sent_analysis.py

# -*- coding: utf-8 -*-
"""
Created on Thu Apr 24 13:33:02 2014

@author: Hussam Hamdan

Modified version by Gaël Guibon (command line interface added + speed optimization (~15% for tweets, ~40% for reviews))
"""
#ignore scikit warnings
def warn(*args, **kwargs):
    pass
import warnings
warnings.warn = warn

import re
import happy_tokeniser as htok
import z_score as zscore
from collections import OrderedDict
import utility as ut
import get_senti_from_lexicon as sentii
from sklearn.externals import joblib
from sklearn import svm, cross_validation
from sklearn.datasets import load_svmlight_file
tokeniser = htok.Tokenizer(preserve_case=False)
import shutil, tempfile, os, json, argparse, sys, time

# argparse added by Gael Guibon
parser = argparse.ArgumentParser(description='echo by Hussam Hamdam. Forked by Gaël Guibon in order to add a CLI and speed optimization\n Sentiment analysis classifier by polarity.')
parser.add_argument('-train','--train', metavar='TRAIN', type=str, help='train file path')
parser.add_argument('-c','--corpus', metavar='MODES', type=str, help='modes; "txt" for text or "tw" for tweets)')
parser.add_argument('-test','--test', metavar='TEST', type=str, help='test file path')
parser.add_argument('-f','--feature', metavar='FEATS', type=str, help='type of features : "zs" for z-score, "pol" for polarity or "dic" for twitterDictionary or combine them "zs+pol+dic"')
parser.add_argument('-t', '--trainingFlag', action='store_true', help='use this flag to enable training')
parser.add_argument('-v', '--verbose', action='store_true', help='use this flag to enable progressionBar (will slightly slow computation)')
parser.add_argument('-o','--output', metavar='OUTPUT', type=str, help='output file path')

args = parser.parse_args()


class Echo():

    def __init__(self, training=False, vocab_path='./data/reviewvocab.txt', model_path='./model/review.pkl'):
        if training:
            print "echo start whithout model loaded"
        else:
            print "Model Loading ..."
            self.vocabhash = self.loadVocbFile(vocab_path)
            self.classifier1 = joblib.load(model_path)
            self.rep = {"\t": " ", "\r": "", "\n": "", "\\u002c": ",", "\\u2019": "'", "\\u2013": "-" ,  "\\u2013": "-"}
            self.rep = dict((re.escape(k), v) for k, v in self.rep.iteritems())
            self.pattern = re.compile("|".join(self.rep.keys()))

    # ## added by GG to monitor speed
    def stopWatch(self, value):
        '''From seconds to Days;Hours:Minutes;Seconds'''
        valueD = (((value/365)/24)/60)
        Days = int (valueD)
        valueH = (valueD-Days)*365
        Hours = int(valueH)
        valueM = (valueH - Hours)*24
        Minutes = int(valueM)
        valueS = (valueM - Minutes)*60
        Seconds = int(valueS)
        print Days,"days ;",Hours,"hours :",Minutes,"minutes ;",Seconds, "seconds"

    # added by GG to show the progress
    def progressBar( self, value, endvalue, bar_length=100):
        '''print the progress bar given the values.
        default bar_length = 100'''
        percent = float(value) / endvalue
        arrow = '-' * int(round(percent * bar_length)-1) + '>'
        spaces = ' ' * (bar_length - len(arrow))
        sys.stdout.write("\rPercent: [{0}] {1}%".format(arrow + spaces, int(round(percent * 100))))
        sys.stdout.flush()

    def readFile(self, filename):
        content = open(filename, "r").readlines()
        data = []
        labels = []
        i = 0
        count = [0, 0, 0]
        for line in content:
            i += 1
            #line=unicodedata.normalize('NFKD', line).encode('ascii','ignore')
            line = line.split("\t")
            line[3] = line[3].replace("\n", "")
            line[2] = 2 if line[2] == "positive" else 0 if line[2] == "negative" else 1
            count[line[2]] += 1
            data.append(line[3])
            labels.append(line[2])
        return data, labels

    def loadDic(self, path, corpus):
        if corpus=='tw':
            f = open(path + "data/twitterDic.txt", "r")
        elif corpus=='txt':
            f = open(path + "data/twitterDic.txt", "r")
        dicth = dict()
        lines = f.readlines()
        for line in lines:
            if line.find(" = ") != -1:
                line = line.split(" =")
                dicth[line[0].lower()] = line[1].replace("\n", "").lower()
        return dicth

