用机器学习打造聊天机器人(四) 代码篇

def build_feature(self, sentence, w_i_dict):
    """
    根据词汇表构造句子向量，其中用到的'w_i_dict'参数会通过以下方法先构造好：
    # 构建训练语料库
    build_corpus_vocabulary()
    # 训练语料库分词
    cut_corpus_vocabulary()
    # 构建训练语料库词汇反向索引
    word_index_dict_ = load_vocabulary()
    # 存储训练语料库词汇反向索引
    dump_word_index(word_index_dict_)
    :param sentence: 句子
    :param w_i_dict: 词汇-位置索引字典
    :return: one-hot 向量
    """
    # 分词
    sentence_seg = jieba.cut(sentence)
    # 用0初始化one-hot向量，维数为词汇表的词的个数
    sen_vec = np.zeros(len(w_i_dict))
    # 词汇表的词的列表
    w_i_dict_keys = w_i_dict.keys()
    # one-hot向量对应词在词典中的位置至1
    for word in sentence_seg:
        if w_i_dict_keys.__contains__(word):
            sen_vec[w_i_dict[word]] = 1
    return sen_vec

def sum_vecs_avg(self, text):
    """
    根据词向量模型构建句子向量
    :param text: 句子
    :return:
    """
    # 加载词向量模型
    word_vec_model = ModelsLoader().sf_words_vec_model
    # 用0值初始化一个同维数的向量，如果你知道你的词向量模型是多少维的，可以直接指定，不用采用下面的野路子
    vec = np.zeros(word_vec_model['是'].shape[0])
    # 分词
    words_list = list(jieba.cut(text))
    for w in words_list:
        try:
            # 将所有词的向量累加
            vec = vec + word_vec_model[w]
        except KeyError as e:
            logging.warning('词‘%s’，不在词向量模型词汇表中', w)
            continue
        except ValueError as e:
            logging.error('Error:', e)
            break
    # 计算平均向量
    vec = vec / len(words_list)
    return vec

def train_clf(self):
    """
    基于贝叶斯算法训练意图分类器，并存储为文件，以便下次使用
    :return: 
    """
    dump_path = "%s/classifier_mnb.m" % get_resources_trained_models()
    # 加载训练样本数据
    features_np, labels_np = load_train_data()
    features_np = np.array(features_np)
    labels_np = np.array(labels_np)
    # 开始训练
    starttime = datetime.datetime.now()
    print("开始训练分类器...")
    # 使用多项式朴素贝叶斯算法训练模型
    clf = MultinomialNB(alpha=0.1, fit_prior=True, class_prior=None)
    # 从第10个开始纳入训练，前10将做为验证集评估模型的表现
    clf.fit(features_np[10:], labels_np[10:])
    endtime = datetime.datetime.now()
    print("===========训练耗时: %s" % (endtime - starttime).seconds)
    # 评估分类器在验证集上的表现
    print("评估结果：%s" % clf.score(features_np[:10], labels_np[:10]))
    self.clf_nb = clf
    # 存储分类器
    dump_clf(self)
    print("分类器存储位置：%s" % dump_path)
    return self

def predict(self, feature_vec, clf):
    """
    预测(基于贝叶斯模型)
    :param feature_vec: 输入句子的特征向量
    :param clf: 训练好的贝叶斯模型
    :return:
    """
    proba_pred_np = clf.clf_nb.predict_proba(np.array([feature_vec]))[0]
    logging.debug("预测结果的概率：%s", proba_pred_np)
    # 加载类别集合
    labels_set = load_labels_set()
    label_score_list = []
    for i, num in enumerate(proba_pred_np):
        # if num != 0.00000000e+00:
        if num >= current_app.config['THRESHOLD_INTENT_RECOGNITION']:
            label_score_list.append((labels_set[i], num))
    if len(label_score_list) == 0:  # 正常阈值下没有匹配项，就降级匹配
        logging.debug("意图识别在正常分数阈值下没有匹配到任何项，进行降级匹配...")
        for i, num in enumerate(proba_pred_np):
            # if num != 0.00000000e+00:
            if num >= current_app.config['MINIMUM_THRESHOLD_INTENT_RECOGNITION']:
                label_score_list.append((labels_set[i], num))
    rs = sorted(label_score_list, key=lambda item: item[1], reverse=True)
    return rs, [c for c, v in rs]

def train_clf(self):
    """
    训练分类器(基于中心向量的方式)
    :return: 
    """
    data = DataLoader().load_train_data()
    logging.info("开始训练...")
    _, labels_centroids_dict = self.cal_centroid_vec(data)
    self.labels_centroids_dict = labels_centroids_dict
    self.labels = list(labels_centroids_dict.keys())
    logging.info("训练完成！")
    # 存储分类器模型
    self.dump(self)
    return self

