From f3ee36cb955783280b29fa2a898761a9e75a4c87 Mon Sep 17 00:00:00 2001 From: Carsten Date: Thu, 3 May 2018 14:00:58 +0200 Subject: [PATCH] task 2 angefangen --- Carsten_Solutions/Exercise 1.ipynb | 431 ++++++++++++++++++++++++++--- 1 file changed, 388 insertions(+), 43 deletions(-) diff --git a/Carsten_Solutions/Exercise 1.ipynb b/Carsten_Solutions/Exercise 1.ipynb index 4f910d0..f4c77c0 100644 --- a/Carsten_Solutions/Exercise 1.ipynb +++ b/Carsten_Solutions/Exercise 1.ipynb @@ -9,12 +9,13 @@ }, { "cell_type": "code", - "execution_count": 39, + "execution_count": 23, "metadata": { "collapsed": true }, "outputs": [], "source": [ + "import numpy as np\n", "import nltk\n", "from nltk import word_tokenize, pos_tag" ] @@ -29,7 +30,7 @@ }, { "cell_type": "code", - "execution_count": 40, + "execution_count": 24, "metadata": { "collapsed": true }, @@ -49,14 +50,14 @@ }, { "cell_type": "code", - "execution_count": 41, + "execution_count": 25, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ - "{'word': 'bims', 'length': 4, 'is_capitalized': False, 'prefix-1': 'b', 'suffix-1': 's', 'prev_word': 'i', 'next_word': 'der'}\n" + "{'word': 'bims', 'length': 4, 'is_capitalized': False, 'prefix-1': 'b', 'suffix-1': 's', 'prev_word': 'i', 'next_word': 'der', 'kindOfCamelCase': False, 'includesSpace': False}\n" ] } ], @@ -69,7 +70,9 @@ " 'prefix-1': sentence[index][0],\n", " 'suffix-1': sentence[index][-1],\n", " 'prev_word': '' if index == 0 else sentence[index - 1],\n", - " 'next_word': '' if index == len(sentence) - 1 else sentence[index + 1]\n", + " 'next_word': '' if index == len(sentence) - 1 else sentence[index + 1],\n", + " 'kindOfCamelCase': sentence[index][1:].lower() != sentence[index][1:],\n", + " 'includesSpace': True if ((' ') in sentence[index]) else False #depemds on tokenizer\n", " }\n", "\n", "print(features(\"halli hallo i bims der Programmierer\".strip().split(\" \"), 3))" @@ -84,18 +87,27 @@ }, { "cell_type": "code", - "execution_count": null, + "execution_count": 26, "metadata": { "collapsed": true }, "outputs": [], - "source": [] + "source": [ + "#max entropie pre trained pos tag\n", + "#see Calculate performance 1.2" + ] }, { "cell_type": "markdown", "metadata": {}, "source": [ - "### 3. model3.x = rule-based classifiers (x = 1 to 5)" + "### 3. model3.x = rule-based classifiers (x = 1 to 5)\n", + "1. DefaultTagger that simply tags everything with the same tag\n", + "2. RegexpTagger that applies tags according to a set of regular expressions\n", + "3. N-Gram (n-gram tagger is a generalization of a unigram tagger whose context is the current word together with the part-of-speech tags of the n-1 preceding token)\n", + " + UnigramTagger\n", + " + BigramTagger\n", + " + TrigramTagger" ] }, { @@ -105,7 +117,18 @@ "collapsed": true }, "outputs": [], - "source": [] + "source": [ + "#used from description for RegexpTagger\n", + "patterns = [(r'.*ing$', 'VBG'), (r'.*ed$', 'VBD'), (r'.*es$', 'VBZ'), (r'.*ould$', 'MD'), (r'.*\\'s$', 'NN$'), \n", + " (r'.*s$', 'NNS'), (r'^-?[0-9]+(.[0-9]+)?$', 'CD'), (r'.