[Day 13] Google Natural Language - 子系列最終章

Published: Sep 21, 2019 Joseph

我們今天玩的是AutoML Natural Language，可以透過GCP左邊的sidebar，找到Natural Language打開，或直接點這個連結
GCP

AutoML Natural Language的收費方式記得先看這裡：https://cloud.google.com/natural-language/automl/pricing?_ga=2.125750730.-1764044900.1562297905

這邊很有意思的是在AutoML Text Classification跟AutoML Sentiment Analysis(圖中兩個紅色的項目)在寫文章的時候，如果你驗證的不是google的第一個帳號，而是選其他帳號登入的話，允許授權後會一直跳500錯誤。
switch account
500 error

沒關係，這部份我們就選AutoML Entity Extraction，就玩他這個功能，至少他的授權是成功的。
然後我們就照這份Quickstarts來操作，裡面也有範例檔案gs://cloud-ml-data/NL-entity/dataset.csv

那我們就來New一個dataset吧！
new dataset

Import是吃一個CSV檔案，也有提供檔案格式，我們載下來的檔案長這個樣子

1
2
3

TRAIN,gs://cloud-ml-data/NL-entity/train.jsonl
TEST,gs://cloud-ml-data/NL-entity/test.jsonl
VALIDATION,gs://cloud-ml-data/NL-entity/validation.jsonl%

前面的SET可加可不加，但後面要個別帶入jsonl的資料集。輸入後長下面這樣
import

經過一段時間的Processing text items(大概10分鐘)，就會有結果
text
至於這些label怎麼來的？我們就來看看jsonl裡面的內容吧。

{"annotations": [{"text_extraction": {"text_segment": {"end_offset": 54, "start_offset": 27}}, "display_name": "SpecificDisease"}, {"text_extraction": {"text_segment": {"end_offset": 173, "start_offset": 156}}, "display_name": "SpecificDisease"}, {"text_extraction": {"text_segment": {"end_offset": 179, "start_offset": 176}}, "display_name": "SpecificDisease"}, {"text_extraction": {"text_segment": {"end_offset": 246, "start_offset": 243}}, "display_name": "Modifier"}, {"text_extraction": {"text_segment": {"end_offset": 340, "start_offset": 337}}, "display_name": "Modifier"}, {"text_extraction": {"text_segment": {"end_offset": 698, "start_offset": 695}}, "display_name": "Modifier"}], "text_snippet": {"content": "1301937\tMolecular basis of hexosaminidase A deficiency and pseudodeficiency in the Berks County Pennsylvania Dutch.\tFollowing the birth of two infants with Tay-Sachs disease ( TSD ) , a non-Jewish , Pennsylvania Dutch kindred was screened for TSD carriers using the biochemical assay . A high frequency of individuals who appeared to be TSD heterozygotes was detected ( Kelly et al . , 1975 ) . Clinical and biochemical evidence suggested that the increased carrier frequency was due to at least two altered alleles for the hexosaminidase A alpha-subunit . We now report two mutant alleles in this Pennsylvania Dutch kindred , and one polymorphism . One allele , reported originally in a French TSD patient ( Akli et al . , 1991 ) , is a GT-- > AT transition at the donor splice-site of intron 9 . The second , a C-- > T transition at nucleotide 739 ( Arg247Trp ) , has been shown by Triggs-Raine et al . ( 1992 ) to be a clinically benign \" pseudodeficient \" allele associated with reduced enzyme activity against artificial substrate . Finally , a polymorphism [ G-- > A ( 759 ) ] , which leaves valine at codon 253 unchanged , is described  .\n "}}

jsonl就是把json一列一列表達的一種格式

這是train.jsonl裡的第一筆資料，就是對content裡的文字把開始跟結束index的範圍框起來，給他一個label。舉例來說{"end_offset": 54, "start_offset": 27}給他SpecificDisease這個label，而27~54就是hexosaminidase A deficiency這個字，看明白點就是下圖的結果。