class AINewsCorpus: """ A corpus is a set of news articles (each with a title, content, and categories) that are used for training and comparison purposes. For training, the corpus provides the training examples. For comparison, the corpus provides the data for various measures like word frequency. This is important in the prediction process: we only want to predict a new article's categories based on word frequencies, and other measures, from the corpus; we don't want articles that have not been "vetted" (articles not part of the corpus) to contribute to these measures. A corpus can be "loaded" via C{load_corpus()} or "restored" via C{restore_corpus()}. The difference is the following: when loading a corpus, word frequencies are measured and stored in the database table C{wordlist_eval}; when restoring a corpus, word frequencies are simply retrieved from the database table C{wordlist}. In other words, we load a corpus when we are training or evaluating our training procedures, and we restore a corpus when we are predicting. """ def __init__(self): self.txtpro = AINewsTextProcessor() self.cache_urls = {} #: A dictionary of word=>word freq in corpus self.dftext = {} #: A dictionary of word=>wordid self.idwords = {} #: A dictionary of wordid=>word self.wordids = {} self.db = AINewsDB() self.categories = ["AIOverview","Agents", "Applications", \ "CognitiveScience", "Education", "Ethics", "Games", "History", \ "Interfaces", "MachineLearning", "NaturalLanguage", "Philosophy", \ "Reasoning", "Representation", "Robots", "ScienceFiction", \ "Speech", "Systems", "Vision"] self.retained_db_docs = None self.restore_corpus() def compare_articles(self, article1, article2): dupcount1 = len(article1['duplicates']) dupcount2 = len(article2['duplicates']) relevance1 = article1['source_relevance'] relevance2 = article2['source_relevance'] cat_count1 = len(article1['categories']) cat_count2 = len(article2['categories']) if cmp(dupcount1, dupcount2) == 0: if cmp(relevance1, relevance2) == 0: return cmp(cat_count1, cat_count2) else: return cmp(relevance1, relevance2) else: return cmp(dupcount1, dupcount2) def get_tfidf(self, urlid, wordfreq): """ Helper function to retrieve the tfidf of each word based on the urlid. @param urlid: target news story's urlid. @type urlid: C{int} """ if urlid in self.cache_urls: return self.cache_urls[urlid] wordid_freq_pairs = {} for word in wordfreq: if word in self.dftext: wordid_freq_pairs[self.idwords[word]] = (wordfreq[word], self.dftext[word]) data = {} distsq = 0.0 for wordid in wordid_freq_pairs: tfidf = math.log(wordid_freq_pairs[wordid][0] + 1, 2) * \ (math.log(self.corpus_count + 1, 2) - \ math.log(wordid_freq_pairs[wordid][1] + 1, 2)) data[wordid] = tfidf distsq += tfidf * tfidf dist = math.sqrt(distsq) if dist > 1.0e-9: for key in data: data[key] /= dist self.cache_urls[urlid] = data return data def cos_sim(self, tfidf1, tfidf2): """ A helper function to compute the cos simliarity between news story and centroid. @param tfidf1: target news story tfidf vector. @type tfidf1: C{dict} @param tfidf2: centroid tfidf vector. @type tfidf2: C{dict} """ sim = 0.0 for key in tfidf1: if key in tfidf2: word = self.wordids[key] a = tfidf1[key] b = tfidf2[key] sim += a * b return sim def get_article(self, urlid, corpus=False): row = None if corpus: table = 'cat_corpus' cat_table = 'cat_corpus_cats' row = self.db.selectone("""select u.url, u.title, u.content from %s as u where u.urlid = %s""" % (table, urlid)) else: table = 'urllist' cat_table = 'categories' row = self.db.selectone("""select u.url, u.title, u.content, u.summary, u.pubdate, u.crawldate, u.processed, u.published, u.source, u.source_relevance, u.source_id, u.tfpn, u.image_url from %s as u where u.urlid = %s""" % \ (table, urlid)) if row != None and row[2] is not None: content = row[2] wordfreq = self.txtpro.simpletextprocess(urlid, content) summary = "" if not corpus: summary = row[3] processed = False if not corpus and row[6] == 1: processed = True published = False if not corpus and row[7] == 1: published = True pubdate = "" if not corpus: pubdate = row[4] crawldate = "" if not corpus: crawldate = row[5] source = "" if not corpus: source = row[8] tfpn = "xx" if not corpus: tfpn = row[11] source_relevance = 0 if row[9]: source_relevance = int(row[9]) categories = [] cat_rows = self.db.selectall("""select category from %s where urlid = %s""" % (cat_table, urlid)) for cat_row in cat_rows: categories.append(cat_row[0]) return { 'urlid': urlid, 'url': row[0], 'title': row[1], 'content': content, 'summary': summary, 'pubdate': pubdate, 'crawldate': crawldate, 'processed': processed, 'published': published, 'source': source, 'source_relevance': source_relevance, 'source_id': row[10], 'categories': categories, 'duplicates': [], 'tfpn': tfpn, 'wordfreq': wordfreq, 'image_url': row[12], 'tfidf': self.get_tfidf(urlid, wordfreq) } else: return None def get_articles_daterange(self, date_start, date_end): articles = {} rows = self.db.selectall( """select urlid from urllist where pubdate >= %s and pubdate <= %s""", (date_start, date_end)) for row in rows: articles[row[0]] = self.get_article(row[0]) return articles def get_articles_idrange(self, urlid_start, urlid_end, corpus=False): articles = {} rows = self.db.selectall( """select urlid from urllist where urlid >= %s and urlid <= %s""", (urlid_start, urlid_end)) for row in rows: art = self.get_article(row[0], corpus) if art is not None: articles[row[0]] = art return articles def get_unprocessed(self): articles = {} rows = self.db.selectall( "select urlid from urllist where processed = 0") for row in rows: articles[row[0]] = self.get_article(row[0]) return articles def get_publishable(self): articles = [] rows = self.db.selectall( "select urlid from urllist where " "publishable = 1 and published = 0 and pubdate != '0000-00-00'") for row in rows: articles.append(self.get_article(row[0])) return articles def get_published(self): articles = [] rows = self.db.selectall( "select urlid from urllist where published = 1") for row in rows: articles.append(self.get_article(row[0])) return articles def mark_processed(self, articles): for article in articles: self.db.execute( "update urllist set processed = 1 where urlid = %s", article['urlid']) def mark_publishable(self, articles): for article in articles: self.db.execute( "update urllist set publishable = 1 where urlid = %s", article['urlid']) def mark_published(self, articles): for article in articles: self.db.execute( "update urllist set published = 1 where urlid = %s", article['urlid']) def restore_corpus(self): self.wordids = {} self.dftext = {} rows = self.db.selectall("select rowid, word, dftext from wordlist") for row in rows: self.wordids[row[0]] = row[1] self.idwords[row[1]] = row[0] self.dftext[row[1]] = row[2] self.corpus_count = self.db.selectone( "select count(*) from cat_corpus")[0] def add_freq_index(self, urlid, wordfreq, categories=[]): for word in wordfreq: self.wordcounts.setdefault(word, 0) self.wordcounts[word] += 1 def commit_freq_index(self, table): self.dftext = {} self.wordids = {} for word in self.wordcounts: rowid = self.db.execute("insert into "+table+" (word, dftext) " + \ "values(%s, %s)", (word, self.wordcounts[word])) self.wordids[rowid] = word self.idwords[word] = rowid self.dftext[word] = self.wordcounts[word] self.wordcounts = {} def load_corpus(self, ident, pct, debug=False, retain=False): if debug: print "Loading corpus..." source = ident.split(':')[0] name = ident.split(':')[1:] if source == "file": docs = self.load_file_corpus(name, debug) elif source == "db": docs = self.load_db_corpus(name, debug, retain) if debug: print random.shuffle(docs) offset = int(len(docs) * pct) if debug: print "Selecting random %d%% of corpus (%d docs)." % \ (pct * 100, offset) # sort train_corpus by urlid train_corpus = sorted(docs[0:offset], key=operator.itemgetter(0)) self.corpus_count = len(train_corpus) # sort predict_corpus by urlid predict_corpus = sorted(docs[offset:offset+int(len(docs)*0.1)], \ key=operator.itemgetter(0)) self.db.execute("delete from wordlist_eval") self.db.execute("alter table wordlist_eval auto_increment = 0") self.wordids = {} self.wordcounts = {} self.cache_urls = {} for c in train_corpus: self.add_freq_index(c[0], c[1], c[2].split()) if debug: sys.stdout.write('.') sys.stdout.flush() self.commit_freq_index('wordlist_eval') return (train_corpus, predict_corpus) def load_file_corpus(self, name, debug=False): wordsfile = paths['corpus.corpus_other'] + name[0] + ".mat.clabel" f = open(wordsfile, 'r') self.wordids = {} wordid = 1 for line in f: self.wordids[int(wordid)] = line.strip() wordid += 1 catsfile = paths['corpus.corpus_other'] + name[0] + ".mat.rlabel" f = open(catsfile, 'r') cats = {} uniqcats = set() docid = 0 for line in f: cats[docid] = line.strip() uniqcats.add(line.strip()) docid += 1 self.categories = list(uniqcats) matfile = paths['corpus.corpus_other'] + name[0] + ".mat" f = open(matfile, 'r') f.readline() # ignore first line docs = [] docid = 0 for line in f: wordfreq = {} for (wordid, freq) in izip(*[iter(line.split())] * 2): wordfreq[self.wordids[int(wordid)]] = int(float(freq)) docs.append((docid, wordfreq, cats[docid])) docid += 1 if debug: sys.stdout.write('.') sys.stdout.flush() return docs def load_db_corpus(self, name, debug=False, retain=False): rows = self.db.selectall("""select c.urlid, c.content, group_concat(cc.category separator ' ') from %s as c, %s as cc where c.urlid = cc.urlid group by c.urlid order by c.urlid desc""" % (name[0], name[1])) if debug: print "Processing %d articles..." % len(rows) if retain and self.retained_db_docs != None: return self.retained_db_docs docs = [] for row in rows: wordfreq = self.txtpro.simpletextprocess(row[0], row[1]) if wordfreq.N() > 0 and 'NotRelated' not in row[2].split(' '): docs.append((row[0], wordfreq, row[2])) if debug: sys.stdout.write('.') sys.stdout.flush() if retain: self.retained_db_docs = docs return docs
class AINewsWekaClassifier: def __init__(self): self.txtpro = AINewsTextProcessor() def __save_bag_of_words(self, tid, fieldidx): # find all unique words in the arff 'title' field, remove stop # words, perform stemming, collect their frequencies phrases = [] f = arff.load(open("%s%d.arff" % (paths['weka.training_arff_dir'], tid), 'r')) for record in f['data']: phrases.append(record[fieldidx]) bag = self.txtpro.simpletextprocess(0, ' '.join(phrases)) smallerbag = FreqDist() i = 0 for word in bag: if i == 1000: break smallerbag[word] = bag[word] i += 1 p = open("%sbag_of_words-%d.pickle" % (paths['weka.bag_of_words_dir'], fieldidx), 'w') pickle.dump(smallerbag, p) p.close() def __prepare_arff(self, tid): p = open("%sbag_of_words-0.pickle" % paths['weka.bag_of_words_dir'], 'r') bag_title = pickle.load(p) p.close() p = open("%sbag_of_words-1.pickle" % paths['weka.bag_of_words_dir'], 'r') bag_body = pickle.load(p) p.close() data = {'attributes': [], 'data': [], 'description': u'', 'relation': tid} for word in bag_title: data['attributes'].append(("title-%s" % word, 'NUMERIC')) for word in bag_body: data['attributes'].append(("body-%s" % word, 'NUMERIC')) data['attributes'].append(('class', ['yes', 'no'])) f = arff.load(open("%s%d.arff" % (paths['weka.training_arff_dir'], tid), 'r')) for record in f['data']: record_bag_title = self.txtpro.simpletextprocess(0, record[0]) record_bag_body = self.txtpro.simpletextprocess(0, record[1]) record_data = [] # iterate through original bag, figure out freq in this record's bag for word in bag_title: if word in record_bag_title: record_data.append(record_bag_title[word]) else: record_data.append(0) for word in bag_body: if word in record_bag_body: record_data.append(record_bag_body[word]) else: record_data.append(0) record_data.append(record[2]) data['data'].append(record_data) fnew = open("%s%d-wordvec-nonsparse.arff" % \ (paths['weka.training_arff_dir'], tid), 'w') arff.dump(fnew, data) fnew.close() # convert to sparse format Popen(("java -cp %s weka.filters.unsupervised.instance.NonSparseToSparse " + "-i %s%d-wordvec-nonsparse.arff -o %s%d-wordvec.arff") % \ (paths['weka.weka_jar'], paths['weka.training_arff_dir'], tid, paths['weka.training_arff_dir'], tid), shell = True).communicate() remove("%s%d-wordvec-nonsparse.arff" % (paths['weka.training_arff_dir'], tid)) # 1. load unprocessed arff files, from just one tid, from family_resemblance export # 2. gather all titles, parse into a bag of words # 3. save bag of words (list? need to keep the order) in a pickle file # 4. write new sparse arff files for each tid using this sorted bag of words def __get_tids(self): tids = [] files = listdir(paths['weka.training_arff_dir']) for f in files: m = re.match(r'^(\d+).arff$', f) if m: if m.group(1) == '0': continue tids.append(int(m.group(1))) return tids def train(self): tids = self.__get_tids() # all tid arffs have same entries, so use the first to grab the bag of words print "Saving bag of words..." self.__save_bag_of_words(tids[0], 0) self.