*', 'NN')]\n", + "\n", + "#train taggers\n", + "def_model = nltk.DefaultTagger('NN')\n", + "regexp_model = nltk.RegexpTagger(patterns)\n", + "uni_model = nltk.UnigramTagger(training_sentences_X1)\n", + "bi_model = nltk.BigramTagger(training_sentences_X1)\n", + "tri_model = nltk.TrigramTagger(training_sentences_X1)" + ] }, { "cell_type": "markdown", @@ -156,7 +179,7 @@ }, { "cell_type": "code", - "execution_count": 42, + "execution_count": 27, "metadata": {}, "outputs": [ { @@ -183,7 +206,7 @@ }, { "cell_type": "code", - "execution_count": 43, + "execution_count": 28, "metadata": {}, "outputs": [ { @@ -209,14 +232,70 @@ }, { "cell_type": "code", - "execution_count": 44, + "execution_count": 64, + "metadata": {}, + "outputs": [ + { + "data": { + "text/plain": [ + "'#import pandas as pd\\n#df = pd.read_table(\"/Users/Carsten/GitRepos/NLP-LAB/Carsten_Solutions/sets/croatia/set.hr.conll\")\\n#df.head()\\n\\n#x3 = other language\\n#from croatia:\\n#by ZˇeljkoAgic ́,⋆NikolaLjubesˇic ́ http://www.lrec-conf.org/proceedings/lrec2014/pdf/690_Paper.pdf\\n#licenses: https://creativecommons.org/licenses/by-sa/4.0/\\ncorp = nltk.corpus.ConllCorpusReader(root=\"/Users/Carsten/GitRepos/NLP-LAB/Carsten_Solutions/sets/croatia/\", fileids=[\"set.hr.conll\"], columntypes=(\\'ignore\\',\\'ignore\\',\\'pos\\',\\'ignore\\',\\'ignore\\',\\'ignore\\',\\'ignore\\',\\'ignore\\',\\'ignore\\',\\'ignore\\'))\\nprint(corp.tagged_sents[-100])\\n#from croatia:\\n#by ZˇeljkoAgic ́,⋆NikolaLjubesˇic ́ http://www.lrec-conf.org/proceedings/lrec2014/pdf/690_Paper.pdf\\n#licenses: https://creativecommons.org/licenses/by-sa/4.0/'" + ] + }, + "execution_count": 64, + "metadata": {}, + "output_type": "execute_result" + } + ], + "source": [ + "'''#import pandas as pd\n", + "#df = pd.read_table(\"/Users/Carsten/GitRepos/NLP-LAB/Carsten_Solutions/sets/croatia/set.hr.conll\")\n", + "#df.head()\n", + "\n", + "#x3 = other language\n", + "#from croatia:\n", + "#by ZˇeljkoAgic ́,⋆NikolaLjubesˇic ́ http://www.lrec-conf.org/proceedings/lrec2014/pdf/690_Paper.pdf\n", + "#licenses: https://creativecommons.org/licenses/by-sa/4.0/\n", + "corp = nltk.corpus.ConllCorpusReader(root=\"/Users/Carsten/GitRepos/NLP-LAB/Carsten_Solutions/sets/croatia/\", fileids=[\"set.hr.conll\"], columntypes=('ignore','ignore','pos','ignore','ignore','ignore','ignore','ignore','ignore','ignore'))\n", + "print(corp.tagged_sents[-100])\n", + "#from croatia:\n", + "#by ZˇeljkoAgic ́,⋆NikolaLjubesˇic ́ http://www.lrec-conf.org/proceedings/lrec2014/pdf/690_Paper.pdf\n", + "#licenses: https://creativecommons.org/licenses/by-sa/4.0/'''\n" + ] + }, + { + "cell_type": "code", + "execution_count": 30, "metadata": { "collapsed": true }, "outputs": [], "source": [ - "#nltk.download('brown')\n", - "#x3 = other language" + "#? nltk.corpus.ConllCorpusReader" + ] + }, + { + "cell_type": "code", + "execution_count": 52, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "\n", + "50472\n", + "[('So', 'ADV'), ('kann', 'VMFIN'), ('man', 'PIS'), ('Marsilius', 'NE'), ('von', 'APPR'), ('Padua', 'NE'), ('so', 'ADV'), ('wenig', 'ADV'), ('zu', 'APPR'), ('einem', 'ART'), ('Vorläufer', 'NN'), ('moderner', 'ADJA'), ('Volkssouveränität', 'NN'), ('machen', 'VVINF'), ('wie', 'KOKOM'), ('Rousseau', 'NE'), ('zum', 'APPRART'), ('Verkünder', 'NN'), ('eines', 'ART'), ('``', '$('), ('Zurück', 'NN'), ('zur', 'APPRART'), ('Natur', 'NN'), (\"''\", '$('), ('.', '$.')]