Chenhui Chu , Toshiaki Nakazawa , Daisuke Kawahara, Sadao Kurohashi

1

Exploiting Shared Chinese Characters in Chinese Word Segmentation Optimization for Chinese-Japanese Machine Translation

Chenhui Chu, Toshiaki Nakazawa, Daisuke Kawahara, Sadao Kurohashi

Graduate School of Informatics, Kyoto University

EAMT2012 (2012/05/28)

Outline

• Word Segmentation Problems• Common Chinese Characters• Chinese Word Segmentation Optimization• Experiments• Discussion• Related Work• Conclusion and Future Work

2

Outline


3

Word Segmentation for Chinese-Japanese MT

4

Mr. Kosaka is the founder of The Japan Society for Clinical Anesthesiologists.

小坂 / 先生 / は / 日本 / 臨床 / 麻酔 / 学会 / の / 創始 / 者 / である / 。小 / 坂 / 先生 / 是 / 日本 / 临床 / 麻醉 / 学会 /的 / 创始者 / 。Zh:

Ja:

Ref:

小坂先生は日本臨床麻酔学会の創始者である。小坂先生是日本临床麻醉学会的创始者。Zh:

Ja:

Word Segmentation Problems in Chinese-Japanese MT

• Unknown words– Affect segmentation accuracy and consistency

• Word segmentation granularity– Affect word alignment

5

小坂 / 先生 / は / 日本 / 臨床 / 麻酔 / 学会 / の / 創始 / 者 / である / 。小 / 坂 / 先生 / 是 / 日本 / 临床 / 麻醉 / 学会 / 的 / 创始者 / 。Mr. Kosaka is the founder of The Japan Society for Clinical Anesthesiologists.

Zh:

Ja:

Ref:

Outline

• Introduction• Common Chinese Characters• Chinese Word Segmentation Optimization• Experiments• Discussion• Related Work• Conclusion and Future Work

6

Chinese Characters

• Chinese characters are used both in Chinese (Hanzi) and Japanese (Kanji)

• There are many common Chinese characters between Hanzi and Kanji

• We made a common Chinese characters mapping table for 6,355 JIS Kanji (Chu et al., 2012)

7

Common Chinese Characters Related Studies

• Automatic sentence alignment task (Tan et al., 1995)

• Dictionary construction (Goh et al., 2005)• Word level semantic relations investigation

(Huang et al., 2008)• Phrase alignment (Chu et al., 2011)• This study exploits common Chinese characters

in Chinese word segmentation optimization for MT

8

Outline


9

Reason for Chinese Word Segmentation Optimization

• Segmentation for Japanese is easier than Chinese, because Japanese uses Kana other than Chinese characters

• F-score for Japanese segmentation is nearly 99% (Kudo et al., 2004), while that for Chinese is still about 95% (Wang et al., 2011)

• Therefore, we only do word segmentation optimization for Chinese, and keep the Japanese segmentation results

10

11

Parallel Training Corpus

① Chinese Lexicons Extraction

Chinese Annotated Corpus for Chinese Segmenter

② Chinese Lexicons Incorporation

③ Short Unit Transformation

Optimized Chinese Segmenter

Training

Short Unit Chinese Corpus for Chinese Segmenter

Common Chinese Characters

Chinese Lexicons

System Dictionary of Chinese Segmenter

System Dictionary withChinese Lexicons

12







Training



Chinese Lexicons




• Step 1: Segment Chinese and Japanese sentences in the parallel training corpus

• Step 2: Convert Japanese Kanji tokens into Chinese using the mapping table we made (Chu et al., 2012)

• Step 3: Extract the converted tokens as Chinese lexicons if they exist in the corresponding Chinese sentence

13

Extraction Example

14

小坂 / 先生 / は / 日本 / 臨床 / 麻酔 / 学会 / の / 創始 / 者 / である / 。

小 / 坂 / 先生 / 是 / 日本 / 临床 / 麻醉 / 学会 / 的 / 创始者 / 。

Mr. Kosaka is the founder of The Japan Society for Clinical Anesthesiologists.