    def getpolarityid(self, polarity):
        map_polarities = {"0":"negative", "1":"neutral", "2":"positive"}
        return map_polarities[str(int(polarity))]

    def getPolaritynam(self, polarity):
        map_polarities = {"0":"negative", "1":"neutral", "2":"positive"}
        return map_polarities[str(int(polarity))]

    def splitfun4tweetGG(self, doc):
        doc = doc.decode('utf8', errors='ignore').encode('ascii', errors='ignore')

        doc = self.pattern.sub(lambda m: self.rep[re.escape(m.group(0))], doc.lower())

        y = re.findall(r'@[a-z0-9]+', doc)
        for y1 in y:
            doc = doc.replace(y1, "uuser")
        y = re.findall(r'http://[a-z0-9./]+', doc)
        for y1 in y:
            doc = doc.replace(y1, "http")
        res = tokeniser.tokenize(doc)

        ## removed these two test conditions
        if "\u2013" in res:
            print res
        if "\\u2013" in res:
            print res
            print "h"

        ## faster append + faster string concatenation
        bires = []
        append_bires = bires.append
        for i in xrange(len(res)-1):
            append_bires(  " ".join([ res[i], res[i + 1] ])  )
        return res + bires

    def predictzvalues(self, text, zdict_3, threshold=3):
        sumz = [0, 0, 0, 0]
        text = self.splitfun4tweetGG(text)
        for token in text:
            if (len(token) > 1):
                for i in xrange(len(zdict_3)):
                    if ((token in zdict_3[0]) and zdict_3[i][token]>=threshold):
                        sumz[i] += 1
        return sumz

    def getTrainingFile(self, data, labels, vocabfile, textfile, prepolarity, Z_score, POS, TwitterDict, corpus, tokenlngth=1, zthreshold=3, path='./'):
        '''This function take the data and their labels with some options in order to construct a training file
        data: the orginal textual tweets
        labels: the polarity of each tweet
        vocabfile: in which the vocabulary is indexed and saved
        textfile: the output training file
        prepolarity,Z_score,POS,TwitterDict : if we want using the prepolarity, z score, pos, twitter dictionary features.
        it is not recommended to use POS.'''
        f = open(textfile, "w")
        vocf = open(vocabfile, "w")
        vocdict = dict()
        i = 0
        if prepolarity:
            vocdict["positive1"] = i; i+=1
            vocdict["negative1"] = i; i+=1
            vocdict["neutral1"] = i; i+=1
            lexdic = sentii.loadLexicon()
            ludic = sentii.loadLU()
        if Z_score:
            vocdict["zscore1"] = i; i+=1
            vocdict["zscore2"] = i; i+=1
            vocdict["zscore3"] = i; i+=1
            vocdict["allzscore"] = i; i+=1
            z_dict = zscore.loadzscore(corpus, path)
        if POS:
            vocdict["pos0"] = i; i+=1
            vocdict["pos1"] = i; i+=1
            vocdict["pos2"] = i; i+=1
            vocdict["pos3"] = i; i+=1
            vocdict["pos4"] = i; i+=1

        # tweet dictionary loading (expression,emotion icons)
        if TwitterDict:
            dicht = self.loadDic(path, corpus)
        vocid = i