__save_bag_of_words(tids[0], 1) for tid in sorted(tids): print "Preparing tid %d" % tid self.__prepare_arff(tid) for tid in sorted(tids): print "Spread subsampling for tid %d" % tid Popen(("java -cp %s weka.filters.supervised.instance.SpreadSubsample " + "-M 1.0 -X 0.0 -S 1 -c last " + "-i %s%d-wordvec.arff -o %s%d-wordvec-subsample.arff") % \ (paths['weka.weka_jar'], paths['weka.training_arff_dir'], tid, paths['weka.training_arff_dir'], tid), shell = True).communicate() print "Training random forests for tid %d" % tid Popen(("java -cp %s %s %s -v " + "-t %s%d-wordvec-subsample.arff -d %s%d.model") % \ (paths['weka.weka_jar'], config['weka.classifier'], config['weka.classifier_params'], paths['weka.training_arff_dir'], tid, paths['weka.training_arff_dir'], tid), shell = True, stdout = PIPE).communicate() print out def train_experiment(self): model_scores = {} models = {'random-forest': ('weka.classifiers.trees.RandomForest', '-I 20 -K 0'), 'naive-bayes': ('weka.classifiers.bayes.NaiveBayes', ''), 'bayesnet': ('weka.classifiers.bayes.BayesNet', ''), 'j48': ('weka.classifiers.trees.J48', ''), 'knn': ('weka.classifiers.lazy.IBk', '-K 3')} tids = self.__get_tids() # all tid arffs have same entries, so use the first to grab the bag of words print "Saving bag of words..." self.__save_bag_of_words(tids[0], 0) self.__save_bag_of_words(tids[0], 1) for tid in sorted(tids): print "Preparing tid %d" % tid self.__prepare_arff(tid) for tid in sorted(tids): print "Spread subsampling for tid %d" % tid Popen(("java -cp %s weka.filters.supervised.instance.SpreadSubsample " + "-M 1.0 -X 0.0 -S 1 -c last " + "-i %s%d-wordvec.arff -o %s%d-wordvec-subsample.arff") % \ (paths['weka.weka_jar'], paths['weka.training_arff_dir'], tid, paths['weka.training_arff_dir'], tid), shell = True).communicate() for tid in sorted(tids): model_scores[tid] = {} for model in models.keys(): print "Training %s for tid %d" % (models[model][0], tid) (out, _) = Popen(("java -cp %s %s %s -v " + "-t %s%d-wordvec-subsample.arff -d %s%d.model") % \ (paths['weka.weka_jar'], models[model][0], models[model][1], paths['weka.training_arff_dir'], tid, paths['weka.training_arff_dir'], tid), shell = True, stdout = PIPE).communicate() correct = 0.0 for line in out.splitlines(): m = re.search(r'Correctly Classified Instances\s+\d+\s+(.*) %', line) if m: correct = float(m.group(1)) break model_scores[tid][model] = correct with open('training_experiment.csv', 'w') as csvfile: writer = csv.writer(csvfile) writer.writerow(['model', 'tid', 'correct']) for tid in model_scores.keys(): for model in model_scores[tid].keys(): writer.writerow([model, tid, model_scores[tid][model]]) def __predict_arff(self): tids = self.__get_tids() # the testing file should always be 0.arff self.__prepare_arff(0) predictions = {} for tid in sorted(tids): predictions[tid] = [] print "Predicting tid %d" % tid (out, err) = Popen(("java -cp %s %s " + "-T %s0-wordvec.arff -l %s%d.model -p last") % \ (paths['weka.weka_jar'], config['weka.classifier'], paths['weka.training_arff_dir'], paths['weka.training_arff_dir'], tid), shell = True, stdout = PIPE).communicate() for line in out.splitlines(): m = re.search(r'2:no\s+[12]:(no|yes)\s+\+?\s+(\d+\.?\d*)', line) if m: answer = False if m.group(1) == 'yes': answer = True conf = float(m.group(2)) if conf < 0.75: answer = False predictions[tid].append((answer, conf)) return predictions def predict(self, articles): # modifies the provided articles dict data = {'attributes': [('title', 'STRING'), ('body', 'STRING'), ('class', ['yes', 'no'])], 'data': [], 'description': u'', 'relation': '0'} for urlid in sorted(articles.keys()): title = re.sub(r'\W', ' ', articles[urlid]['title']) body = re.sub(r'\W', ' ', articles[urlid]['summary']) data['data'].append([title, body, 'no']) # make the testing file 0.arff fnew = open("%s0.arff" % paths['weka.training_arff_dir'], 'w') arff.dump(fnew, data) fnew.close() predictions = self.__predict_arff() for urlid in sorted(articles.keys()): articles[urlid]['categories'] = [] tids = self.__get_tids() for tid in sorted(tids): for (i, urlid) in enumerate(sorted(articles.keys())): if predictions[tid][i][0]: articles[urlid]['categories'].append(str(tid))
