.1, +-- filepath >= 1.3.0.1, +-- text >= 0.11.3.0, +-- base >= 4.6 && <= 6, +-- deepseq, +-- split >= 0.1.2.3, +-- tokenize >= 0.2.0 + +-- ghc-options: -Wall -main-is Bench test-suite tests diff --git a/data/models/README b/data/models/README --- a/data/models/README +++ b/data/models/README @@ -40,3 +40,9 @@ real 0m20.730s user 0m20.298s sys 0m0.290s + +--------------------------------------------------------------------- +wikiner.ner.model.gz +--------------------------------------------------------------------- + +A Chunker that chunks to NER tags; trained on aij-wikiner-en-wp2 diff --git a/data/models/wikiner.ner.model.gz b/data/models/wikiner.ner.model.gz new file mode 100644 # file too large to diff: data/models/wikiner.ner.model.gz diff --git a/src/NLP/Corpora/Conll.hs b/src/NLP/Corpora/Conll.hs --- a/src/NLP/Corpora/Conll.hs +++ b/src/NLP/Corpora/Conll.hs @@ -16,6 +16,20 @@ import qualified NLP.Types.Tags as T import NLP.Types.General +-- | Named entity categories defined for the Conll 2003 task. +data NERTag = PER + | ORG + | LOC + | MISC + deriving (Read, Show, Ord, Eq, Generic, Enum, Bounded) + +instance Arbitrary NERTag where + arbitrary = elements [minBound..] + +instance Serialize NERTag +instance T.NERTag NERTag + +-- | Phrase chunk tags defined for the Conll task. data Chunk = ADJP | ADVP | CONJP @@ -35,7 +49,6 @@ instance Serialize Chunk -instance Serialize Tag instance T.Tag Tag where fromTag = showTag @@ -56,6 +69,7 @@ instance Arbitrary Tag where arbitrary = elements [minBound ..] +instance Serialize Tag readTag :: Text -> Either Error Tag readTag "#" = Right Hash diff --git a/src/NLP/Corpora/WikiNer.hs b/src/NLP/Corpora/WikiNer.hs new file mode 100644 --- /dev/null +++ b/src/NLP/Corpora/WikiNer.hs @@ -0,0 +1,136 @@ +{-# LANGUAGE OverloadedStrings #-} +{-# LANGUAGE DeriveGeneric #-} +-- | A parser for the Wiki NER work presented in: +-- +-- @Article{nothman2012:artint:wikiner, +-- author = {Joel Nothman and Nicky Ringland and Will Radford and Tara Murphy and James R. Curran}, +-- title = {Learning multilingual named entity recognition from {Wikipedia}}, +-- journal = {Artificial Intelligence}, +-- publisher = {Elsevier}, +-- volume = {194}, +-- pages = {151--175}, +-- year = {2012}, +-- doi = {10.1016/j.artint.2012.03.006}, +-- url = {http://dx.doi.org/10.1016/j.artint.2012.03.006} +-- } +-- +-- And provided here: http://schwa.org/projects/resources/wiki/Wikiner +-- +-- The format does not appear to be documented, but it looks like: +-- +-- * One sentence per line. +-- +-- * Tagged tokens are separated by spaces +-- +-- * Items in a tagged token are separated by vertical bars ('|') +-- +-- * Each line of `n` text tokens contains 3*n items, starting with a +-- text token, a POS tag, then a IOB tag with one of the NER classes +-- +-- For example, the sentence: +-- The Oxford Companion to Philosophy says, "there is no single defining position that all anarchists hold, and those considered anarchists at best sharae a certain family resemblance." +-- +-- Is rendered as: +-- The|DT|I-MISC Oxford|NNP|I-MISC Companion|NNP|I-MISC to|TO|I-MISC Philosophy|NNP|I-MISC says|VBZ|O ,|,|O "|LQU|O there|EX|O is|VBZ|O no|DT|O single|JJ|O defining|VBG|O position|NN|O that|IN|O all|DT|O anarchists|NNS|O hold|VBP|O ,|,|O and|CC|O those|DT|O considered|VBN|O anarchists|NNS|O at|IN|O best|JJS|O share|NN|O a|DT|O certain|JJ|O family|NN|O resemblance|NN|O .