使用TF-IDF和BM25提取文章关键词
评估方法:
人工从文章中提取1-5个关键词,和机器提取的关键词做比较
召回 = 机器提词∩人工提词 / 人工提词
准确 = 机器提词∩人工提词 / 机器提词
TF-IDF
原理参考:http://www.ruanyifeng.com/blog/2013/03/tf-idf.html
实现参考:tf-idf-keyword
其他参考: 使用不同的方法计算TF-IDF值
第一版 标题和正文加权计算tf-idf
主要策略
(1)使用nlpc切词服务(可用jieba切词代替)+TF-IDF提取关键词。
(2)去除停用词
(3)按照体裁+年级分成若干类型,来训练模型,示例用高中+叙事类,取了20000条数据训练
(4)对标题进行加权,标题的每个词汇频率+6,再合一起计算tf-idf
(5)按照权重取前4个关键词,在这4个关键词中对于权重小于 频率(5)*平均IDF/总词数 的进行过滤
注:以上数据均为调节后最优解
代码实现
config.py
program = 'composition_term_weight' logger = logging.getLogger(program) logging.basicConfig(format='%(asctime)s: %(levelname)s: %(message)s', stream=sys.stderr, datefmt='%a, %d %b %Y %H:%M:%S') logging.root.setLevel(level=logging.INFO)
IDFLoader.py
class IDFLoader(object): """词典加载类""" def __init__(self, idf_path): self.idf_path = idf_path self.idf_freq = {} # idf self.mean_len = 0 #平均长度 self.mean_idf = 0.0 # 均值 self.load_idf() def load_idf(self): """从文件中载入idf""" cnt = 0 with open(self.idf_path, 'rb') as f: for line in f: try: word, freq = line.strip().decode('utf-8', errors='ignore').split(' ') if word == 'LEN_AVG': self.mean_len = int(freq) break self.idf_freq[word] = float(freq) cnt += 1 except Exception as e: # logger.error('load_idf error: ' + e.message + ' line: ' + line.decode('utf-8', errors='ignore')) continue self.mean_idf = sum(self.idf_freq.values()) / cnt logger.info('Vocabularies %s loaded: %d mean_idf: %d' % (self.idf_path, cnt, self.mean_idf))
class TfIdf(object): """TF-IDF""" # 对正文进行过滤 p_cut = re.compile(r'[a-zA-Z0-9]', re.VERBOSE) # 对标题进行过滤 p_title = re.compile(r'作文|\d+字|.年级|_', re.VERBOSE) # 过滤常用标点符号等,也可以放到停用词表中 ignored = ['', ' ', '', '。', ':', ',', ')', '(', '!', '?', '”', '“', '"', '―', '.', '说', '好', '时'] # 主题最小出现次数,用于过滤权重不达标的关键词 min_times = 5.0 # 标题加权次数 title_add_times = 6.0 # 取关键词的个数 words_num = 4 def __init__(self): # 1. 获取停用词库 my_stop_words_path = 'stop_words.utf8.txt' self.stop_words_dict = [] with open(my_stop_words_path, 'rb') as fr: for line in fr.readlines(): self.stop_words_dict.append(line.strip()) def my_cut(self, inTxt): """切词""" inTxt = self.p_cut.sub('', str(inTxt)) words_list = [] # 由于性能问题,一句一句的切词 for l in inTxt.split('。'): # NLPC切词服务,可用jieba切词代替 r = cut(l) if r is not None: words_list += r return [w for w in words_list if w not in self.stop_words_dict and w not in self.ignored and len(w.strip()) > 0] def get_tfidf(self, idf_loader, title, content): """计算文章tf-idf""" filter_title = self.p_title.sub('', title.encode('utf-8', errors='ignore')) title_words = self.my_cut(filter_title) corpus0 = title_words + self.my_cut(content) freq = {} for w in corpus0: freq[w] = freq.get(w, 0.0) + 1.0 # 对标题进行加权 for w in title_words: logger.info(freq[w]) freq[w] = freq.get(w, 0.0) + self.title_add_times logger.info(freq[w]) total = sum(freq.values()) for k in freq: # 计算 TF-IDF freq[k] *= idf_loader.idf_freq.get(k, idf_loader.mean_idf) / total return sorted(freq.items(), key=lambda d: d[1], reverse=True), len(corpus0), title_words def get_term_weight(self, idf_loader, title, content): """获得term权重""" result, words_number, title_words = self.get_tfidf(idf_loader, title, content) bound = self.min_times * idf_loader.mean_idf / words_number machine_words = [item for item in result[:4] if item[1] > bound] # machine_words = [item for item in result[:self.words_num]] if