Download presentation
Presentation is loading. Please wait.
1
臺灣大學資訊工程學系 高紹航 臺灣大學外國語文學系 高照明
詞義辨識 機器學習演算法特徵的選取與組合 臺灣大學資訊工程學系 高紹航 臺灣大學外國語文學系 高照明 本文由高照明老師與高紹航同學撰寫並發表於 ROCLING 2007， 再由 Patty Liu於SLP Lab meeting時報告 Presented by Patty Liu
2
Outline Word Sense Disambiguation Senseval-2 Bayesian Classification
Forward Sequential Selection Algorithm The features we applied Result
3
Word Sense Disambiguation
A word may have more then one sense Ex. Bank-銀行, 河堤, 庫 The task of WSD is to automatically identify the correct sense in a given context.
4
Senseval-2 Published in 2001 Senseval-2 English lexical sample
73 different target words, including nouns, verbs , and adjectives.
5
Senseval Senseval-1 Senseval-2 Senseval-3 Semeval-1 / Senseval-4 Date
Senseval-1 Senseval-2 Senseval-3 Semeval-1 / Senseval-4 Date 1998/9 2001/7 2004 /7 2007/7 Location Sussex , England Toulouse, France Barcelona, Spain Prague, Czech Republic Content English, French, and Italian included tasks for Basque, Chinese, Czech, Danish, Dutch, English, Estonian, Italian, Japanese, Korean, Spanish, Swedish included 14 different tasks for core word sense disambiguation, as well as identification of semantic roles, multilingual annotations, logic forms, subcategorization acquisition Note in conjunction with ACL 2001 in conjunction with ACL 2004 in conjunction with ACL 2007
6
Senseval-2 Competition
7
Corpora of Senseval-2 Competition
Example： ..\corpora\english-lex-sample\train\eng-lex-sample.training.xml <instance id="art.40001" docsrc="bnc_ACN_245"> <answer instance="art.40001"senseid="art%1:06:00::"/> <context>Their multiscreen projections of slides and film loops have featured in orbital parties, at the Astoria and Heaven, in Rifat Ozbek's 1988/89 fashion shows, and at Energy's recent Docklands all-dayer.From their residency at the Fridge during the first summer of love, Halo used slide and film projectors to throw up a collage of op-art patterns, film loops of dancers like E-Boy and Wumni, and unique fractals derived from video feedback.&bquo;We're not aware of creating a visual identify for the house scene, because we're right in there.We see a dancer at a rave, film him later that week, and project him at the next rave.&equo;[hi]Ben Lewis [/hi] Halo can be contacted on [ptr][/p] [caption] <head>Art</head>you can dance to from the creative group called Halo [/caption] [/div2] [div2] [head] </context> </instance>
8
Bayesian Classification
Suppose the target word has k senses, s1, s2, …, sk Find s’ such that is maximum, c is the context or features of the target word
9
P(s’|c)
10
Forward Sequential Selection Algorithm
Used in feature selection First let First add the best feature into S and then iteratively add into S the best feature in the remaining feature set until the performance cannot be improved. The final S is approximately the best feature set
11
The features we applied
We tried 9 feature, named F1, F2, …F9
12
The features we applied-F1
The words around the target word excluding stop words such as “is”, “a” Best window size is 3 Window Size Precision (%) 1 52.7 2 54.2 3 54.6 4 5 54.1
13
The features we applied-F2
Similar to F1, but include the information of relative position of the target word Include stop words For example, “The art of design” {(The, -1), (of, 1), (design, 2)}
14
Window Size Test of F2 Best window size is 1 Window Size Precision (%)
54.9 2 53.6 3 51.1 4 47.9
15
The features we applied-F3
Similar to F2, but use part-of-speech instead “The art of design” design: (n, 2) Best window size is 1 Window Size Precision (%) 1 44.6 2 35.5 3 30.7 4 27.7
16
The features we applied-F4