        for i in xrange(len(data)):
            text = data[i]
            polarity = labels[i]
            if POS:
                pos = ut.getPosTags(text)
                postags = [0, 0, 0, 0, 0]
                paspect = str(pos)
                postags[0] = len(paspect.split("'NN")) - 1
                postags[1] = len(paspect.split("'JJ")) - 1
                postags[2] = len(paspect.split("'RB'")) - 1
                postags[3] = len(paspect.split("'VB")) - 1
                postags[4] = len(paspect.split("'CC'")) - 1
            if Z_score:
                zsum = self.predictzvalues(text, z_dict, zthreshold)
            # add the emotions
            if TwitterDict:
                for k, v in dicht.iteritems():
                    if text.find(k) != -1:
                        text += " " + v
            newtxt = self.splitfun4tweetGG(text)
            instance = ""
            sentencedict = dict()
            priorpol = [0, 0, 0, 0]
            for tok in newtxt:
                if len(tok) > tokenlngth:
                    if prepolarity:
                        pol1 = sentii.getWordsenti(tok, ludic)
                        pol2 = sentii.getWordsenti(tok, lexdic)
                        priorpol[pol2 + 1] += 1  # 61 +1 alone
                        priorpol[pol1 + 1] += 1
                    index = -1
                    if vocdict.has_key(tok):
                        index = vocdict[tok]
                    else:
                        vocdict[tok] = vocid
                        index = vocid
                        vocid += 1
                    if sentencedict.has_key(index):
                        sentencedict[index] += 1
                    else:
                        sentencedict[index] = 1
            i = 0
            instance = ""
            if prepolarity:
                instance += " %d:%d %d:%d %d:%d" % (i, priorpol[0], i+1, priorpol[1], i+2, priorpol[2])
                i = i + 3
            if Z_score:
                instance += " %d:%d %d:%d %d:%d %d:%d" % (i, zsum[0], i+1, zsum[1], i+2, zsum[2], i+3, zsum.index(max(zsum)))
                i = i + 4
            if POS:
                instance += " %d:%d %d:%d %d:%d %d:%d %d:%d" % (i, postags[0], i+1, postags[1], i+2, postags[2], i+3, postags[3], i+4, postags[4])
                i = i + 5
            d_sorted_by_value = OrderedDict(sorted(sentencedict.items(), key=lambda x: x[0]))
            for k, v in d_sorted_by_value.iteritems():
                instance += " " + str(k) + ":" + str(v)
            f.write(str(polarity) + instance + "\n")
        line = ""
        d_sorted_by_value = OrderedDict(sorted(vocdict.items(), key=lambda x: x[1]))
        for k, v in d_sorted_by_value.iteritems():
            line += str(k.encode("ascii", "ignore")) + "\t" + str(v) + "\n"
        vocf.write(line)
        vocf.close()
        f.close()

    # duplicate for monitoring and speed purpose by GG
    def getTestFile(self, path, inputfile, outputsvmfile, inputvocab, classifier, prepolarity, Z_score, POS, TwitterDict, corpus, zthreshold=3, tokenlngth=1):
        '''This function generates the test file, it takes the test file path and the vocabulary file (inputvocab), the classifier (classifier) and the options
        in order to predict the polarity of each tweet in the test file'''

        # pol is always false in getresult() --> removed
        if prepolarity:
            lexdic = sentii.loadLexicon()
            ludic = sentii.loadLU()
            lexswn, lexposswn = sentii.loadswn()

        # zs is always True in getresult() --> condition removed
        if Z_score:
            z_dict = zscore.loadzscore(corpus, path)
        
        # removed because never used
        if TwitterDict:
            dicht = self.loadDic(path, corpus)
        index = 0
        f = open(inputfile, "r")
        lines = f.readlines()
        all_txt = ""

        ## added in order to fasten the concatenation by using cpython
        all_txt_list = list()
        append = all_txt_list.append
        lower = str.lower

        resfile = open(outputsvmfile, "w")
        resfile.close()
        with open(outputsvmfile, "a") as resfile:

            total = len(lines)
            for indexLine, row1 in enumerate(lines):
                # self.progressBar(indexLine, total)
                row1 = row1.replace("\n", "").split("\t")
                if len(row1) < 1 :  exit(0)
                id1 = row1[0]
                id2 = row1[1]

                sentence = lower(row1[3])

                zsum = [0, 0, 0, 0]
                
                ## zs if always true in getresult --> removed the condition
                if Z_score:
                    zsum = self.predictzvalues(sentence, z_dict, zthreshold)
                # zsum = self.predictzvalues(sentence, z_dict, zthreshold)

                rowhash = {}
                count = 0

                ## dic is always false in getresult --> condition never used --> removed
                if TwitterDict:
                    for k, v in dicht.iteritems():
                        if sentence.find(k) != -1:
                            sentence += " " + v

                newt = self.splitfun4tweetGG(sentence)
                priorpol = [0, 0, 0, 0]
                postags = [0, 0, 0, 0, 0]

