自然语言处理（NLP）：07 fastText训练中文模型-文本分类

fastText 另外两种安装方式

conda install 方式：速度慢

https://anaconda.org/conda-forge/fasttext

windows 版本下可以通过whl安装（fasttext‑0.9.1‑cp36‑cp36m‑win32.whl） ，windows 下可以使用这个安装

https://www.lfd.uci.edu/~gohlke/pythonlibs/#fasttext

fastText 训练

import fastText
import fastText
import pandas as pd
import numpy as np
from sklearn.metrics import confusion_matrix,precision_recall_fscore_support

# 训练
'''

dtrain.txt 和dtest.txt 数据格式 如下：

__label__2 中新网 日电 日前 上海 国际
__label__0 两人 被捕 警方 指控 非法
__label__3 中旬 航渡 过程 美军 第一
__label__1 强强 联手 背后 品牌 用户 双赢
'''
model = fastText.train_supervised(
    '../data/dtrain.txt',
    lr=0.1,
    dim=200,
    epoch=50,
    neg=5,
    wordNgrams=2,
    label="__label__"
    )
# 预测
result = model.test('../data/dtest.txt')
print('y_pred = ',y_pred)

# 保存model
model_path = '../model/fastText_model.pkl'
model.save_model(model_path)

# 计算分类的metrics
#绘制precision、recall、f1-score、support报告表
def eval_model(y_true, y_pred, labels):
    # 计算每个分类的Precision, Recall, f1, support
    p, r, f1, s = precision_recall_fscore_support(y_true, y_pred)
    # 计算总体的平均Precision, Recall, f1, support
    tot_p = np.average(p, weights=s)
    tot_r = np.average(r, weights=s)
    tot_f1 = np.average(f1, weights=s)
    tot_s = np.sum(s)
    res1 = pd.DataFrame({
        u'Label': labels,
        u'Precision': p,
        u'Recall': r,
        u'F1': f1,
        u'Support': s
    })
    res2 = pd.DataFrame({
        u'Label': ['总体'],
        u'Precision': [tot_p],
        u'Recall': [tot_r],
        u'F1': [tot_f1],
        u'Support': [tot_s]
    })
    res2.index = [999]
    res = pd.concat([res1, res2])
    return res[['Label', 'Precision', 'Recall', 'F1', 'Support']]

cate_dic = {'entertainment': 0, 'technology': 1, 'sports': 2, 'military': 3, 'car': 4}
dict_cate = dict(('__label__{}'.format(v),k) for k,v in cate_dic.items())
y_true= []
y_pred = []
with open('../data/dtest.txt','r',encoding='utf-8') as f:
    for line in f.readlines():
        line = line.strip()
        splits = line.split(" ")
        label = splits[0]
        words = [" ".join(splits[1:])]
        label = dict_cate[label]
        y_true.append(label)
        y_pred_results = clf.predict(words)[0][0][0]
        y_pred.append(dict_cate[y_pred_results])
print("y_true = ",y_true[:5])
print("y_pred = ",y_pred[:5])
print('y_true length = ',len(y_true))
print('y_pred length = ',len(y_pred))

print('keys = ',list(cate_dic.keys()))

y_true =  ['sports', 'car', 'car', 'technology', 'entertainment']
y_pred =  ['sports', 'car', 'car', 'technology', 'entertainment']
y_true length =  87581
y_pred length =  87581
keys =  ['entertainment', 'technology', 'sports', 'military', 'car']

eval_model(y_true,y_pred,list(cate_dic.keys()))

	Label	Precision	Recall	F1	Support
0	entertainment	0.934803	0.827857	0.878086	8400
1	technology	0.906027	0.923472	0.914666	26696
2	sports	0.881885	0.911727	0.896558	11555
3	military	0.943886	0.931749	0.937778	22476
4	car	0.857226	0.873252	0.865165	18454
999	总体	0.905035	0.904294	0.904270	87581

模拟在线预测

# 加载模型
model_path = '../model/fastText_model.pkl'
clf = fastText.load_model(model_path)

cate_dic = {'entertainment': 0, 'technology': 1, 'sports': 2, 'military': 3, 'car': 4}
print(cate_dic)

dict_cate = dict(('__label__{}'.format(v),k) for k,v in cate_dic.items()) 
print(dict_cate)

{'entertainment': 0, 'technology': 1, 'sports': 2, 'military': 3, 'car': 4}
{'__label__0': 'entertainment', '__label__1': 'technology', '__label__2': 'sports', '__label__3': 'military', '__label__4': 'car'}

