NAIST-NTT System Description for Patent Translation Task at NTCIR-7
-- Semi-supervised Learning of Bilingual Lexicon for Technical Terms from Wikipedia -Mamoru Komachi (NAIST, Japan), Masaaki Nagata (NTT, Japan), Yuji Matsumoto (NAIST, Japan)
Background
The cost of hand-tagging resources is crucial for MT
Wikipedia has become one of the important source for knowledge acquisition
Wikipedia’s link structure has attracted attention in NLP fields, and the link structure turns out to be useful for extracting bilingual
lexicon from Wikipedia [Adafre and Rijke 2006; Erdmann 2008]
This study aims at:
Domain adaptation (word sense disambiguation per domain) of term extraction
Automatic refinement of translation pairs extracted from Wikipedia
Approach
Assumption: one sense per domain [Thelen and Riloff 2002]
→Use “relatedness” measure in link analysis to select the most relevant sense to the domain
Ranked
bilingual
lexicon
Wikipedia
Use interlingual links in Wikipedia to extract translation pairs
Create a bipartite graph using Wikipedia abstract and interlingual links
Manually select seed translation pairs from training corpora
Select top-ranked translation pairs given seed translation pairs
(Plant, 植物)
Seed
translation
pairs
Corpora
(Target
domain)
Graph Laplacian
Regularized Laplacian matrix
L = D- A

n

organism

n

1
R

(

L
)

(
I

L
)

(Plant, 工場)
Wikipedia
English
abstracts
n

0
System overview
A: adjacency matrix of the graph
D: (diagonal) degree matrix
β:parameter
Each column of
Rβ gives the rankings relative to a node
施設
Wikipedia
Japanese
abstracts
(Library, 図書館)
building
Bipartite graph construction steps
1. Add translation pairs (en, ja) as white nodes
2. Add bag-of-content words appearing abstracts of both languages as black nodes (Airport, 空港)
3. Add edges from translation pairs to co-occurring patterns (black nodes)
Bipartite graph from Wikipedia link structure
Experiment
Extract bilingual lexicon from Wikipedia: 222,739 translation pairs (197,770 after filtering)
Split the bilingual lexicon randomly into 8 sub-lexicons, and choose 5 seeds for each sub lexicon (total 8 * 5 = 40 seeds)
After applying the regularized Laplacian kernel, collect the top 10%, 50% and 75% of the ranked list for each sub lexicon
Accumulate sub lexicons by taking the intersection of the 8 collected lists to obtain final bilingual lexicon
Random samples of extracted lexicon
Wikipedia # of words (OOV coverage)
samples
10%
11,970 (1.9%)
(natural selection, 自然選択説), (scrabble, スクラブル), (phase transition, 相転移), (diamond, ダイアモンド), (videocassette recorder, ビデオテープレコーダ)
50%
75.420 (7.7%)
(movement for multiparty democracy, 複数政党制民主主義運動), (fentanyl, フェンタニル) [an opioid analgestic], (sigma sagittarii, ヌンキ) [the second brightest star system in the constellation Sagittarius]),
(shintaro abe, 安倍晋太郎) [the former prime minister of Japan], (nippon television, 日本テレビ放送網)
75%
113,277 (11.5%)
(pride final conflict 2003, pride grandprix 2003 決勝戦) [a mixed martial arts event held by PRIDE Fighting Championships], (uglyness, 醜), (palma il vecchio, パルマ・イル・ヴェッキオ) [an Italian painter],
(jean gilles, ジャン・ジル) [a French composer; a French soldier], (amiloride, アミロライド) [a potassium-sparing diuretic]
100%
197,770 (13.5%)
(brilliant corners, ブリリアント・コーナーズ) [an album by a jazz musician], (charly mottet シャーリー・モテ) [a French former professional cyclist], (deep purple in rock, ディープ・パープル・イン・ロック)
[an album by an English rock band], (june 2003, 「最近の出来事」2003年6月) [navigational entry for events happened in June 2003], (moanin', モーニン) [a jazz album]
filtered
24.969
(1, 1年) [year], (UTC+9, UTC+9) [Japanese side contains only alphanumeric characters], (Aera, AERA) [case-insensitive match] (大岡越前, 大岡越前) [garbage in English side], (image:himeji castle
frontview.jpg, himeji castle frontview.jpg) [Wikipedia format navigational links], (user:eririnrinrin, eririnrinrin) [Wikipedia specific entries]
Performance of Wikipedia dict.
Single-ref
Fmlrun-int
JE
EJ
JE
EJ
Baseline (WMT08) 26.39
28.25
25.34
27.19
Wikipedia (10%)
N/A
27.47
N/A
N/A
Wikipedia (50%)
N/A
27.46
N/A
N/A
Wikipedia (75%)
N/A
27.42
N/A
N/A
Wikipedia (100%)
26.48
27.28
25.48
28.15
Dataset: NTCIR-7 Patent Translation Task (1.8 million parallel sentences)
Add the extracted bilingual lexicon to the training corpus to learn the
translation probability between translation pairs
The extracted bilingual lexicon slightly improves BLUE score
However, increasing the extracted lexicon constantly degrades BLEU score
for EJ on single-reference setting → Need re-examination
Conclusion and Future Work
Demonstrated that a large scale bilingual dictionary can be extracted from Wikipedia
Improved the quality of extracted bilingual lexicon by applying a graph kernel
Plan to integrate graph-based word sense disambiguation into statistical machine translation framework
Nara Institute of Science and Technology
Sample seeds
English
Japanese
Thermal spray
溶射
Epoxy
エポキシ樹脂
Single crystal
単結晶
Laser cooling
レーザー冷却
Centrifugal compressor
遠心圧縮機
ダウンロード

Semi-supervised Learning of Bilingual Lexicon for Technical Terms