len(machine_words) < 1: # 如果一个term都没有,则把标题拿出来 machine_words = [item for item in result if item[1] in title_words] data = [] offset = 0 for i, word in enumerate(machine_words): data.append('%s:%d:%s' % (word[0], offset, str(round(word[1], 4)))) offset += len(word[0].decode('utf-8', errors='ignore')) return data def getCorpus(self, data_path): """获取词表""" count = 0 corpus_list = [] with open(data_path, 'rb') as f: for line in f: info = json.loads(line.decode('utf-8', errors='ignore')) sentence = self.p_title.sub('', info.get('title').encode('utf-8', errors='ignore')) + '。' + info.get( '@merge_text').encode('utf-8', errors='ignore') r = self.my_cut(sentence) if not r: continue corpus_list.append(r) count += 1 if count % 1000 == 0: logger.info("processd " + str(count) + " segment_sentence") return corpus_list def train(self, dir_name, data_path): """训练模型""" idf_path = 'data/%s/idf.txt' % dir_name documents = self.getCorpus(data_path) id_freq = {} i = 0 len_sum = 0 for doc in documents: len_sum += len(doc) doc = set(doc) for x in doc: id_freq[x] = id_freq.get(x, 0) + 1 if i % 1000 == 0: logger.info('Documents processed: ' + str(i) + ', time: ' + str(datetime.datetime.now())) i += 1 del documents with open(idf_path, 'wb') as f: for key, value in id_freq.items(): f.write(key + ' ' + str(math.log(i / value, 2)) + '\n') logger.info(str(i) + ' ' + str(len_sum)) f.write('LEN_AVG ' + str(len_sum / i)) def test_one(self, dir_name, method='tfidf'): """单个测试""" idf_loader = IDFLoader('data/%s/idf.txt' % dir_name) for item in sys.stdin: info = json.loads(item.decode('utf-8', errors='ignore')) title = info['title'] content = info['@merge_text'] if method == 'tfidf': result, words_number, title_words = self.get_tfidf(idf_loader, title, content) else: result, words_number, title_words = self.get_bm25(idf_loader, title, content) bound = self.min_times * idf_loader.mean_idf / words_number print '_____words_number bound_____' print words_number, bound print '_____tfidf_result_____' for item in result[:20]: print item[0].encode('utf-8', errors='ignore'), item[1]
经调优,最优解为:min_times=5 title_add_times=6.0 words_num=4
结果
人工抽样评估了100个 TF-IDF召回率:0.2778 TF-IDF准确率:0.2778
BM25
算法参考: 搜索中的权重度量利器: TF-IDF和BM25
第一版
TfIdf.py 增加方法:
def get_bm25(self, idf_loader, title, content): """计算bm25""" k = 1.2 # 用来限制TF值的增长极限 b = 0.75 # b是一个常数,它的作用是规定L对评分的影响有多大。 # L是文档长度与平均长度的比值 EPSILON = 0.25 # 如果idf词表中没有,则平均idf*该值 filter_title = self.p_title.sub('', title.encode('utf-8', errors='ignore')) title_words = self.my_cut(filter_title) corpus0 = title_words + self.my_cut(content) freq = {} for w in corpus0: freq[w] = freq.get(w, 0.0) + 1.0 # 对标题进行加权 for w in title_words: freq[w] = freq.get(w, 0.0) + self.title_add_times total = sum(freq.values()) logger.info(str((k, b, total, idf_loader.mean_len))) for i in freq: tf = freq[i] / total idf = idf_loader.idf_freq.get(i, idf_loader.mean_idf * EPSILON) freq[i] = idf * ((k + 1) * tf) / (k * (1.0 - b + b * (total / idf_loader.mean_len)) + tf) return sorted(freq.items(), key=lambda d: d[1], reverse=True), len(corpus0), title_words
经调优,最优解为:min_times=2.5 title_add_times=6.0 words_num=4 k=1.2 b=0.75 EPSILON=0.25
结果
人工抽样评估了100个 BM25召回率:0.2889 BM25准确率:0.3333
神龙|纯净稳定代理IP免费测试>>>>>>>>天启|企业级代理IP免费测试>>>>>>>>IPIPGO|全球住宅代理IP免费测试