自然言語処理 webスクレイピングによるトレーニングデータの集め方

自然言語処理 webスクレイピングによるトレーニングデータの集め方

今更ながらbotを作りたいなと思ったのですが、機械学習によりモデルを作るためのトレーニングデータを集めるところに苦労したため、トレーニングデータを集めるところだけにフォーカスを当ててて自分が実行した手段を公開したいと思います。使用した言語はpythonです。

今回自分がとったデータの集め方は、「http://ssmania.info/category/ 」 のサイトにアップされているキャラクターのやり取りをもとにトレーニングデータを抽出するやり方です。

全体の流れとしては、 1.リンクを集める 「http://ssmania.info/category/ 」 のサイトに掲載されているURLのリンクを一つのページにまとめる 2.リンクからやり取りを集める まとめたURLからページのやり取りを「input.txt」「output.txt」二つのファイルに保存する。 3.やり取りを単語ごとに分ける MeCabを使用して「input.txt」「output.txt」の二つの文章データを単語ごとに分ける。

前提条件

OS:Windows10 言語:Python 3.6.4

目次

1.リンクを集める 2.リンクからやり取りを集める 3.やり取りを単語ごとに分ける 4.結果

1.リンクを集める

「http://ssmania.info/category/ 」 のサイトの収集したいやり取りのアニメを選択する。 例1)以下のようにデータを取得したいアニメを選択する。 例2)以下のページに移動したら、URLをコピーしてプログラムCollectLinks.pyの11行目の[self.base_url='※※※'] の※※※のところにペーストする。

#!/Users/igaki/.pyenv/shims/python
# -*- coding: utf-8 -*-

import requests
from bs4 import BeautifulSoup
import types

class CharaScreipingu(object):
    def __init__(self):
        self.exclusion_url = "http://ssmania.info/"
        self.base_url ="http://ssmania.info/category/765%E3%83%97%E3%83%AD"

    def _boardLinkSearch(self):
        self.board_list_url = self.base_url+"?page={}"
        page = 1
        while True:
            link = list()
            print(page)
            board_top_html = requests.get(self.board_list_url.format(page))
            board_top_soup = BeautifulSoup(board_top_html.content, "html.parser")
            main_contents = board_top_soup.find(id="contents")
            if "該当記事はありませんでした" == board_top_soup.find(class_="alink").text:
                print("finish")
                break
            for a in main_contents.find_all('a'):
                tmp_link = a.get('href')
                if self.exclusion_url not in tmp_link:
                    link.append(a.get('href'))
            link_list = "\n".join(link)
            link_list = "\n"+link_list
            with open("./charaLink.txt", "a") as f:
                f.write(link_list)

            page += 1


def main():
    scrpeing = CharaScreipingu()
    scrpeing._boardLinkSearch()

if __name__ == "__main__":
    main()
    print("search Finish")

2.リンクからやり取りを集める

1で収集したリンクが保存されているファイルからキャラクター同士のやり取りを収集する。

#!/Users/igaki/.pyenv/shims/python
# -*- coding: utf-8 -*-

import requests
from bs4 import BeautifulSoup
import types


class WordScreipingu(object):
    def __init__(self):
        self.urlList ={"blog.livedoor.jp":self._livedoor,"ssbiyori.blog.fc2.com":self._ssbiyori,"yomicom.jp":self._yomicom,"potittoss.blog.jp":self._potittoss,"ss-m.net":self._ss_m,"s2-log.com":self._s2_log,"ss-navi.com":self._ss_navi,"horahorazoon.blog134.fc2.com":self._horahorazoon,"ayame2nd.blog.jp":self._ayame2nd,"ssimas72.blog.jp":self._ssimas72,"elephant.2chblog.jp":self._elephant,"morikinoko.com":self._morikinoko,"amnesiataizen.blog.fc2.com":self._amnesiataizen,"ssblog614.blog.fc2.com":self._ssblog614,"sssokuhou.com":self._sssokuhou,"invariant0.blog130.fc2.com":self._invariant0,"darusoku.xyz":self._darusoku,"ss-station.2chblog.jp":self._ss_station,"minnanohimatubushi.2chblog.jp":self._minnanohimatubushi,"ssmansion.xyz":self._ssmansion,"142ch.blog90.fc2.com":self._142ch,"tangaron3.sakura.ne.jp":self._tangaron3,"ssmaster.blog.jp":self._ssmaster,"dousoku.net":self._dousoku,"ssflash.net":self._ssflash,"lclc.blog.jp":self._lclc,"www.lovelive-ss.com":self._www,"maoyuss.blog.fc2.com":self._maoyuss,"ssspecial578.blog135.fc2.com":self._ssspecial578}

    def _fileRead(self):
        with open("./charaLink.txt", "r") as f:
            num = 1
            for link in f.readlines():
                self.urlList[link.split("/")[2]](link)
                num += 1

