第四章 實驗與結果
第四節 分析與討論
(一) 實驗結果分析
從第三節的四組實驗結果分析,發現Precision 的數值都無法大幅提升,所以
43
本研究請醫學專家人工分辨句子,原始句子43 句的分辨結果顯示 41 句為有關係,
2 句為無關係,這代表系統大多數判斷正確,原標示為負向的測試句子應該也是 屬於正向的測試集,只是剛好這些疾病與藥物的配對沒有出現在Clinical trials 的 標準答案內。探討此現象形成原因,主要是因為醫學文章大部分只會介紹疾病與
44
表28 實驗名稱中的(1)、(2)、Positive、Negative 說明如下:(1)代表標準答案 內的正向測試集、(2)代表標準答案外的正向測試集、Positive 代表為該實驗中系
45
表 29 中的整體 Recall 數值大多數超過 70%,最高可達 87%,只有實驗編號 1、2、3、5 的 Recall 數值小於 70%,本研究分兩種比較方式來說明,第一種是有 無使用卡方再分類的詞彙;第二種是有無去掉<2 的詞彙。
表 29 中的無卡方去掉<2 之組合(實驗編號 1、2)的關聯詞彙有 8 個,而有卡 方去掉<2 之組合(實驗編號 3、4、5、6)的關聯詞彙有 9 個和 10 個(如表 21 所示),
關聯的詞彙數量差異不大,但是Recall 也有差距 10%以上,這代表使用卡方再次 分類的詞彙,對於實驗Recall 數值很有幫助。
表 29 中的無卡方去掉<2 之組合(實驗編號 1、2)的關聯詞彙有 8 個,而無卡 方之組合(實驗編號 7、8)的關聯詞彙有 37 個,代表去掉<2 的關聯詞彙,會因為 詞彙量減少過多而造成Recall 值有所差異。
以直條圖統整單一詞性之Recall、F-Score 如圖 20、21 所示,x 軸為實驗 1. ~ 實驗16.,y 軸各為 Recall、F-Score 之百分比。
46
47
48
49
幅度最大。如單一名詞的無卡方實驗(如表 29 所示)與先看名詞再看動詞的無卡方 之實驗(如表 31 所示),因為動詞輔助判斷所以大幅度提升 Recall 數值。
以直條圖統整同時使用兩個詞性之Recall、F-Score 如圖 22、23 所示,x 軸為 實驗1. ~實驗 16.,y 軸各為 Recall、F-Score 之百分比。
圖22:同時使用兩個詞性的 Recall 直條圖
55%
60%
65%
70%
75%
80%
85%
90%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
先看動詞的極性再看名詞的極性 先看名詞的極性再看名詞的極性
50
圖23:同時使用兩個詞性的 F-Score 直條圖
圖22 可以觀察出先看動詞的極性再看名詞的極性,在實驗編號 2、5、8、10、
14 的 F-Score 數值比先看名詞的極性再看動詞的極性還要高,這代表標準答案外 的正向測試集以動詞為主較能提高效能。
經過更新後的實驗結果,Recall 的數值都能大幅提升,但是有一個狀況需深 入討論,就是原本判斷極性的規則。
原始的規則為同一組疾病與藥物配對的句子中,只要有一句被系統判斷成正 向,則把同一組疾病與藥物配對的測試句子中無法分類極性的句子都當成有關聯 (即正向),若是原本被系統判斷成負向,則不會改變極性,在此定義何謂無法判斷 極性的句子。原始本研究是採取多數表決,如果同一個句子的關聯詞彙比負向多 就當成有關聯,相反即為無關聯性,但是無法判斷結果會有兩種原因造成,第一 種是句子中沒有關聯、無關聯性的詞彙,第二種則是關聯詞彙的數量和無關聯性 詞彙的數量相同。針對此現象,本研究新增一個規則如下:當結果無法判斷時,
70.0%
75.0%
80.0%
85.0%
90.0%
95.0%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
先看動詞的極性再看名詞的極性 先看名詞的極性再看名詞的極性
51
先觀察該句是否有系統訓練得到的關聯單字,如果有就把該句判斷為有關係。如
“Sixty-five patients diagnosed with advanced <disease>PD<disease> were recruited in an open longitudinal 36-month study; 35 treated with <drug>levodopa<drug>-carbidopa intestinal gel (LCIG) and 30 were candidates for switching from oral PD treatment to LCIG.”,紅色標註為關聯詞彙;藍色標註為無關聯性詞彙,此句原本因為平手而
52
53
表33 中的實驗編號 1、3、5、9、11、13 的 Recall 數值比新增規則前微幅上 升,其原因與表31 相同,也是因為新增規則判斷所造成。
(二) 更新後的負向測試資料
前面的實驗結果顯示Precision 達到 82%以上,Recall 達到 63%以上,F-score
達到72%以上,只是這些都是正向資料的測試結果,為了探討負向資料測試的實 effects of antihypotensive drugs in patients with Parkinson's disease and thus no specific therapeutic strategy can be recommended.
Amphetamine PMID 25090657
Methamphetamine (METH) belongs to Amphetamine-type stimulants, METH abusers are at high risk of neurodegenerative disorders, including Parkinson's disease (PD).
Buspirone PMID 3026624
Buspirone is ineffective in the treatment of Parkinson's disease, but at anxiolytic doses (less than 40 mg/day) does not adversely affect parkinsonian disability.
54 parkinsonian signs and symptoms in outpatients with Parkinson's disease.
Haloperidol PMID 15909390
Haloperidol is the treatment of first choice for delirium due to somatic illness, except in patients with delirium due to drug use or medication, Parkinson's disease or Lewy body dementia.
Cinnarizine PMID 3302112
Cinnarizine should be added to the list of drugs capable of aggravating Parkinson's disease.
Physostigmine PMID 3766092
Physostigmine is contra-indicated in patients having symptoms and signs similar to those of Parkinson's disease, and the dose of physostigmine should also be reduced to 0.5 mg.
Vitamin E PMID 10209671
We studied the plasma chain-breaking antioxidants alpha carotene, beta carotene, lycopene, Vitamin A, Vitamin C, Vitamin E and a measure of total antioxidant capacity, TAC, in 79 patients with Alzheimer's disease (AD), 37 patients with vascular dementia (VaD), 18 patients with Parkinson's disease and dementia (PDem), and 58 matching controls, together with 41 deficiency only four had Parkinson's disease.
Norepinephrine PMID 26373628
Norepinephrine deficiency in NOH can arise peripherally, due to cardiovascular sympathetic denervation (as in pure autonomic failure, Parkinson's disease, and a variety of neuropathies), or centrally, due to a failure of viscerosensory signals to generate adequate sympathetic traffic to intact sympathetic nerve endings (as in multiple system atrophy).
55
(三) 負向測試資料分析
這10 句使用卡方值為 0.2 的關聯、無關聯性詞彙測試,結果系統判定 2 句為 負向,8 句為無法判定極性,Precision 為 100%,Recall 為 20%,F-Score 為 33.33%,
把這10 句負向測試句加上標準答案內、外的正向測試集讓系統判斷,則實驗結果 Precision 數值一樣為 100%,此結果代表本研究提出的規則樣式對於準確度有良 好的幫助。
56