class AINewsCorpus: """ A corpus is a set of news articles (each with a title, content, and categories) that are used for training and comparison purposes. For training, the corpus provides the training examples. For comparison, the corpus provides the data for various measures like word frequency. This is important in the prediction process: we only want to predict a new article's categories based on word frequencies, and other measures, from the corpus; we don't want articles that have not been "vetted" (articles not part of the corpus) to contribute to these measures. A corpus can be "loaded" via C{load_corpus()} or "restored" via C{restore_corpus()}. The difference is the following: when loading a corpus, word frequencies are measured and stored in the database table C{wordlist_eval}; when restoring a corpus, word frequencies are simply retrieved from the database table C{wordlist}. In other words, we load a corpus when we are training or evaluating our training procedures, and we restore a corpus when we are predicting. """ def __init__(self): self.txtpro = AINewsTextProcessor() self.cache_urls = {} #: A dictionary of word=>word freq in corpus self.dftext = {} #: A dictionary of word=>wordid self.idwords = {} #: A dictionary of wordid=>word self.wordids = {} self.db = AINewsDB() self.categories = ["AIOverview","Agents", "Applications", \ "CognitiveScience", "Education", "Ethics", "Games", "History", \ "Interfaces", "MachineLearning", "NaturalLanguage", "Philosophy", \ "Reasoning", "Representation", "Robots", "ScienceFiction", \ "Speech", "Systems", "Vision"] self.sources = {} rows = self.db.selectall("select parser, relevance from sources") for row in rows: self.sources[row[0].split('::')[0]] = int(row[1]) self.retained_db_docs = None self.restore_corpus() def get_relevance(self, publisher): if re.search(r'via Google News', publisher): publisher = 'GoogleNews' return self.sources[publisher] def compare_articles(self, article1, article2): dupcount1 = len(article1['duplicates']) dupcount2 = len(article2['duplicates']) if article1['publisher'].find('User submitted') != -1: relevance1 = 200 else: relevance1 = self.get_relevance(article1['publisher']) if article2['publisher'].find('User submitted') != -1: relevance2 = 200 else: relevance2 = self.get_relevance(article2['publisher']) cat_count1 = len(article1['categories']) cat_count2 = len(article2['categories']) if cmp(dupcount1, dupcount2) == 0: if cmp(relevance1, relevance2) == 0: return cmp(cat_count1, cat_count2) else: return cmp(relevance1, relevance2) else: return cmp(dupcount1, dupcount2) def get_tfidf(self, urlid, wordfreq): """ Helper function to retrieve the tfidf of each word based on the urlid. @param urlid: target news story's urlid. @type urlid: C{int} """ if urlid in self.cache_urls: return self.cache_urls[urlid] wordid_freq_pairs = {} for word in wordfreq: if word in self.dftext: wordid_freq_pairs[self.idwords[word]] = (wordfreq[word], self.dftext[word]) data = {} distsq = 0.0 for wordid in wordid_freq_pairs: tfidf = math.log(wordid_freq_pairs[wordid][0] + 1, 2) * \ (math.log(self.corpus_count + 1, 2) - \ math.log(wordid_freq_pairs[wordid][1] + 1, 2)) data[wordid] = tfidf distsq += tfidf * tfidf dist = math.sqrt(distsq) if dist > 1.0e-9: for key in data: data[key] /= dist self.cache_urls[urlid] = data return data def cos_sim(self, tfidf1, tfidf2): """ A helper function to compute the cos simliarity between news story and centroid. @param tfidf1: target news story tfidf vector. @type tfidf1: C{dict} @param tfidf2: centroid tfidf vector. @type tfidf2: C{dict} """ sim = 0.0 for key in tfidf1: if key in tfidf2: word = self.wordids[key] a = tfidf1[key] b = tfidf2[key] sim += a*b return sim def get_article(self, urlid, corpus = False): row = None if corpus: table = 'cat_corpus' cat_table = 'cat_corpus_cats' row = self.db.selectone("""select u.url, u.title, u.content from %s as u where u.urlid = %s""" % (table, urlid)) else: table = 'urllist' cat_table = 'categories' row = self.db.selectone("""select u.url, u.title, u.content, u.summary, u.pubdate, u.crawldate, u.processed, u.published, u.publisher from %s as u where u.urlid = %s""" % \ (table, urlid)) if row != None and row[2] is not None: wordfreq = self.txtpro.simpletextprocess(urlid, row[2]) summary = "" if not corpus: summary = row[3] processed = False if not corpus and row[6] == 1: processed = True published = False if not corpus and row[7] == 1: published = True