Zh:

Ja:

Ref:

小坂 / 先生 / は / 日本 / 临床 / 麻醉 / 学会 / の / 创始 / 者/ である / 。Ja:

Kanji tokens conversion

Check

小坂先生日本临床麻醉学会创始者 Chinese Lexicons :

Extraction

15







Training



Chinese Lexicons




• Using a system dictionary is helpful for Chinese word segmentation (Low et al., 2005; Wang et al., 2011)

• We incorporate the extracted lexicons into the system dictionary of a Chinese segmenter

• Assign POS tags by converting the POS tags assigned by the Japanese segmenter using POS tags mapping table between Chinese and Japanese

16

17

CTB (Penn Chinese Treebank)

JUMAN (Kurohashi et al., 1994)

AD(adverb) 副詞 (adverb)CC(coordinating conjunction) 接続詞 (conjunction)CD(cardinal number) 名詞 (noun)[ 数詞 (numeral noun)]FW(foreign words) 未定義語 (undefined word)[ アルファベット (alphabet)] IJ(interjection) 感動詞 (interjection)M(measure word) 接尾辞 (suffix)[ 名詞性名詞助数辞 (measure word suffix)] NN(common noun) 名詞 (noun)[ 普通名詞 (common noun)/ サ変名詞 (sahen

noun)/形式名詞 (formal noun)/ 副詞的名詞 (adverbial noun)], 接尾辞 (suffix)[ 名詞性名詞接尾辞 (noun suffix)/名詞性特殊接尾辞 (special noun suffix)]

NR(proper noun) 名詞 (noun)[ 固有名詞 (proper noun)/ 地名 (place name)/人名 (person name)/ 組織名 (organization name)]

NT(temporal noun) 名詞 (noun)[ 時相名詞 (temporal noun)]PU(punctuation) 特殊 (special word) VA(predicative adjective) 形容詞 (adjective)VV(other verb) 動詞 (verb)/ 名詞 (noun)[ サ変名詞 (sahen noun)]

18







Training



Chinese Lexicons




• Adjusting Chinese word segmentation to make tokens 1-to-1 mapping as many as possible between a parallel sentences can improve alignment accuracy (Bai et al., 2008)

• Wang et al. (2010) proposed a short unit standard for Chinese word segmentation, which can reduce the number of 1-to-n alignments and improve MT performance

19

Our Method

• We transform the annotated training data of Chinese segmenter utilizing the extracted lexicons

20

从 _P/ 有效性 _NN / 高 _VA/ 的 _DEC/ 格要素_NN /…

从 _P/ 有效 _NN/性 _NN /高 _VA/的 _DEC/ 格 _NN/要素_NN /…

Short:

CTB:

Lexicon: 有效 (effective) Lexicon : 要素 (element)

From case element with high effectiveness …Ref:

Constraints

• We do not use the extracted lexicons that are composed of only one Chinese character

21

歌 (song)・・・歌颂(praise)

歌 (song)/颂 (praise)

long token extracted lexicons short unit tokens

Outline


22

Two Kinds of Experiments

• Experiments on Moses • Experiments on Kyoto example-based

machine translation (EBMT) system Nakazawa and Kurohashi, 2011)– A dependency tree-based decoder

23

Experimental Settings on MOSES (1/2)

24

Parallel Training Corpus Zh-Ja paper abstract corpus (680k sentences)

Chinese Annotated Corpus NICT Chinese Treebank (same domain of parallel corpus, 9,792 sentences)CTB 7 (31,131 sentences)

Chinese Segmenter In-house Chinese segmenterJapanese Segmenter JUMAN (Kurohashi et al., 1994)MT system Moses with default options, except for

the distortion limit (6→20)

Experimental Settings on MOSES (2/2)

• Baseline: Only using the lexicons extracted from Chinese annotated corpus

• Incorporation: Incorporate the extracted Chinese lexicons

• Short unit: Incorporate the extracted Chinese lexicons and train the Chinese segmenter on the short unit training data

25

Results of Chinese-to-Japanese Translation Experiments on MOSES

BLEU NICT Chinese Treebank CTB 7Baseline 34.92 36.64Incorporation 36.19 37.39Short unit 36.82 38.59

26

• CTB 7 shows better performance because the size is more than 3 times of NICT Chinese Treebank

• Lexicons extracted from a paper abstract domain also work well on other domains (i.e. CTB 7)

Results of Japanese-to-Chinese Translation Experiments on MOSES

27

BLEU NICT Chinese Treebank CTB 7Baseline 31.42 31.83Incorporation 31.24 32.34Short unit 31.95 31.95

• Not significant compared to Zh-to-Ja, because our proposed approach does not change the segmentation results of input Japanese sentences

Experimental Settings on EBMT (1/2)

28

Parallel Training Corpus Zh-Ja paper abstract corpus (680k sentences)

Chinese Annotated Corpus CTB 7 (31,131 sentences)Chinese Segmenter In-house Chinese segmenterJapanese Segmenter JUMAN (Kurohashi et al., 1994)MT system Kyoto example-based machine

translation (EBMT) system

Experimental Settings on EBMT (2/2)

