中国管理科学 | 使用业绩说明会文本数据测量上市公司前瞻性信息

pip3 install openpyxl
pip3 install pandas

import pandas as pd

df = pd.read_excel('业绩说明会问答05-23.xlsx')
df['text'] = df['提问内容']+df['回答内容']
df.dropna(subset=['text'], inplace=True)
df.head()

with open('业绩说明会05-23.txt', 'a+', encoding='utf-8') as f:
    
    #为了保证问答上下文一致， 问答要放在相邻处
    text = ''.join(df['text'].to_list())
    f.write(text)

pip3 install cntext==1.8.6

import cntext as ct
import os


#Init W2VModels. Support English and Chinese
model = ct.W2VModels(cwd=os.getcwd(), 
                     lang='chinese')  #corpus data w2v_corpus.txt


#训练结束后，「业绩说明会05-23.100.6.bin」会出现在「output/Word2Vec」文件夹内 
model.train(input_txt_file='业绩说明会05-23.txt', 
            model_name='业绩说明会05-23.100.6.bin')

seedwords = ['计划', '预计', '未来', '目标', '可能', '如果', '机遇', '预期', '挑战', '预测', '今后', '目的', '契机', '前景', '希望', '展望', '相信', '愿景', '期待', '明年', '期望']

from gensim.models import KeyedVectors

def load_w2v(w2v_path):
    """
    Load word2vec model

    Args:
        w2v_path (str): path of word2vec model

    Returns:
        model: word2vec model
    """
    print('Loading word2vec model...')
    model = KeyedVectors.load(w2v_path)
    return model


wv = load_w2v('Embeddings/业绩说明会05-23.100.6.bin')
wv2 = load_w2v('Embeddings/mda01-21.200.6.bin')

#词汇量
print(len(wv.index_to_key))
print(len(wv2.index_to_key))

#查询某词的词向量
wv.get_vector('创新')

#查询多个词的词向量
#wv.get_mean_vector(['创新', '研发'])

def expand_dictionary(wv, seedwords, topn=100):
    """
    According to the seed word file, select the top n words with the most similar semantics and save them in the directory save_dir.
    
    Args:
        wv (Word2VecKeyedVectors): the word embedding model
        seedwords (list): 种子词
        topn (int, optional): Set the number of most similar words to retrieve to topn. Defaults to 100.
        save_dir (str, optional): the directory to save the candidate words. Defaults to 'Word2Vec'.
    
    Returns:
    """
    simidx_scores = []

    similars_candidate_idxs = [] #the candidate words of seedwords
    dictionary = wv.key_to_index
    seedidxs = [] #transform word to index
    for seed in seedwords:
        if seed in dictionary:
            seedidx = dictionary[seed]
            seedidxs.append(seedidx)
    for seedidx in seedidxs:
        # sims_words such as [('by', 0.99984), ('or', 0.99982), ('an', 0.99981), ('up', 0.99980)]
        sims_words = wv.similar_by_word(seedidx, topn=topn)
        #Convert words to index and store them
        similars_candidate_idxs.extend([dictionary[sim[0]] for sim in sims_words])
    similars_candidate_idxs = set(similars_candidate_idxs)
    
    for idx in similars_candidate_idxs:
        score = wv.n_similarity([idx], seedidxs)
        simidx_scores.append((idx, score))
    simidxs = [w[0] for w in sorted(simidx_scores, key=lambda k:k[1], reverse=True)]

    simwords = [str(wv.index_to_key[idx]) for idx in simidxs][:topn]

    resultwords = []
    resultwords.extend(seedwords)
    resultwords.extend(simwords)
    
    return resultwords


#为了节省板面，这里设置为50
#论文中经过筛选留下174个词，实际上topn应该远大于174，
expand_dictionary(wv=wv, 
                  #前瞻性种子词
                  seedwords= ['计划', '预计', '未来', '目标', '可能', '如果', '机遇', '预期', '挑战', '预测', '今后', '目的', '契机', '前景', '希望', '展望', '相信', '愿景', '期待', '明年', '期望'],
                  topn=50)