                ## pos is always false in getresult --> condition never used --> removed
                if POS:
                    aspect = ut.getPosTags(sentence)
                    paspect = str(aspect)
                    postags[0] = len(paspect.split("'NN")) - 1
                    postags[1] = len(paspect.split("'JJ")) - 1
                    postags[2] = len(paspect.split("'RB'")) - 1
                    postags[3] = len(paspect.split("'VB")) - 1
                    postags[4] = len(paspect.split("'CC'")) - 1

                for token in newt:
                    if (len(token) > tokenlngth):
                        ## pol is always false in getresult --> condition never used --> removed
                        if prepolarity:
                            pol1 = sentii.getWordsenti(token, ludic)
                            pol2 = sentii.getWordsenti(token, lexdic)
                            priorpol[pol2 + 1] += 1  # 61 +1 alone
                            priorpol[pol1 + 1] += 1  # alone 61 +1
                        count += 1
                        if inputvocab.has_key(token):
                            index = inputvocab[token]
                        # filling the hashtable for each file index:number of occurence
                            if rowhash.has_key(index):
                                rowhash[index] += 1
                            else:
                                rowhash[index] = 1
                x_test = [0 for i in xrange(len(inputvocab))]
                d_sorted_by_value = OrderedDict(sorted(rowhash.items(), key=lambda x: x[0]))
                for k, v in d_sorted_by_value.iteritems():
                    x_test[k] = v

                i = 0

                ## pol is always false in getresult() --> removed because never used
                if prepolarity:
                    x_test[i], x_test[i+1], x_test[i+2] = (priorpol[0], priorpol[1], priorpol[2])
                    i = i + 3

                ## zs is always true in getresult() --> condition removed
                if Z_score:
                    x_test[i], x_test[i+1], x_test[i+2], x_test[i+3] = zsum[0], zsum[1], zsum[2], zsum.index(max(zsum))
                    i = i + 4

                ## pos is always false in getresult() --> removed because never used
                if POS:
                    x_test[i], x_test[i+1], x_test[i+2], x_test[i+3], x_test[i+4] = postags[0], postags[1], postags[2], postags[3], postags[4]
                    i = i + 5

                y_pred = classifier.predict(x_test)
                append(   "\t".join([ id1, id2, self.getPolaritynam(y_pred[0]), row1[3] ]) + "\n" )
                resfile.write( "\t".join([ id1, id2, self.getPolaritynam(y_pred[0]), row1[3] ]) + "\n" )

    def getFileName(self, prepolarity, z_score, POS, Twittdic):
        fn = "t"
        if z_score:
            fn += "-z"
        if prepolarity:
            fn += "-pol"
        if POS:
            fn += "-pos"
        if Twittdic:
            fn += "-dic"
        return fn


    def loadVocbFile(self, vocfile):
        hasht = dict()
        f = open(vocfile, "r")
        lines = f.readlines()
        i = 0
        for index, line in enumerate(lines):
            voc = line.split("\t")
            hasht[voc[0]] = index
        f.close()
        return hasht

    # duplicate version for speed and monitoring purpose by GG
    def writeInputFile(self, txt_lst, filename):

        with open(filename, "a") as myfile:
            for i, line in enumerate(txt_lst):
                # self.progressBar(i, len(txt_lst))
                myfile.write( "NA\t%s\tunknwn\t%s\n" % (str(txt_lst.index(line)), line) )

    ## added a modified version by GG to try going faster and monitoring
    def getResult(self, txt_lst, corpus='txt', option='zs', training='N'):
        """Tagged a list of sentence in positive, negative or neutral opinion
        @param : List[string] ; the text (list of sentence) that you want to tagged. Each value of the list is a sentence.
        @return: List[(set)] ; which the set is (sentence, opinion)"""

        print 1
        inp_dir = tempfile.mkdtemp(prefix = 'data_sent_analysis')
        print 2
        out_dir = tempfile.mkdtemp(prefix='eval_sent_anlysis')
        print 3
        inputf = os.path.join(inp_dir, 'tmpdata.txt')
        print 4, "list st.encode\n"
        lst_tmp = [st.encode('utf-8') for st in txt_lst]
        txt_lst = None # empty the list to save memory
        print 5, "writeInputFile\n"
        self.writeInputFile(lst_tmp, inputf)
        lst_tmp = None # empty the list to save memory
        print 6, "outputf\n"
        outputf = os.path.join(out_dir, 'tmpeval.txt')
        print 7, "Predicting ...\n"
        pol = pos = dic = False
        zs = True
        print 8, "getTestFile\n"
        self.getTestFile('./', inputf, outputf, self.vocabhash, self.classifier1, pol, zs, pos, dic, corpus)
        print 9, "noutputf\n"
        with open(outputf) as f:
            resultat = []