pubdate = "" if not corpus: pubdate = row[4] crawldate = "" if not corpus: crawldate = row[5] publisher = "" if not corpus: publisher = row[8] categories = [] cat_rows = self.db.selectall("""select category from %s where urlid = %s""" % (cat_table, urlid)) for cat_row in cat_rows: categories.append(cat_row[0]) return {'urlid': urlid, 'url': row[0], 'title': row[1], 'content': trunc(row[2], max_pos=3000), 'content_all': row[2], 'summary': summary, 'pubdate': pubdate, 'crawldate': crawldate, 'processed': processed, 'published': published, 'publisher': publisher, 'categories': categories, 'duplicates': [], 'wordfreq': wordfreq, 'tfidf': self.get_tfidf(urlid, wordfreq)} else: return None def get_articles_daterange(self, date_start, date_end): articles = {} rows = self.db.selectall("""select urlid from urllist where pubdate >= %s and pubdate <= %s""", (date_start, date_end)) for row in rows: articles[row[0]] = self.get_article(row[0]) return articles def get_articles_idrange(self, urlid_start, urlid_end, corpus = False): articles = {} rows = self.db.selectall("""select urlid from urllist where urlid >= %s and urlid <= %s""", (urlid_start, urlid_end)) for row in rows: art = self.get_article(row[0], corpus) if art is not None: articles[row[0]] = art return articles def get_unprocessed(self): articles = {} rows = self.db.selectall("select urlid from urllist where processed = 0") for row in rows: articles[row[0]] = self.get_article(row[0]) return articles def get_publishable(self): articles = [] rows = self.db.selectall("select urlid from urllist where " "publishable = 1 and published = 0 and pubdate != '0000-00-00'") for row in rows: articles.append(self.get_article(row[0])) return articles def get_published(self): articles = [] rows = self.db.selectall("select urlid from urllist where published = 1") for row in rows: articles.append(self.get_article(row[0])) return articles def mark_processed(self, articles): for article in articles: self.db.execute("update urllist set processed = 1 where urlid = %s", article['urlid']) def mark_publishable(self, articles): for article in articles: self.db.execute("update urllist set publishable = 1 where urlid = %s", article['urlid']) def mark_published(self, articles): for article in articles: self.db.execute("update urllist set published = 1 where urlid = %s", article['urlid']) def restore_corpus(self): self.wordids = {} self.dftext = {} rows = self.db.selectall("select rowid, word, dftext from wordlist") for row in rows: self.wordids[row[0]] = row[1] self.idwords[row[1]] = row[0] self.dftext[row[1]] = row[2] self.corpus_count = self.db.selectone("select count(*) from cat_corpus")[0] def add_freq_index(self, urlid, wordfreq, categories = []): for word in wordfreq: self.wordcounts.setdefault(word, 0) self.wordcounts[word] += 1 def commit_freq_index(self, table): self.dftext = {} self.wordids = {} for word in self.wordcounts: rowid = self.db.execute("insert into "+table+" (word, dftext) " + \ "values(%s, %s)", (word, self.wordcounts[word])) self.wordids[rowid] = word self.idwords[word] = rowid self.dftext[word] = self.wordcounts[word] self.wordcounts = {} def load_corpus(self, ident, pct, debug = False, retain = False): if debug: print "Loading corpus..." source = ident.split(':')[0] name = ident.split(':')[1:] if source == "file": docs = self.load_file_corpus(name, debug) elif source == "db": docs = self.load_db_corpus(name, debug, retain) if debug: print random.shuffle(docs) offset = int(len(docs)*pct) if debug: print "Selecting random %d%% of corpus (%d docs)." % \ (pct * 100, offset) # sort train_corpus by urlid train_corpus = sorted(docs[0:offset], key=operator.itemgetter(0)) self.corpus_count = len(train_corpus) # sort predict_corpus by urlid predict_corpus = sorted(docs[offset:offset+int(len(docs)*0.1)], \ key=operator.itemgetter(0)) self.db.execute("delete from wordlist_eval") self.db.execute("alter table wordlist_eval auto_increment = 0") self.wordids = {} self.wordcounts = {} self.cache_urls = {} for c in train_corpus: self.add_freq_index(c[0], c[1], c[2].split()) if debug: sys.stdout.write('.') sys.stdout.flush() self.commit_freq_index('wordlist_eval') return (train_corpus, predict_corpus) def load_file_corpus(self, name, debug = False): wordsfile = paths['corpus.corpus_other'] + name[0] + ".mat.clabel" f = open(wordsfile, 