    def _takeCharaWord(self,word):
        char_num = len(word)
        flag = 0
        with open("./input.txt", "a") as inputfile:
            with open("./output.txt", "a") as outputfile:
                for num in range(char_num):
                    if word[num] == "「":
                        num += 1
                        if flag == 0:
                            while num < char_num and word[num] != "」":
                                inputfile.write(word[num])
                                num += 1
                            inputfile.write("\n")
                            flag = 1
                        else:
                            while num < char_num and word[num] != "」":
                                outputfile.write(word[num])
                                num += 1
                            outputfile.write("\n")
                            flag = 0
                    elif word[num] == "『":
                        num += 1
                        if flag%2 == 0:
                            while num < char_num and word[num] != "』":
                                inputfile.write(word[num])
                                num += 1
                            inputfile.write("\n")
                            flag = 1
                        else:
                            while num < char_num and word[num] != "』":
                                outputfile.write(word[num])
                                num += 1
                            outputfile.write("\n")
                            flag = 0

    def _livedoor(self,url):
        livedoor_html = requests.get(url)
        livedoor_html_soup = BeautifulSoup(livedoor_html.content, "html.parser")
        if livedoor_html_soup.find(id="character") == None:
            try:
                self._takeCharaWord(livedoor_html_soup.find(class_="article-body-inner").text)
            except:
                pass

    def _ssbiyori(self,url):
        pass

    def _yomicom(self,url):
        livedoor_html = requests.get(url)
        livedoor_html_soup = BeautifulSoup(livedoor_html.content, "html.parser")
        if livedoor_html_soup.find(_class="sh_heading_main_b") == None:
            try:
                self._takeCharaWord(livedoor_html_soup.find(class_="ently_text").text)
            except:
                pass


    def _potittoss(self,url):
        livedoor_html = requests.get(url)
        livedoor_html_soup = BeautifulSoup(livedoor_html.content, "html.parser")
        if livedoor_html_soup.find(id="character") == None:
            try:
                self._takeCharaWord(livedoor_html_soup.find(class_="article-body-inner").text)
            except:
                pass

    def _ss_m(self,url):
        pass

    def _s2_log(self,url):
        livedoor_html = requests.get(url)
        livedoor_html_soup = BeautifulSoup(livedoor_html.content, "html.parser")
        if livedoor_html_soup.find(id="character") == None:
            try:
                self._takeCharaWord(livedoor_html_soup.find(class_="article-body-inner").text)
            except:
                pass

    def _ss_navi(self,url):
        pass

    def _horahorazoon(self,url):
        pass

    def _ayame2nd(self,url):
        livedoor_html = requests.get(url)
        livedoor_html_soup = BeautifulSoup(livedoor_html.content, "html.parser")
        if livedoor_html_soup.find(id="character") == None:
            try:
                self._takeCharaWord(livedoor_html_soup.find(class_="article-body-inner").text)
            except:
                pass

    def _ssimas72(self,url):
        livedoor_html = requests.get(url)
        livedoor_html_soup = BeautifulSoup(livedoor_html.content, "html.parser")
        if livedoor_html_soup.find(id="character") == None:
            try:
                self._takeCharaWord(livedoor_html_soup.find(class_="article-body-inner").text)
            except:
                pass

    def _elephant(self,url):
        livedoor_html = requests.get(url)
        livedoor_html_soup = BeautifulSoup(livedoor_html.content, "html.parser")
        if livedoor_html_soup.find(id="character") == None:
            self._takeCharaWord(livedoor_html_soup.find(class_="article").text)

    def _morikinoko(self,url):
        livedoor_html = requests.get(url)
        livedoor_html_soup = BeautifulSoup(livedoor_html.content, "html.parser")
        if livedoor_html_soup.find(id="character") == None:
            try:
                self._takeCharaWord(livedoor_html_soup.find(class_="article-body-inner").text)
            except:
                pass

    def _amnesiataizen(self,url):
        pass

    def _ssblog614(self,url):
        pass

    def _sssokuhou(self,url):
        livedoor_html = requests.get(url)
        livedoor_html_soup = BeautifulSoup(livedoor_html.content, "html.parser")
        if livedoor_html_soup.find(id="character") == None:
            try:
                self._takeCharaWord(livedoor_html_soup.find(class_="article-body-inner").text)
            except:
                pass


    def _invariant0(self,url):
        pass

    def _darusoku(self,url):
        pass

    def _ss_station(self,url):
        livedoor_html = requests.get(url)
        livedoor_html_soup = BeautifulSoup(livedoor_html.content, "html.parser")
        if livedoor_html_soup.find(id="character") == None:
            try:
                self._takeCharaWord(livedoor_html_soup.find(class_="article-body-inner").text)
            except:
                pass