def cal_centroid_vec(self, data):
    """
    构建“类别-中心向量”字典
    :param data: {'类别':{examples:'句子样本',centroid:'中心向量'}}
    :return: 
    """
    labels_centroids_dict = {}
    for the_label in data.keys():
        centroid = self.get_centroid(data[the_label]["examples"])
        data[the_label]["centroid"] = centroid
        labels_centroids_dict[the_label] = centroid
    return data, labels_centroids_dict

def get_centroid(self, examples):
    """
    获取当前意图类别的中心向量。中心向量由examples中所有句子向量各分量上的算数平均数表示
    :param examples: 当前类别下的所有样本句子
    :return:
    """
    word_vec_model = ModelsLoader().sf_words_vec_model
    word_dim = word_vec_model['是'].shape[0]
    C = np.zeros((len(examples), word_dim))
    for idx, text in enumerate(examples):
        C[idx, :] = self.sum_vecs_avg(text)
    centroid = np.mean(C, axis=0)
    assert centroid.shape[0] == word_dim
    return centroid
    
def predict(self, feature_vec, clf):
    """
    预测意图类别（基于向量夹角余弦值）
    :param feature_vec: 输入句子的特征向量
    :param clf: 从接口继承下来的参数，这里用不到
    :return: 
    """
    intents = self.labels
    # 分数计算规则：计算新句子的向量和当前意图类别的中心向量的夹角余弦值，下面其实可以改进以下，用矩阵并行计算代替for循环，但是因为类别目前不多，影响暂时不大。
    scores = [(label_, np.dot(feature_vec, self.labels_centroids_dict[label_]) / (
            np.linalg.norm(feature_vec) * np.linalg.norm(self.labels_centroids_dict[label_]))) for label_ in
              intents]
    rs = sorted(scores, key=lambda item: item[1], reverse=True)
    top1scores = rs[0][1]
    top1label = rs[0][0]
    logging.debug("top1的分数：%s,label:%s", top1scores, top1label)
    if top1scores >= current_app.config['THRESHOLD_INTENT_RECOGNITION']:
        rs = rs[:1]
    elif top1scores >= current_app.config['MINIMUM_THRESHOLD_INTENT_RECOGNITION']:
        logging.debug("意图识别在正常分数阈值下没有匹配到任何项，进行降级匹配...")
    elif top1scores < current_app.config['MINIMUM_THRESHOLD_INTENT_RECOGNITION']:
        logging.debug("意图识别在最小分数阈值下没有匹配到任何项...")
        rs = []
    return rs, [c for c, v in rs]

def compare(self, statement, statement_vec):
    """
    比较夹角余弦值
    :param statement: 输入句子对象
    :param statement_vec: 句子样本特征向量，是一个二维list
    :return: 输入句子和各句子样本的相似度构成的二维数组
    """
    statement_text_vec = statement.text_vector
    statement_vec = np.array(statement_vec)
    # 向量化并行计算余弦值
    similarity = np.dot(statement_text_vec, statement_vec.T) / (
                np.linalg.norm(statement_text_vec) * np.linalg.norm(statement_vec, axis=1)).T
    print("similarity.shape %s" % similarity.shape)
    return similarity

def train_sf_chatbot():
    data_root_dir = path_configer.get_classifier_train_samples()
    for file_name in os.listdir(data_root_dir):
        if file_name.startswith("QA_sf_"):
            __train(('%s/%s' % (get_chatter_corpus(), file_name)), file_name[:file_name.find('-')])
        
def __train(corpus_path, collection_name):
    print("开始训练SF...")
    starttime = datetime.now()
    chatbot = SF().chatters[collection_name]
    chatbot.set_trainer(ListTrainer)
    chatbot.train(read_custom(corpus_path))
    print("SF训练完成！")
    endtime = datetime.now()
    print("===========训练耗时: %s秒" % (endtime - starttime).seconds)
    
@singleton
class SF(object):
    def __init__(self):
        logging.info('预加载sf词向量模型...')
        logging.info('预加载SF所有实例...')
        labels = [file_name[:file_name.find("-")] for file_name in os.listdir(path_configer.get_chatter_corpus()) if
                  file_name.startswith("QA_sf_")]
        chatters = {}
        bot_name = current_app.config['DATABASE']
        # 根据不同的类型，创建不同的ChatBot实例
        for label in labels:
            chatters[label] = (
                ChatBot(
                    bot_name,
                    database=bot_name,
                    database_uri=current_app.config['DATABASE_URI'],
                    # 使用合适的词向量模型时开启
                    preprocessors=[
                        'kbqa_sf.train.chatter.sf.sf_preprocessors.sum_vecs_avg'
                    ],
                    statement_comparison_function=WordVecComparator(),
                    # statement_comparison_function=levenshtein_distance,
                    logic_adapters=[{'import_path': 'kbqa_sf.train.chatter.sf.sf_adapter.BestMatchExtLogicAdapter'}],
                    storage_adapter="kbqa_sf.train.chatter.sf.sf_mongo_storage.MongoDatabaseExtAdapter",
                    ext_collection_name=label,
                    read_only=True)
            )
        self.chatters = chatters
        logging.info('SF所有实例预加载完成！')    