|.|O "|RQU|O +-- +-- +-- This module also provides a trained model for NER via the averaged +-- perceptron chunker. This actually kindof works, which is a bit +-- amazing. For example: +-- +-- > import NLP.Corpora.WikiNer +-- > import NLP.POS +-- > import NLP.Chunk +-- > tgr <- defaultTagger +-- > chk <- wikiNerChunker +-- > chunkText tgr chk "Real World Haskell is a book created by Don Stewart, Bryan O'Sullivan, and Jon Goerzen." +-- > "[ORG Real/NNP] [MISC World/NNP] [PER Haskell/NNP] is/VBZ a/DT book/NN created/VBN by/IN [PER Don/NNP Stewart/NNP] ,/, [PER Bryan/NNP O'Sullivan/NNP] ,/, and/CC [PER Jon/NNP Goerzen/NNP] ./." +-- +-- +module NLP.Corpora.WikiNer + ( parseWikiNer + , trainChunker + , wikiNerChunker + , Chunk(..) + ) +where + +import Data.Text (Text) +import qualified Data.Text as T +import qualified Data.Text.IO as T +import Data.Serialize (Serialize) +import GHC.Generics +import System.FilePath (()) +import Text.Read (readEither) +import Test.QuickCheck.Arbitrary (Arbitrary(..)) +import Test.QuickCheck.Gen (elements) + + +import NLP.Chunk (train, loadChunker) +import NLP.Chunk.AvgPerceptronChunker (Chunker(..), mkChunker) +import qualified NLP.Corpora.Conll as Conll +import NLP.ML.AvgPerceptron ( emptyPerceptron ) +import NLP.Types.IOB hiding (parseIOB) +import NLP.Types.General (Error, toEitherErr) +import NLP.Types.Tags + +import Paths_chatter + +parseWikiNer :: Text -> Either Error [[IOBChunk Chunk Conll.Tag]] +parseWikiNer = parseIOB + +-- | Convert wikiNer format to basic IOB (one token perline, space +-- separated tags, and a blank line between each sentence) +parseIOB :: (ChunkTag chunk, Tag tag) => Text -> Either Error [[IOBChunk chunk tag]] +parseIOB input = sequence $ map (parseSentence . toIOBLines) (filter (/="") $ T.lines input) + +-- | Different classes of Named Entity used in the WikiNER data set. +data Chunk = LOC + | MISC + | ORG + | PER + | C_O -- ^ "out" not a chunk. + deriving (Read, Show, Ord, Eq, Generic, Enum, Bounded) + + +instance Arbitrary Chunk where + arbitrary = elements [minBound ..] + +instance Serialize Chunk + +instance ChunkTag Chunk where + fromChunk = T.pack . show + parseChunk txt = toEitherErr $ readEither (T.unpack txt) + notChunk = C_O + +wikiNerChunker :: IO (Chunker Chunk Conll.Tag) +wikiNerChunker = do + dir <- getDataDir + loadChunker (dir "data" "models" "wikiner.ner.model.gz") + +-- | Tranlsate a WikiNER sentence into a list of IOB-lines, for +-- parsing with `parseIOBLine` +toIOBLines :: Text -> [Text] +toIOBLines sent = map (T.replace "|" " ") (T.words sent) + +-- | Train a chunker on a provided corpus. +trainChunker :: [FilePath] -> IO (Chunker Chunk Conll.Tag) +trainChunker corpora = do + content <- mapM T.readFile corpora + + let trainingText = T.intercalate "\n" content + + eiobs = parseWikiNer trainingText + + chunker :: Chunker Chunk Conll.Tag + chunker = mkChunker emptyPerceptron + + case eiobs of + Left err -> do + T.putStrLn err + error (T.unpack err) + Right iobs -> do + print (take 1 iobs) + let chunkSents = map toChunkTree iobs + train chunker chunkSents + diff --git a/src/NLP/Types/IOB.hs