• 预测案例1-汽车类
  摘自今日头条： https://www.toutiao.com/a6714271125473346055/

import jieba
text = "奥迪A3、宝马1系和奔驰A级一直纠缠不休的三个冤家"
words = [word for word in jieba.lcut(text)]
print('words = ',words)
data = " ".join(words)

# predict
results = clf.predict([data])
y_pred = results[0][0][0]
print("y_pred results = ",dict_cate[y_pred])

words =  ['奥迪', 'A3', '、', '宝马', '1', '系', '和', '奔驰', 'A', '级', '一直', '纠缠', '不休', '的', '三个', '冤家']
y_pred results =  car

• 预测案例2-军事类
  摘自今日头条新闻： https://www.toutiao.com/a6714188329937535496/

import jieba
text = "谁说文物只能躺在博物馆，想买一架梦想中的战斗机开着兜风吗？"
words = [word for word in jieba.lcut(text)]
print('words = ',words)
data = " ".join(words)
# predict
results = clf.predict([data])
y_pred = results[0][0][0]
print("y_pred results = ",dict_cate[y_pred])

words =  ['谁', '说', '文物', '只能', '躺', '在', '博物馆', '，', '想', '买', '一架', '梦想', '中', '的', '战斗机', '开着', '兜风', '吗', '？']
y_pred results =  military

• 预测案例3-娱乐类
  我们从 今日头条： https://www.toutiao.com/a6689675139333751299/ 拷贝标题来进行预测

import jieba
text = "陈晓旭：从完美林黛玉到身家过亿后剃度出家，她戏里戏外都是传奇"
words = [word for word in jieba.lcut(text)]
print('words = ',words)
data = " ".join(words)
results = clf.predict([data])
y_pred = results[0][0][0]
print("y_pred results = ",dict_cate[y_pred])

words =  ['陈晓旭', '：', '从', '完美', '林黛玉', '到', '身家', '过', '亿后', '剃度', '出家', '，', '她', '戏里', '戏外', '都', '是', '传奇']
y_pred results =  entertainment

• 预测案例4-体育类
  摘自今日头条：https://www.toutiao.com/a6714266792253981192/

import jieba
text = "男女有别！国乒主力参加马来西亚T2联赛 男队站着吃自助女队吃桌餐"
words = [word for word in jieba.lcut(text)]
print('words = ',words)
data = " ".join(words)
results = clf.predict([data])
y_pred = results[0][0][0]
print("y_pred results = ",dict_cate[y_pred])

words =  ['男女有别', '！', '国乒', '主力', '参加', '马来西亚', 'T2', '联赛', ' ', '男队', '站', '着', '吃', '自助', '女队', '吃', '桌餐']
y_pred results =  sports

• 预测案例5-科技类

import jieba
text = "摩托罗拉One Macro将是最新一款Android One智能手机"
words = [word for word in jieba.lcut(text)]
print('words = ',words)
data = " ".join(words)
results = clf.predict([data])
y_pred = results[0][0][0]
print("y_pred results = ",dict_cate[y_pred])

words =  ['摩托罗拉', 'One', ' ', 'Macro', '将', '是', '最新', '一款', 'Android', ' ', 'One', '智能手机']
y_pred results =  technology

Flask Web 服务在线预测

http://127.0.0.1:5000/v1/p?q=xxxxx

其中： q 是要预测的样本

# -*- coding: UTF-8 -*-
import jieba
import fastText
from flask import Flask
from flask import request

app = Flask(__name__)
model_path = '../model/fastText_model.pkl'
clf = fastText.load_model(model_path)
cate_dic = {'entertainment': 0, 'technology': 1, 'sports': 2, 'military': 3, 'car': 4}
dict_cate = dict(('__label__{}'.format(v), k) for k, v in cate_dic.items())
print(dict_cate)


@app.route('/')
def hello_world():
    return 'Hello World!'


@app.route('/v1/p', methods=['POST', 'GET'])
def predict():
    if request.method == 'POST':
        q = request.form['q']
    else:
        q = request.args.get('q', '')
        print('q = ', q)

    print('input data:', q)
    words = [word for word in jieba.lcut(q)]
    print('words = ', words)
    data = " ".join(words)
    results = clf.predict([data])
    y_pred = results[0][0][0]
    return dict_cate[y_pred]

if __name__ == '__main__':
    app.run()

来源：https://blog.csdn.net/shenfuli/article/details/98882655