\n" + ] + } + ], + "source": [ + "# TODO: loading german corpus \n", + "X3 = nltk.corpus.ConllCorpusReader(root='/Users/Carsten/GitRepos/NLP-LAB/Carsten_Solutions/sets/german/', fileids=['tiger_release_aug07.corrected.16012013.conll09'], columntypes=['ignore', 'words', 'ignore', 'ignore', 'pos'], encoding='utf-8')\n", + "german_tagged_sents = X3.tagged_sents()\n", + "print(type(german_tagged_sents))\n", + "print(len(german_tagged_sents))\n", + "\n", + "print (german_tagged_sents[-100])" ] }, { @@ -246,7 +325,7 @@ }, { "cell_type": "code", - "execution_count": 45, + "execution_count": 32, "metadata": { "collapsed": true }, @@ -259,18 +338,21 @@ }, { "cell_type": "code", - "execution_count": 46, + "execution_count": 65, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ + "\n", "got 3131 training sentences and 783 test sentences\n" ] } ], "source": [ + "#print(type(nltk.corpus.treebank.tagged_sents()))\n", + "\n", "#object including the annotated sentences\n", "annotated_sent = nltk.corpus.treebank.tagged_sents()\n", "\n", @@ -285,7 +367,7 @@ }, { "cell_type": "code", - "execution_count": 47, + "execution_count": 34, "metadata": { "collapsed": true }, @@ -303,8 +385,10 @@ }, { "cell_type": "code", - "execution_count": 48, - "metadata": {}, + "execution_count": 35, + "metadata": { + "collapsed": true + }, "outputs": [], "source": [ "#trainings inputset X and training teacher set y\n", @@ -324,7 +408,7 @@ }, { "cell_type": "code", - "execution_count": 49, + "execution_count": 36, "metadata": {}, "outputs": [ { @@ -350,7 +434,7 @@ }, { "cell_type": "code", - "execution_count": 50, + "execution_count": 37, "metadata": { "collapsed": true }, @@ -360,6 +444,56 @@ "X2, y2 = transform_to_dataset(training_sentences_X2)" ] }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "##### Generate Training and Testdata for X3\n", + "1. split annotaed sentences into training and testdata\n", + "2. split trainingdata into input data and teacherdata\n", + " *input is the feature vector of each word\n", + " *output is a list of POS tags for each word and sentences" + ] + }, + { + "cell_type": "code", + "execution_count": 63, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "\n", + "got 40377 training sentences and 10095 test sentences\n" + ] + } + ], + "source": [ + "#object including the annotated sentences\n", + "annotated_sent = X3.tagged_sents()\n", + "\n", + "#print(type(annotated_sent))\n", + "\n", + "#to split the data, calculate the borders for ratio\n", + "cutoff = int(.8 * len(annotated_sent))\n", + "training_sentences_X3 = annotated_sent[:cutoff]\n", + "test_sentences_X3 = annotated_sent[cutoff:]\n", + "\n", + "#show the amount of sentences\n", + "print(\"got \",len(training_sentences_X3),\" training sentences and \", len(test_sentences_X3), \" test sentences\")" + ] + }, + { + "cell_type": "code", + "execution_count": 66, + "metadata": {}, + "outputs": [], + "source": [ + "#trainings inputset X and training teacher set y\n", + "X3, y3 = transform_to_dataset(training_sentences_X3)" + ] + }, { "cell_type": "markdown", "metadata": { @@ -374,7 +508,7 @@ }, { "cell_type": "code", - "execution_count": 51, + "execution_count": 40, "metadata": { "collapsed": true }, @@ -394,7 +528,7 @@ }, { "cell_type": "code", - "execution_count": 52, + "execution_count": 41, "metadata": { "collapsed": true }, @@ -417,7 +551,7 @@ }, { "cell_type": "code", - "execution_count": 53, + "execution_count": 42, "metadata": {}, "outputs": [ { @@ -425,13 +559,13 @@ "output_type": "stream", "text": [ "training OK\n", - "Accuracy: 0.87983432307\n" + "Accuracy: 0.880632766106\n" ] } ], "source": [ "size=10000\n", - "clf.fit(X[:size], y[:size])\n", + "clf.fit(X1[:size], y1[:size])\n", " \n", "print('training OK')\n", " \n", @@ -446,37 +580,206 @@ "cell_type": "markdown", "metadata": {}, "source": [ - "##### Calculate other performances" + "##### Calculate performance 1.2" ] }, { "cell_type": "code", - "execution_count": 58, + "execution_count": 43, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ - "calculated perfomance 1.4= 0.756485959481\n" + "Accuracy: 0.8936074654423873\n" ] } ], "source": [ + "#extract only the words from feature trainings set\n", + "only_words_X1 = [x['word'] for x in X1_test]\n", + "\n", + "#train with the pos tagger by nltk\n", + "pos_tags_by_pre_trained_pos_tagger = [word_tag_tuple[1] for word_tag_tuple in pos_tag(only_words_X1, lang='eng')]\n", + "\n", + "#calculate performance by comparing each pos tag\n", "performance1_2 = 0\n", + "for index in range(len(pos_tags_by_pre_trained_pos_tagger)):\n", + " if(pos_tags_by_pre_trained_pos_tagger[index]==y1_test[index]):\n", + " performance1_2 += 1\n", + "performance1_2 /= len(pos_tags_by_pre_trained_pos_tagger)\n", "\n", - "performance1_3 = 0\n", + "print(\"Accuracy:\", performance1_2)" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "##### Calculate performance 1.3\n", + "1. DefaultTagger that simply tags everything with the same tag\n", + "2. RegexpTagger that applies tags according to a set of regular expressions\n", + "3. N-Gram (n-gram tagger is a generalization of a unigram tagger whose context is the current word together with the part-of-speech tags of the n-1 preceding token)\n", + " + UnigramTagger\n", + " + BigramTagger\n", + " + TrigramTagger" + ] + }, + { + "cell_type": "code", + "execution_count": 44, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "performance 1.3.1 is: 0.1447677029791906\n", + "performance 1.3.2 is: 0.24232746145017217\n", + "performance 1.3.3 is: 0.8608213982733669\n", + "performance 1.3.4 is: 0.1132791057437996\n", + "performance 1.3.5 is: 0.06736863116922003\n" + ] + } + ], + "source": [ + "#evaluate taggers\n", + "# performance of Default Tagger\n", + "performance1_3_1 = def_model.evaluate(test_sentences_X1)\n", + "print('performance 1.3.1 is: ',performance1_3_1)\n", "\n", - "# performance1_4\n", + "# performance of Regex Tagger\n", + "performance1_3_2 = regexp_model.evaluate(test_sentences_X1)\n", + "print('performance 1.3.2 is: ',performance1_3_2)\n", + "\n", + "# performance of Unigram Tagger\n", + "performance1_3_3 = uni_model.evaluate(test_sentences_X1)\n", + "print('performance 1.3.3 is: ',performance1_3_3)\n", + "\n", + "# performance of Bigram Tagger\n", + "performance1_3_4 = bi_model.evaluate(test_sentences_X1)\n", + "print('performance 