# a：问题对象Statement，q:回答对象Statement
chatbot_.learn_response(a, q) 

@qac.route('/record', methods=['POST'])
def record():
    """
    将要学习的问题、答案、类别，写入文件learn目录下的wait-learn.txt、history-learn.txt
    :return:
    """
    qac_list = request.get_json()
    learn_path = path_configer.get_learn()
    wait_learn_path = "%s/%s" % (learn_path, "wait-learn.txt")
    history_learn_path = "%s/%s" % (learn_path, "history-learn.txt")
    with __record_lock:
        fa_wait = codecs.open(wait_learn_path, "a", encoding="utf-8")
        fa_history = codecs.open(history_learn_path, "a", encoding="utf-8")
        for qac_item in qac_list:
            q = qac_item["q"]
            a = qac_item["a"]
            c = qac_item["c"]
            if 0 < len(a) <= 300 and len(q) > 0 and len(c) > 0:
                content = 'Q %s\nA %s\nC %s\n' % (q, a, c)
                fa_wait.write(content)
                fa_history.write(
                    '%sT %s\n' % (content, time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time()))))
            else:
                return make_response(jsonify({'error': '参数不符合要求，请检查！'}), 400)
        fa_wait.close()
        fa_history.close()
        logging.debug("=========待学习问题记录完成！")
    return "success"


@qac.route('/learn/batch', methods=['GET'])
def learn_batch():
    """
    批量学习给定的问题和答案:
    重命名wait-learn.txt为learning.txt,读取learning.txt的内容进行学习
    :return:
    """
    _learn_new_batch_lock = threading.Lock()
    logging.debug("开始学习...")
    starttime = datetime.datetime.now()
    learn_path = path_configer.get_learn()
    wait_learn_path = "%s/%s" % (learn_path, "wait-learn.txt")
    learning_path = "%s/%s" % (learn_path, "learning.txt")
    with __record_lock:
        if os.path.exists(learning_path):
            # 若上一次的临时文件未能删除，就在这里删除。
            os.remove(learning_path)
            logging.info("=========发现上一次的临时文件未能删除，已删除！")
        if not os.path.exists(wait_learn_path):
            msg = "nothing"
            logging.info(msg)
            return msg
        os.rename(wait_learn_path, learning_path)
        logging.debug("重命名wait-learn.txt为learning.txt ...")
    with _learn_new_batch_lock:
        logging.debug("读取learning.txt的内容进行学习 ...")
        with codecs.open(learning_path, "r", encoding="utf-8") as fr:
            q = fr.readline().strip("\n\r")
            while q != "":
                a = fr.readline().strip("\n\r")
                assert a.strip("\n\r") != "", 'q,a,c格式无法匹配！缺少a！'
                c = fr.readline().strip("\n\r")
                assert c.strip("\n\r") != "", 'q,a,c格式无法匹配！缺少a！'
                # 添加q,a到指定的c类别文件；训练c对应的chatterbot
                logging.debug("添加%s,%s到指定的%s类别文件；训练对应的chatterbot ...", q, a, c)
                # 开始学习
                learn_(q, a, c[c.find(" ") + 1:])
                q = fr.readline().strip("\n\r")
        logging.debug("learning.txt学习全部完成...")
        logging.debug("完整的重新训练分类器模型 ...")
        IntentClassifier().full_retrain_clf()
        logging.debug("构建文本-向量索引文件，并存储 ...")
        IntentClassifier().build_text_vec_indx()
        logging.debug("加载文本向量索引文件 ...")
        IntentClassifier().load_text_vec_indx()
        # 删除临时的学习文件
        os.remove(learning_path)
        endtime = datetime.datetime.now()
        print("===========本次学习耗时: %s秒" % (endtime - starttime).seconds)
        logging.info("=========本次学习已全部完成！")
    return "success"
    
def learn_(q, a, c):
    """
    添加q,a到指定的c类别文件；训练c对应的chatterbot
    :param q: 问题
    :param a: 答案
    :param c: 分类
    :return:
    """
    file_names = [file_name for file_name in os.listdir(path_configer.get_chatter_corpus()) if
                  file_name.startswith(c)]
    if not file_names:
        logging.warning("未知的类别：%s,已忽略", c)
        return
    file_name = file_names[0]
    file_path = "%s/%s" % (path_configer.get_chatter_corpus(), file_name)
    # 追加到c对应的意图分类文件中
    with codecs.open(file_path, "a", encoding="utf-8") as fa:
        if len(q) > 0 and len(a) > 0:
            if os.path.getsize(file_path) == 0:
                fa.write('%s' % q)
            else:
                fa.write('\n%s' % q)
            fa.write('\n%s' % a)
    # 学习问答
    qa_learn(q, a, c)
    return "success"

def qa_learn(q, a, c):
    a_statement = Statement(a)
    q_statement = Statement(q)
    if c.startswith("QA_talk"):
        chat_bot = Talk().chat
    else:
        chat_bot = SF().chatters[c]
    chat_bot.learn_response(a_statement, q_statement)