            append_resultat = resultat.append
            i = 0
            lines = f.readlines()
            total = len(lines)
            
            for line in lines:
                line_in_list = line.split("\t")
                if (line_in_list[2] and line_in_list[3]) is not None:
                    append_resultat(  (line_in_list[2].decode('utf-8'), (line_in_list[3].replace('\n', '')).decode('utf-8'))  )

        os.unlink(outputf)
        os.unlink(inputf)
        shutil.rmtree(inp_dir)
        shutil.rmtree(out_dir)
        return resultat



# modified by Gael Guibon 
if __name__ == '__main__':
    
    # time starting point
    startTime = time.time()

    zs = dic = pol = pos = False
    echo = Echo()
    if args.corpus=='tw':
        filename="./corpus/twitter-train-cleansed-B.txt"
        svmfname = "data/tweet.txt"
        data, labels = echo.readFile(args.train)    
        x = args.feature
        x = x.split('+')
        print x
        y = args.trainingFlag
        if "zs" in x:
            zs = True
        if "pol" in x:
            pol = True
        if "dic" in x:
            dic = True
        if "dic" not in x and "pol" not in x and "zs" not in x: raise NameError('Invalid Feature Option')
        # else: raise NameError('Invalid Feature Option')

        print zs, pol, dic

        if y==False:
            print "Model Loading  ..."
            vocabhash = echo.loadVocbFile("data/tweetvocab"+echo.getFileName(pol,zs,pos,dic)+".txt")
            outf = "eval/hx" + echo.getFileName(pol, zs, pos, dic) + ".txt"
            classifier1 = joblib.load(echo.getFileName(pol, zs, pos, dic) + ".pkl")
        elif y==True:
            print "preprocessing ..."
            echo.getTrainingFile(data, labels, "data/tweetvocab" + echo.getFileName(pol,zs,pos,dic) + ".txt", svmfname, pol, zs, pos, dic, args.corpus)
            vocabhash = echo.loadVocbFile("data/tweetvocab" + echo.getFileName(pol,zs,pos,dic) + ".txt")

            if args.output: 
                outf = os.path.abspath(args.output)
            else:
                outf = "eval/hx" + echo.getFileName(pol, zs, pos, dic) + ".txt"    
            print "training"
            x_train, y_train = load_svmlight_file(svmfname)
            classifier1 = svm.LinearSVC()
            classifier1.fit(x_train, y_train)
            joblib.dump(classifier1, echo.getFileName(pol,zs,pos,dic) + ".pkl")
        else: raise NameError('Invalid Option: training or not training?')

        print "Predicting ..."
        echo.getTestFile('./', args.test, outf, vocabhash, classifier1, pol, zs, pos, dic, args.corpus)
        print "Evaluation ..."
        prog = "perl eval/score-semeval2014-task9-subtaskB.pl " + outf
        import subprocess
        p = subprocess.Popen(prog, shell=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE).communicate()[0]
        print(p)


    elif args.corpus=='txt':
        filename = './corpus/review.txt'
        data, labels = echo.readFile(args.train)
        svmfname = "trainab"
        zthreshold=-0.5
        zs = True # For corpus text: only zscore is avalaible
        print "preprocessing ..."
        echo.getTrainingFile(data, labels, "data/reviewvocab.txt", svmfname, pol, zs, pos, dic, args.corpus, zthreshold)
        vocabhash = echo.loadVocbFile("data/reviewvocab.txt")
        outf = "eval/review-z.txt"
        print "Training ..."
        x_train, y_train = load_svmlight_file(svmfname)
        classifier1 = svm.LinearSVC()
        classifier1.fit(x_train, y_train)
        joblib.dump(classifier1, "review.pkl")
        print "predicting"
        echo.getTestFile('./', args.test, outf, vocabhash, classifier1, pol, zs, pos, dic, args.corpus)
        print "Evaluation ..."
        scores = cross_validation.cross_val_score(classifier1, x_train, y_train, cv=5)
        print scores


    else: raise NameError('Invalid Option : Please use "python sent_analysis.py -h" to see all available options')

    echo.stopWatch(time.time() - startTime)