1.3.4 is: ',performance1_3_4)\n", + "\n", + "# performance of Trigram Tagger\n", + "performance1_3_5 = tri_model.evaluate(test_sentences_X1)\n", + "print('performance 1.3.5 is: ',performance1_3_5)" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "##### Calculate performance 1.4" + ] + }, + { + "cell_type": "code", + "execution_count": 45, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "calculated perfomance 1.4= 0.75680543774\n" + ] + } + ], + "source": [ "size=10000\n", "clf.fit(X2[:size], y2[:size])\n", "X2_test, y2_test = transform_to_dataset(test_sentences_X2)\n", "performance1_4 = clf.score(X2_test, y2_test)\n", - "print(\"calculated perfomance 1.4= \",performance1_4)\n", + "print(\"calculated perfomance 1.4= \",performance1_4)" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "##### Calculate performance 1.5" + ] + }, + { + "cell_type": "code", + "execution_count": 46, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "Accuracy: 0.6044583741861567\n" + ] + } + ], + "source": [ + "#extract only the words from feature trainings set\n", + "only_words_X2 = [x['word'] for x in X2_test]\n", "\n", + "#train with the pos tagger by nltk\n", + "pos_tags_by_pre_trained_pos_tagger = [word_tag_tuple[1] for word_tag_tuple in pos_tag(only_words_X2, lang='eng')]\n", + "\n", + "#calculate performance by comparing each pos tag\n", "performance1_5 = 0\n", + "for index in range(len(pos_tags_by_pre_trained_pos_tagger)):\n", + " if(pos_tags_by_pre_trained_pos_tagger[index]==y2_test[index]):\n", + " performance1_5 += 1\n", + "performance1_5 /= len(pos_tags_by_pre_trained_pos_tagger)\n", "\n", - "performance1_6 = 0" + "print(\"Accuracy:\", performance1_5)" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "##### Calculate performance 1.6" + ] + }, + { + "cell_type": "code", + "execution_count": 47, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "performance 1.6.1 is: 0.10997763652187324\n", + "performance 1.6.2 is: 0.17594438874995869\n", + "performance 1.6.3 is: 0.8773754310202373\n", + "performance 1.6.4 is: 0.3390490564374869\n", + "performance 1.6.5 is: 0.19178610379738467\n" + ] + } + ], + "source": [ + "uni_model = nltk.UnigramTagger(training_sentences_X2)\n", + "bi_model = nltk.BigramTagger(training_sentences_X2)\n", + "tri_model = nltk.TrigramTagger(training_sentences_X2)\n", + "\n", + "#evaluate taggers\n", + "# performance of Default Tagger\n", + "performance1_6_1 = def_model.evaluate(test_sentences_X2)\n", + "print('performance 1.6.1 is: ',performance1_6_1)\n", + "\n", + "# performance of Regex Tagger\n", + "performance1_6_2 = regexp_model.evaluate(test_sentences_X2)\n", + "print('performance 1.6.2 is: ',performance1_6_2)\n", + "\n", + "# performance of Unigram Tagger\n", + "performance1_6_3 = uni_model.evaluate(test_sentences_X2)\n", + "print('performance 1.6.3 is: ',performance1_6_3)\n", + "\n", + "# performance of Bigram Tagger\n", + "performance1_6_4 = bi_model.evaluate(test_sentences_X2)\n", + "print('performance 1.6.4 is: ',performance1_6_4)\n", + "\n", + "# performance of Trigram Tagger\n", + "performance1_6_5 = tri_model.evaluate(test_sentences_X2)\n", + "print('performance 1.6.5 is: ',performance1_6_5)" ] }, { @@ -490,7 +793,7 @@ }, { "cell_type": "code", - "execution_count": 56, + "execution_count": 48, "metadata": {}, "outputs": [ { @@ -499,7 +802,7 @@ "text": [ "3.6.3\n", "checking...