• Baseline: Only using the lexicons extracted from Chinese annotated corpus

• Short unit: Incorporate the extracted Chinese lexicons and train the Chinese segmenter on the short unit training data

29

Results of Translation Experiments on EBMT

30

BLEU Zh-to-Ja Ja-to-ZhBaseline 22.84 19.10Short unit 23.36 19.15

• Translation performance is worse than MOSES, because EBMT suffers from low accuracy of Chinese parser

• Improvement by short unit is not significant because the Chinese parser is not trained on short unit segmented training data

Outline


31

Short Unit Effectiveness on MOSES

32

Baseline (BLEU=49.38)Input: 本 / 论文 / 中 / ， / 提议 / 考虑 / 现存 / 实现 / 方式 / 的 / 功能 / 适应性 / 决定 / 对策 / 目标 / 的 / 保密 / 基本 / 设计法 / 。Output: 本 / 論文 / で / は / ， / 提案 / する / 適応 / 的 / 対策 / を / 決定 / する / セキュリティ / 基本 / 設計 / 法 / を / 考える / 現存 / の / 実現 / 方式 / の / 機能 / を / 目標 / と / して / いる / ．

Short unit (BLEU=56.33)Input: 本 / 论文 / 中 / ， / 提议 / 考虑 / 现存 / 实现 / 方式 / 的 / 功能 / 适应 / 性 / 决定 / 对策 / 目标 / 的 / 保密 / 基本 / 设计 / 法 / 。Output: 本 / 論文 / で / は / ， / 提案 / する / 考え / 現存 / の / 実現 / 方式 / の / 機能 / 的 / 適応 / 性 / を / 決定 / する / 対策 / 目標 / の / セキュリティ / 基本 / 設計 / 法 / を / 提案 / する / ．

Reference本 / 論文 / で / は / ， / 対策 / 目標 / を / 現存 / の / 実現 / 方式 / の / 機能 / 的 / 適合 / 性 / も / 考慮 / して / 決定 / する / セキュリティ / 基本 / 設計 / 法 / を / 提案 / する / ．(In this paper, we propose a basic security design method also consider functional suitability of the existing implementation method for determining countermeasures target.)

Number of Extracted Lexicons

33

Source NICT Treebank CTB 7Chinese annotated corpus 13,471 26,202Short unit Chinese corpus 12,627 25,490

SourceParallel training corpus 18,584

• The number of extracted lexicons deceased after short unit transformation because duplicated lexicons increased

Short Unit Transformation Percentage

• NICT Chinese Treebank– 6,623 tokens out of 257,825 been transformed to

13,469 short unit tokens, the percentage is 2.57% • CTB 7– 19,983 tokens out of 718,716 been transformed to

41,336 short unit tokens, the percentage is 2.78%

34

Short Unit Transformation Problems (1/3)

• Improper transformation problem

35

好意 (favor)・・・不好意思(sorry)

不 (not)/好意 (favor)/思 (think)



36

充电(charge)电器(electric

equipment)・・・

充电器 (charger) 充 (charge)/电器 (electric equipment)

充电 (charge)/器 (device)


• Transformation ambiguity problem


37

实验 (test)・・・被实验者 _NN(test subject)

被 _NN(be)/实验 _NN(test)/者 _NN(person)


The correct POS tag for “ 被 (be)” should be LB (“ 被” in long bei-const)

• POS Tag assignment problem

Outline


38

Bai et al., 2008

• Proposed a method of learning affix rules from a aligned Chinese-English bilingual terminology bank to adjust Chinese word segmentation in the parallel corpus directly

39

Wang et al., 2010

• Proposed a method based on transfer rules and a transfer database. – The transfer rules are extracted from alignment

results of annotated Chinese and segmented Japanese training data

– The transfer database is constructed using external lexicons, and is manually modified

40

Conclusion

• We proposed an approach of exploiting common Chinese characters in Chinese word segmentation optimization for Chinese-Japanese MT

• Experimental results of Chinese-Japanese MT on a phrase-based SMT system indicated that our approach can improve MT performance significantly

41

Future Work

• Solve the short unit transformation problems• Evaluate the proposed approach on parallel

corpus of other domains

42

Chenhui Chu , Toshiaki Nakazawa , Daisuke Kawahara, Sadao Kurohashi

Documents

Transcript of Chenhui Chu , Toshiaki Nakazawa , Daisuke Kawahara, Sadao Kurohashi