'r') self.wordids = {} wordid = 1 for line in f: self.wordids[int(wordid)] = line.strip() wordid += 1 catsfile = paths['corpus.corpus_other'] + name[0] + ".mat.rlabel" f = open(catsfile, 'r') cats = {} uniqcats = set() docid = 0 for line in f: cats[docid] = line.strip() uniqcats.add(line.strip()) docid += 1 self.categories = list(uniqcats) matfile = paths['corpus.corpus_other'] + name[0] + ".mat" f = open(matfile, 'r') f.readline() # ignore first line docs = [] docid = 0 for line in f: wordfreq = {} for (wordid, freq) in izip(*[iter(line.split())]*2): wordfreq[self.wordids[int(wordid)]] = int(float(freq)) docs.append((docid, wordfreq, cats[docid])) docid += 1 if debug: sys.stdout.write('.') sys.stdout.flush() return docs def load_db_corpus(self, name, debug = False, retain = False): rows = self.db.selectall("""select c.urlid, c.content, group_concat(cc.category separator ' ') from %s as c, %s as cc where c.urlid = cc.urlid group by c.urlid order by c.urlid desc""" % (name[0], name[1])) if debug: print "Processing %d articles..." % len(rows) if retain and self.retained_db_docs != None: return self.retained_db_docs docs = [] for row in rows: wordfreq = self.txtpro.simpletextprocess(row[0], row[1]) if wordfreq.N() > 0 and 'NotRelated' not in row[2].split(' '): docs.append((row[0], wordfreq, row[2])) if debug: sys.stdout.write('.') sys.stdout.flush() if retain: self.retained_db_docs = docs return docs
class AINewsWekaClassifier: def __init__(self): self.txtpro = AINewsTextProcessor() def __save_bag_of_words(self, tid, fieldidx): # find all unique words in the arff 'title' field, remove stop # words, perform stemming, collect their frequencies phrases = [] f = arff.load( open("%s%d.arff" % (paths['weka.training_arff_dir'], tid), 'r')) for record in f['data']: phrases.append(record[fieldidx]) bag = self.txtpro.simpletextprocess(0, ' '.join(phrases)) smallerbag = FreqDist() i = 0 for word in bag: if i == 1000: break smallerbag[word] = bag[word] i += 1 p = open( "%sbag_of_words-%d.pickle" % (paths['weka.bag_of_words_dir'], fieldidx), 'w') pickle.dump(smallerbag, p) p.close() def __prepare_arff(self, tid): p = open("%sbag_of_words-0.pickle" % paths['weka.bag_of_words_dir'], 'r') bag_title = pickle.load(p) p.close() p = open("%sbag_of_words-1.pickle" % paths['weka.bag_of_words_dir'], 'r') bag_body = pickle.load(p) p.close() data = { 'attributes': [], 'data': [], 'description': u'', 'relation': tid } for word in bag_title: data['attributes'].append(("title-%s" % word, 'NUMERIC')) for word in bag_body: data['attributes'].append(("body-%s" % word, 'NUMERIC')) data['attributes'].append(('class', ['yes', 'no'])) f = arff.load( open("%s%d.arff" % (paths['weka.training_arff_dir'], tid), 'r')) for record in f['data']: record_bag_title = self.txtpro.simpletextprocess(0, record[0]) record_bag_body = self.txtpro.simpletextprocess(0, record[1]) record_data = [] # iterate through original bag, figure out freq in this record's bag for word in bag_title: if word in record_bag_title: record_data.append(record_bag_title[word]) else: record_data.append(0) for word in bag_body: if word in record_bag_body: record_data.append(record_bag_body[word]) else: record_data.append(0) record_data.append(record[2]) data['data'].append(record_data) fnew = open("%s%d-wordvec-nonsparse.arff" % \ (paths['weka.training_arff_dir'], tid), 'w') arff.dump(fnew, data) fnew.close() # convert to sparse format Popen(("java -cp %s weka.filters.unsupervised.instance.NonSparseToSparse " + "-i %s%d-wordvec-nonsparse.arff -o %s%d-wordvec.arff") % \ (paths['weka.weka_jar'], paths['weka.training_arff_dir'], tid, paths['weka.training_arff_dir'], tid), shell = True).communicate() remove("%s%d-wordvec-nonsparse.arff" % (paths['weka.training_arff_dir'], tid)) # 1. load unprocessed arff files, from just one tid, from family_resemblance export # 2. gather all titles, parse into a bag of words # 3. save bag of words (list? need to keep the order) in a pickle file # 4. write new sparse arff files for each tid using this sorted bag of words def __get_tids(self): tids = [] files = listdir(paths['weka.training_arff_dir']) for f in files: m = re.match(r'^(\d+).arff$', f) if m: if m.group(1) == '0': continue tids.append(int(m.group(1))) return tids def train(self): tids = self.