\n", - "[('Hello', 'CS'), ('world', 'NN'), (',', ','), ('lets', 'NNS'), ('do', 'DO'), ('something', 'PN'), ('awesome', 'NN'), ('today', 'NR'), ('!', 'CD')]\n" + "[('Hello', 'VBD-HL'), ('world', 'VBD'), (',', ','), ('lets', 'NNS'), ('do', 'DO'), ('something', 'PN'), ('awesome', 'NN'), ('today', 'NR'), ('!', 'CD')]\n" ] } ], @@ -532,7 +835,7 @@ }, { "cell_type": "code", - "execution_count": 57, + "execution_count": 49, "metadata": { "scrolled": true }, @@ -553,7 +856,7 @@ "" ] }, - "execution_count": 57, + "execution_count": 49, "metadata": {}, "output_type": "execute_result" } @@ -566,8 +869,8 @@ "import plotly.graph_objs as go\n", "\n", "data = [go.Bar(\n", - " x=['performance 1.1', 'performance 1.2', 'performance 1.3', 'performance 1.4', 'performance 1.5' , 'performance 1.6'],\n", - " y=[performance1_1, performance1_2, performance1_3, performance1_4, performance1_5, performance1_6]\n", + " x=['performance 1.1', 'performance 1.2', 'performance 1.3.1', 'performance 1.3.2', 'performance 1.3.3', 'performance 1.3.4', 'performance 1.3.5', 'performance 1.4', 'performance 1.5' , 'performance 1.6.1', 'performance 1.6.2', 'performance 1.6.3', 'performance 1.6.4', 'performance 1.6.5'],\n", + " y=[performance1_1, performance1_2, performance1_3_1, performance1_3_2, performance1_3_3, performance1_3_4, performance1_3_5, performance1_4, performance1_5, performance1_6_1, performance1_6_2, performance1_6_3, performance1_6_4, performance1_6_5]\n", " )]\n", "\n", "py.iplot(data, filename='basic-bar')" @@ -582,12 +885,55 @@ "### Results for Task 2\n", "* train your model with standard features (plot a graph with all classifiers x results)\n", " * performance 2.1 = model4 in X3\n", - " * performance 2.2 = model5 in X3" + " * model 4 your POS tagger model (not english)\n", + " * performance 2.2 = model5 in X3\n", + " * pre-trained POS tagger model using RDRPOSTagger 1 or TreeTagger 2 (not english)" ] }, { "cell_type": "code", - "execution_count": 60, + "execution_count": 70, + "metadata": { + "collapsed": true + }, + "outputs": [], + "source": [ + "clf = Pipeline([\n", + " ('vectorizer', DictVectorizer(sparse=False)),\n", + " ('classifier', DecisionTreeClassifier(criterion='entropy'))\n", + "])" + ] + }, + { + "cell_type": "code", + "execution_count": 76, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "training OK\n", + "Accuracy: 0.836976962858\n" + ] + } + ], + "source": [ + "size=10000\n", + "clf.fit(X3[:size], y3[:size])\n", + " \n", + "print('training OK')\n", + " \n", + "X3_test, y3_test = transform_to_dataset(test_sentences_X3)\n", + "\n", + "performance2_1 = clf.score(X3_test, y3_test)\n", + "\n", + "print(\"Accuracy:\", performance2_1)" + ] + }, + { + "cell_type": "code", + "execution_count": 77, "metadata": {}, "outputs": [ { @@ -606,13 +952,12 @@ "" ] }, - "execution_count": 60, + "execution_count": 77, "metadata": {}, "output_type": "execute_result" } ], "source": [ - "performance2_1 = 0\n", "performance2_2 = 0\n", "\n", "data = [go.Bar(\n",