b/src/NLP/Types/IOB.hs --- a/src/NLP/Types/IOB.hs +++ b/src/NLP/Types/IOB.hs @@ -4,6 +4,7 @@ import Prelude hiding (print) import Control.Applicative ((<$>), (<*>)) import Data.Maybe (mapMaybe) +import Data.Monoid ((<>)) import Data.Text (Text) import qualified Data.Text as T @@ -14,6 +15,18 @@ import NLP.Types.Tree import NLP.Types.General (Error) + +-- TODO: This module needs to be rewritten to parse IOB represented +-- data as a tree, then once that tree is created, establish the +-- semantic types at the proper levels. +-- +-- I think the levels should look something like this: +-- +-- 0: Tokens +-- 1: POS Tags +-- 2/3: Chunks +-- 3/2: NER tags + -- | Data type to indicate IOB tags for chunking data IOBChunk chunk tag = BChunk (POS tag) chunk -- ^ Beging marker. | IChunk (POS tag) chunk -- ^ In chunk tag @@ -46,10 +59,12 @@ -- parseIOBLine :: (ChunkTag chunk, Tag tag) => Text -> Either Error (IOBChunk chunk tag) parseIOBLine txt = - let (tokTxt:tagTxt:iobTxt:_) = T.words txt - token = Token tokTxt - tag = POS (parseTag tagTxt) token - in iobBuilder iobTxt tag + case T.words txt of + (tokTxt:tagTxt:iobTxt:_) -> + let token = Token tokTxt + tag = POS (parseTag tagTxt) token + in iobBuilder iobTxt tag + _ -> Left ("not enough words in IOB line: \"" <> txt <> "\"") iobBuilder :: (ChunkTag c, Tag t) => Text -> (POS t -> Either Error (IOBChunk c t)) iobBuilder iobTxt | "I-" `T.isPrefixOf` iobTxt = \tag -> (IChunk tag) <$> chunk diff --git a/src/NLP/Types/Tags.hs b/src/NLP/Types/Tags.hs --- a/src/NLP/Types/Tags.hs +++ b/src/NLP/Types/Tags.hs @@ -9,11 +9,22 @@ import qualified Data.Text as T import Data.Text.Encoding (encodeUtf8, decodeUtf8) import GHC.Generics +import Text.Read (readEither) + import Test.QuickCheck (Arbitrary(..), NonEmptyList(..)) import Test.QuickCheck.Instances () -import NLP.Types.General (Error) +import NLP.Types.General (Error, toEitherErr) + +-- | The class of named entity sets. This typeclass can be defined +-- entirely in terms of the required class constraints. +class (Ord a, Eq a, Read a, Show a, Generic a, Serialize a) => NERTag a where + fromNERTag :: a -> Text + fromNERTag = T.pack . show + + parseNERTag :: Text -> Either Error a + parseNERTag txt = toEitherErr $ readEither $ T.unpack txt -- | The class of things that can be regarded as 'chunks'; Chunk tags -- are much like POS tags, but should not be confused. Generally, diff --git a/tests/src/Bench.hs b/tests/src/Bench.hs deleted file mode 100644 --- a/tests/src/Bench.hs +++ /dev/null @@ -1,62 +0,0 @@ -{-# LANGUAGE PackageImports #-} -{-# LANGUAGE OverloadedStrings #-} -module Bench where - -import Data.Text (Text) -import qualified Data.Text as T -import qualified Data.Text.IO as T - -import Control.DeepSeq -import Criterion.Main -import Criterion.Config (defaultConfig, Config(..), ljust) -import Criterion (bench, bgroup, Benchmark) - -import NLP.POS (tagText) -import NLP.POS.AvgPerceptronTagger (trainNew, mkTagger) -import Corpora -import NLP.Corpora.Email - -import qualified NLP.Similarity.VectorSimBench as VS - -myConfig :: Config -myConfig = defaultConfig { - -- Always GC between runs. - cfgPerformGC = ljust True - } - -main :: IO () -main = do --- postagBench <- posTagging - muc3_1 <- VS.muc3_01 - muc3_2 <- VS.muc3_02 - muc3_3 <- VS.muc3_03 - pTxt <- plugArchiveText - let len = length pTxt - plugTxt = take (len `div` 4) pTxt - plugStr = map T.unpack plugTxt - deepseq plugStr $ defaultMainWith myConfig (return ()) - [ - -- bgroup "POS Tagging" [] -- postagBench - -- , bgroup "Similarity" $ VS.benchmarks (muc3_1++muc3_2) muc3_3 - ] - --- posTagging :: IO [Benchmark] --- posTagging = do --- ca01 <- T.readFile brownCA01 --- ca02 <- T.readFile (brownCAFiles!!1) --- let ca1_2 = T.unlines [ca01, ca02] --- return [ bench "Train Brown ca01" $ trainNew ca01 --- , bench "Train & test Brown ca01" $ trainAndTag ca01 "the dog jumped" - --- , bench "Train Brown ca02" $ trainNew ca02 --- , bench "Train & test Brown ca02" $ trainAndTag ca02 "the dog jumped" - --- , bench "Train Brown ca01-02" $ trainNew ca1_2 --- , bench "Train & test Brown ca01-02" $ trainAndTag ca1_2 "the dog jumped" --- ] - -trainAndTag :: Text -> Text -> IO Text -trainAndTag corpus input = do - tagger <- trainNew corpus - return $ tagText (mkTagger tagger Nothing) input - diff --git a/tests/src/NLP/Similarity/VectorSimBench.hs b/tests/src/NLP/Similarity/VectorSimBench.hs deleted file mode 100644 --- a/tests/src/NLP/Similarity/VectorSimBench.hs +++ /dev/null @@ -1,65 +0,0 @@ -{-# LANGUAGE PackageImports #-} -{-# LANGUAGE OverloadedStrings #-} -module NLP.Similarity.VectorSimBench where - -import Data.List.Split (splitWhen) -import Data.Text (Text) -import qualified Data.Text as T -import qualified Data.Text.IO as T -import Criterion (bench, whnf, Benchmark) - -import NLP.Tokenize.Text (tokenize) -import NLP.Similarity.VectorSim -import NLP.Types (mkCorpus, Corpus) - -benchmarks :: [[Text]] -> [[Text]] -> [Benchmark] -benchmarks docs testDocs = let - corpus = mkCorpus docs - in [ bench "Doc 1-2 vs 3-4" $ whnf (similarity corpus (concat $ take 2 testDocs)) - ((testDocs!!2) ++ (testDocs!!3)) - , bench "Doc 1-5 vs 6-10" $ whnf (similarity corpus (concat $ take 5 testDocs)) - (concat $ take 5 $ drop 5 testDocs) - , bench "all pairs of 1-5" $ whnf (docsRunAllPairs corpus) (take 5 testDocs) - - , bench "TV all pairs of 1-5" $ whnf (tvDocsRunAllPairs corpus) (take 5 testDocs) - ] - -docsRunAllPairs :: Corpus -> [[Text]] -> Double -docsRunAllPairs _ [] = 0 -docsRunAllPairs corpus (d:ds) = let - firstRow = foldl (\v doc -> v + similarity corpus d doc) 0 ds - in firstRow + (docsRunAllPairs corpus ds) - -tvDocsRunAllPairs :: Corpus -> [[Text]] -> Double -tvDocsRunAllPairs corpus docs = runVectors (map (mkVector corpus) docs) - where - runVectors :: [TermVector] -> Double - runVectors [] = 0 - runVectors (d:ds) = let - firstRow = foldl (\v doc -> v + tvSim d doc) 0 ds - in firstRow + (runVectors ds) - - -readMucCorpus :: String -> IO [[Text]] -readMucCorpus file = do - content <- T.readFile ("./tests/resources/corpora/muc3_4/"++file) - let - docMarker :: Text -> Bool - docMarker txt = "DEV-MUC3-" `T.isPrefixOf` txt - - docLines :: [[Text]] - docLines = splitWhen docMarker $ T.lines content - - documents :: [Text] - documents = map T.unlines docLines - - return $ map tokenize documents - -muc3_01 :: IO [[Text]] -muc3_01 = readMucCorpus "dev-muc3-0001-0100" - -muc3_02 :: IO [[Text]] -muc3_02 = readMucCorpus "dev-muc3-0101-0200" - -muc3_03 :: IO [[Text]] -muc3_03 = readMucCorpus "dev-muc3-0201-0300"