['计划',
 '预计',
 ......
 '几年',
 '积极影响',
 '有何',
 '谢谢您提问',
 '今后',
 '这块',
 '近几年',
 '近两年',
 '请问李',
 '裁员',
 '亮点',
 '准备采取',
 '将会',
 '接下来',
 '有何规划',
 '前景',
 '管理层是否',
 '未来几年',
 '有没有新',
 '发展状况',
 '一块',
 '当前',
 '很大',
 '这块业务',
 'LNG船',
 '具体措施您好',
 '当下',
 '是否能够',
 '明后',
 '一个台阶',
 '是否符合',
 '巨大',
 '预判',
 '对此',
 '未来三年',
 '资本开支',
 '不少',
 '未来是否',
 '这方面',
 '看法',
 '今年以来',
 '疫情结束',
 '想知道',
 '取得不错',
 '谈谈',
 '一步',
 '今年是否',
 '发展前景',
 '东宝',
 '现状']

expand_dictionary(wv=wv2, 
                  #前瞻性种子词
                  seedwords= ['计划', '预计', '未来', '目标', '可能', '如果', '机遇', '预期', '挑战', '预测', '今后', '目的', '契机', '前景', '希望', '展望', '相信', '愿景', '期待', '明年', '期望'],
                  topn=50)

['计划',
 '预计',
  ......
 '相信',
 '将会',
 '未来',
 '希望',
 '预见',
 '预期',
 '可能',
 '必将',
 '应该',
 '未来几年',
 '今后',
 '有望',
 '目标',
 '这一',
 '当前',
 '当下',
 '无疑',
 '期望',
 '接下来',
 '意味着',
 '背景',
 '期待',
 '近期',
 '下一阶段',
 '机会',
 '看到',
 '预示',
 '能够',
 '短期内',
 '未来一段时间',
 '将来',
 '展望未来',
 '必须',
 '真正',
 '眼光',
 '必然',
 '还会',
 '预计',
 '未来十年',
 '机遇',
 '可能性',
 '后续',
 '潜在',
 '决心',
 '信心',
 '仍然',
 '非常',
 '这为',
 '未来五年',
 '短时间']

import cntext as ct

#cntext已内置了论文中的174个前瞻性词集
fls_words = ct.load_pkl_dict('Chinese_FLS.pkl')['Chinese_FLS']
fls_words

['计划',
 '预计',
 '未来',
 '目标',
 ......
 '企业宗旨',
 '宗旨',
 '该愿景',
 '愿望',
 '心愿',
 '盼望',
 '祝愿',
 '今年年底',
 '今年底',
 '明年初',
 '第二季度',
 '上半年',
 '下半年',
 '本月底',
 '下周',
 '马上',
 '厚望',
 '期盼',
 '鞭策',
 '梦想',
 '愿']

df2 = df.groupby(['股票代码', '会计年度'])['text'].apply(''.join).reset_index()
df2.dropna(subset=['text'], inplace=True)

df2

import cntext as ct
import jieba


fls_words = ct.load_pkl_dict('Chinese_FLS.pkl')['Chinese_FLS']
stopwords = ct.load_pkl_dict('STOPWORDS.pkl')['STOPWORDS']['chinese']

def compute_fls(text):
    num = 0
    words = jieba.lcut(text)
    words = [w for w in words if w not in stopwords]
    for word in words:
        if word in fls_words:
            num+=1
    #+1是为了防止分母为0的情况
    return num/(len(words)+1)

df2['Forward'] = df2['text'].apply(compute_fls)

df2.head()

print('Forward最小值: ', df2['Forward'].min())
print('Forward中位数: ', df2['Forward'].median())
print('Forward均值: ', df2['Forward'].mean())
print('Forward最大值: ', df2['Forward'].max())
print('Forward标准层: ', df2['Forward'].std())