    def _minnanohimatubushi(self,url):
        livedoor_html = requests.get(url)
        livedoor_html_soup = BeautifulSoup(livedoor_html.content, "html.parser")
        if livedoor_html_soup.find(id="character") == None:
            try:
                self._takeCharaWord(livedoor_html_soup.find(class_="article-body-inner").text)
            except:
                pass

    def _ssmansion(self,url):
        pass

    def _142ch(self,url):
        pass

    def _tangaron3(self,url):
        pass

    def _ssmaster(self,url):
        livedoor_html = requests.get(url)
        livedoor_html_soup = BeautifulSoup(livedoor_html.content, "html.parser")
        if livedoor_html_soup.find(id="character") == None:
            try:
                self._takeCharaWord(livedoor_html_soup.find(class_="article-body-inner").text)
            except:
                pass

    def _dousoku(self,url):
        livedoor_html = requests.get(url)
        livedoor_html_soup = BeautifulSoup(livedoor_html.content, "html.parser")
        if livedoor_html_soup.find(id="character") == None:
            try:
                self._takeCharaWord(livedoor_html_soup.find(class_="article-body-inner").text)
            except:
                pass

    def _ssflash(self,url):
        livedoor_html = requests.get(url)
        livedoor_html_soup = BeautifulSoup(livedoor_html.content, "html.parser")
        if livedoor_html_soup.find(id="character") == None:
            try:
                self._takeCharaWord(livedoor_html_soup.find(class_="article-body-inner").text)
            except:
                pass

    def _lclc(self,url):
        livedoor_html = requests.get(url)
        livedoor_html_soup = BeautifulSoup(livedoor_html.content, "html.parser")
        if livedoor_html_soup.find(id="character") == None:
            try:
                self._takeCharaWord(livedoor_html_soup.find(class_="article-body-inner").text)
            except:
                pass

    def _www(self,url):
        pass

    def _maoyuss(self,url):
        pass

    def _ssspecial578(self,url):
        pass

    def _boardLinkSearch(self):
        self.board_list_url = "http://ssmania.info/category/765%E3%83%97%E3%83%AD?page={}"
        page = 1
        while True:
            link = list()
            print(page)
            board_top_html = requests.get(self.board_list_url.format(page))
            board_top_soup = BeautifulSoup(board_top_html.content, "html.parser")
            main_contents = board_top_soup.find(id="contents")
            if "該当記事はありませんでした" == board_top_soup.find(class_="alink").text:
                print("finish")
                break
            for a in main_contents.find_all('a'):
                tmp_link = a.get('href')
                if self.exclusion_url not in tmp_link:
                    link.append(a.get('href'))
            link_list = "\n".join(link)
            link_list = "\n"+link_list
            with open("./charaLink.txt", "a") as f:
                f.write(link_list)

            page += 1

def temp():
    link_list = list()
    with open("./charaLink.txt", "r") as f:
        for link in f:
            link_list.append(link)
    link_type = list()
    for link in link_list:
        if link.split("/")[2] not in link_type:
            link_type.append(link.split("/")[2])
    print(len(link_type))
    for link in link_type:
        print(link)

def main():
    scrpeing = WordScreipingu()
    scrpeing._fileRead()



if __name__ == "__main__":
    main()
    print("search Finish")

3.やり取りを単語ごとに分ける

seq2seqにより学習するためには文章を単語ごとに分ける必要があります。そのため今回はMeCabというオープンソースである形態素解析器を使用しました。 取得したデータが格納されている「input.txt」と「output.txt」からデータを抽出して「input_result.txt」「output_result.txt」にそれぞれ出力しました。

# -*- coding: utf-8 -*-
import MeCab
import sys

mode = MeCab.Tagger("-Ochasen")

num = 0
with open("./input.txt","r",encoding="utf-8_sig") as f:
    for line in f.readlines():
        num += 1
        line_result = mode.parse(line).split("\n")
        with open("./input_result.txt","a",encoding="utf-8") as result_file:
            for word in line_result:
                if word.split("\t")[0] != "EOS":
                    result_file.write(word.split("\t")[0])
                    result_file.write(" ")
            result_file.write("\n")

num = 0
with open("./output.txt","r",encoding="utf-8_sig") as f:
    for line in f.readlines():
        num += 1
        line_result = mode.parse(line).split("\n")
        with open("./output_result.txt","a",encoding="utf-8") as result_file:
            for word in line_result:
                if word.split("\t")[0] != "EOS":
                    result_file.write(word.split("\t")[0])
                    result_file.write(" ")
            result_file.write("\n")

4.結果

結果以下のような内容のデータが収集できました。

...
うーん 、 ブラック の 缶 珈琲 って なんとなく 味気 ない 気 が し て  
で さぁ 春香  
今日 な んで 事務所 来 た の ？  
いや 、 今日 春香 オフ じゃん  
… それ 、 先週 別 の 日 に なっ たって 言っ た と 思う ん だ けど …  
...