__get_tids() # all tid arffs have same entries, so use the first to grab the bag of words print "Saving bag of words..." self.__save_bag_of_words(tids[0], 0) self.__save_bag_of_words(tids[0], 1) for tid in sorted(tids): print "Preparing tid %d" % tid self.__prepare_arff(tid) for tid in sorted(tids): print "Spread subsampling for tid %d" % tid Popen(("java -cp %s weka.filters.supervised.instance.SpreadSubsample " + "-M 1.0 -X 0.0 -S 1 -c last " + "-i %s%d-wordvec.arff -o %s%d-wordvec-subsample.arff") % \ (paths['weka.weka_jar'], paths['weka.training_arff_dir'], tid, paths['weka.training_arff_dir'], tid), shell = True).communicate() print "Training random forests for tid %d" % tid Popen(("java -cp %s %s %s -v " + "-t %s%d-wordvec-subsample.arff -d %s%d.model") % \ (paths['weka.weka_jar'], config['weka.classifier'], config['weka.classifier_params'], paths['weka.training_arff_dir'], tid, paths['weka.training_arff_dir'], tid), shell = True, stdout = PIPE).communicate() print out def train_experiment(self): model_scores = {} models = { 'random-forest': ('weka.classifiers.trees.RandomForest', '-I 20 -K 0'), 'naive-bayes': ('weka.classifiers.bayes.NaiveBayes', ''), 'bayesnet': ('weka.classifiers.bayes.BayesNet', ''), 'j48': ('weka.classifiers.trees.J48', ''), 'knn': ('weka.classifiers.lazy.IBk', '-K 3') } tids = self.__get_tids() # all tid arffs have same entries, so use the first to grab the bag of words print "Saving bag of words..." self.__save_bag_of_words(tids[0], 0) self.__save_bag_of_words(tids[0], 1) for tid in sorted(tids): print "Preparing tid %d" % tid self.__prepare_arff(tid) for tid in sorted(tids): print "Spread subsampling for tid %d" % tid Popen(("java -cp %s weka.filters.supervised.instance.SpreadSubsample " + "-M 1.0 -X 0.0 -S 1 -c last " + "-i %s%d-wordvec.arff -o %s%d-wordvec-subsample.arff") % \ (paths['weka.weka_jar'], paths['weka.training_arff_dir'], tid, paths['weka.training_arff_dir'], tid), shell = True).communicate() for tid in sorted(tids): model_scores[tid] = {} for model in models.keys(): print "Training %s for tid %d" % (models[model][0], tid) (out, _) = Popen(("java -cp %s %s %s -v " + "-t %s%d-wordvec-subsample.arff -d %s%d.model") % \ (paths['weka.weka_jar'], models[model][0], models[model][1], paths['weka.training_arff_dir'], tid, paths['weka.training_arff_dir'], tid), shell = True, stdout = PIPE).communicate() correct = 0.0 for line in out.splitlines(): m = re.search( r'Correctly Classified Instances\s+\d+\s+(.*) %', line) if m: correct = float(m.group(1)) break model_scores[tid][model] = correct with open('training_experiment.csv', 'w') as csvfile: writer = csv.writer(csvfile) writer.writerow(['model', 'tid', 'correct']) for tid in model_scores.keys(): for model in model_scores[tid].keys(): writer.writerow([model, tid, model_scores[tid][model]]) def __predict_arff(self): tids = self.__get_tids() # the testing file should always be 0.arff self.__prepare_arff(0) predictions = {} for tid in sorted(tids): predictions[tid] = [] print "Predicting tid %d" % tid (out, err) = Popen(("java -cp %s %s " + "-T %s0-wordvec.arff -l %s%d.model -p last") % \ (paths['weka.weka_jar'], config['weka.classifier'], paths['weka.training_arff_dir'], paths['weka.training_arff_dir'], tid), shell = True, stdout = PIPE).communicate() for line in out.splitlines(): m = re.search(r'2:no\s+[12]:(no|yes)\s+\+?\s+(\d+\.?\d*)', line) if m: answer = False if m.group(1) == 'yes': answer = True conf = float(m.group(2)) if conf < 0.75: answer = False predictions[tid].append((answer, conf)) return predictions def predict(self, articles): # modifies the provided articles dict data = { 'attributes': [('title', 'STRING'), ('body', 'STRING'), ('class', ['yes', 'no'])], 'data': [], 'description': u'', 'relation': '0' } for urlid in sorted(articles.keys()): title = re.sub(r'\W', ' ', articles[urlid]['title']) body = re.sub(r'\W', ' ', articles[urlid]['summary']) data['data'].append([title, body, 'no']) # make the testing file 0.arff fnew = open("%s0.arff" % paths['weka.training_arff_dir'], 'w') arff.dump(fnew, data) fnew.close() predictions = self.__predict_arff() for urlid in sorted(articles.keys()): articles[urlid]['categories'] = [] tids = self.__get_tids() for tid in sorted(tids): for (i, urlid) in enumerate(sorted(articles.keys())): if predictions[tid][i][0]: articles[urlid]['categories'].append(str(tid))