Ngrams containing the target word . “The art of design” {(The-art), (art-of), (The-art-of), (art-of-design), (The-art-of-design)} Best window size is 3 Window Size Precision (%) 1 48.2 2 56.9 3 57.8 4 5
17
The features we applied-F5
Similar to F4, but use part-of-speech instead such as (n-prep-n) for art-of-design Best window size is 4 Window Size Precision (%) 1 48.2 2 52.1 3 53.8 4 54.2 5 54.1
18
The features we applied-F6
Use word sketch in the sketch engine to extract all possible collocations involving the target word Best window size is 5 Best dependency type is {modifiers, object, n_modifier, a_modifier, and/or, modifier}
19
Window Size Test of F6 Precision(%) Window Size 1 50.5 11 51.6 2 51.1
12 51.4 3 13 4 51.8 14 5 52.0 15 50.8 6 7 8 51.5 9 10
20
Minimum Salience Test of F6
Precision(%) 0.0 52.0 1.0 51.8 2.0 3.0 51.3
21
F6 Step1 Type Precision(%) object 49.2 and/or 49.4 object_of 47.5 pp*
subject 48.4 possessor 46.6 subject_of 48.0 possessed 47.6 a_modifier 48.1 modifier n_modifier 49.0 part* 48.5 modifies 50.1 *comp_of 48.3 *comp 48.2
22
F6 Step2 Type Precision(%) object 51.1 and/or 50.8 object_of 50.1 pp*
50.9 subject 50.2 possessor subject_of possessed 50.0 a_modifier 50.3 modifier n_modifier 50.7 part* *comp *comp_of
23
F6 Step3 Type Precision(%) object_of 51.2 and/or 51.5 subject 51.1 pp*
51.4 subject_of possessor a_modifier 51.3 possessed 51.0 n_modifier modifier *comp part* *comp_of
24
F6 Step4 Type Precision(%) object_of 51.6 and/or 51.7 subject 51.5 pp*
subject_of possessor a_modifier possessed *comp 51.4 modifier *comp_of part* 51.3
25
F6 Step5 Type Precision(%) object_of 51.6 and/or 51.9 subject 51.7 pp*
51.8 subject_of possessor *comp_of possessed *comp modifier part*
26
F6 Step6 Type Precision(%) object_of 51.9 pp* subject possessor
subject_of possessed *comp_of modifier 52.0 *comp part*
27
F6 Step7 Type Precision(%) object_of 52.0 pp* 51.9 subject possessor
subject_of possessed *comp_of part* *comp
28
Word Sketch in the Sketch Engine
29
The features we applied-F7
Use the Stanford parser to identify the dependency relations, ex: object_of, modifies The Precision is 54.6% . Stanford was developed by Klein and Manning in 2003
30
Stanford Parser
31
Some output of Stanford parser
det(government-2, The-1) nsubj(established-4, government-2) advmod(established-4, first-3) amod(system-8, modern-5) amod(system-8, criminal-6) nn(system-8, investigation-7) dobj(established-4, system-8) prep(system-8, in-9) pobj(in-9, )
32
The features we applied-F8
Use the top HowNet semantic features of the word before and after the target word. The Precision is 47.2% . HowNet is developed by董振東. Example: Hownet representations of 醫生 ‘doctor’ {human| 人:HostOf={Occupation| 職 位},domain={medical| 醫},{doctor| 醫 治:agent={~}}}
33
The features we applied-F9
First use Stanford parser to identify words which have dependency relations with the target word. Then use the top HowNet semantic features as feature The Precision is 54.1% .
34
Which features are more important in WSD?
Type Result (%) F1 54.6 F2 54.9 F3 44.6 F4 57.8 F5 54.2 F6 52.0 F7 F8 47.2 F9 54.1
35
Results Step F1 F2 F3 F4 F5 F6 F7 F8 F9 1st 54.6 54.9 44.6 57.8 54.2
52.0 47.2 54.1 2nd 58.9 58.5 56.2 56.8 58.2 59.6 59.2 3rd 60.7 60.1 58.1 60.2 59.1 59.7 4th 61.2 58.8 60.6 60.4 5th 60.3 59.4 60.8 61.1
36
Result Best feature set is {F1, F2, F4, F7} The Precision 61.2%
The best performance in Senseval-2 is 64.2%.
37
Conclusion In terms of the collocation types, the feature of object and modifier play more important roles than subject in WSD.
38
Future Research New features and other machine learning algorithms like SVM and CRF might improve the performance.
Similar presentations
