Elamur Research⌕
☰
home / papers / 2023 / 基于依存句法的中文语义模型及语义提取方法

基于依存句法的中文语义模型及语义提取方法

Summary

Chinese semantic extraction usually relies on intent recognition and slot filling. These approaches need large annotated training sets, and the resulting models transfer poorly to new tasks. The authors propose a rule-driven alternative built on dependency parses, using word segmentation, POS tags and dependency arcs produced by the LTP toolkit. They define four semantic representations for different extraction scenarios: a key-value model, a slot model, a tree model and a path model. These are paired with three extraction procedures:

  • keyword-driven inspection of the parse tree,
  • recursive subject–verb–object decomposition,
  • matching of user-written templates against root-to-leaf dependency paths.

The evaluation uses 5,000 manually labelled BCC corpus sentences built around five Chinese interrogative pronouns. On a small, imbalanced 1,000-sentence subset for 多少, the template-based method surpasses seven deep-learning text classifiers once more than about 25 templates are configured. The work matters for low-resource settings because its rules need little labelled data and can be reused across domains. The trade-off is manual linguistic effort to write those rules.

Key points

  • Four models and three algorithms.
    • Models: key-value, slot ($+POS placeholders), tree (five clause patterns plus three modifier relations), and path (templates over dependency paths).
    • Algorithms: keyword-based, tree-based and extensible template-based extraction.
  • Parser and knowledge base.
    • Parsing relies on LTP with HIT/CDT-style relations (SBV, VOB, ATT, POB, etc.), not UD. The paper cites 89.5% parsing accuracy on the Chinese Dependency Treebank (CDT).
    • A hand-built knowledge base maps 2,513 common words to "semantic ontology" entries with importance and frequency attributes.
  • Dataset. The BCC corpus (Beijing Language and Culture University) supplied 1,000 sentences each for 怎么, 什么, 哪, 多少 and 谁. Senses were manually annotated; 怎么 was split into six sense classes, including combined ones.
  • Effect of template count. Accuracy of the template algorithm rises with the number of templates and flattens beyond roughly 20, by which point about 90% of sentences are handled. The remaining errors are mostly edge cases.
  • Comparison with deep-learning classifiers.
    • Setup: the 多少 subset (six classes, 3:1:1 train/dev/test split) against TextCNN, TextRNN, FastText, TextRCNN, TextRNN_Att, DPCNN and Transformer.
    • Baselines: the best was TextRNN at 86.72%, most others reached 85.94%, and FastText reached 85.16%.
    • Result: the rule-based method exceeds them with more than 25 templates.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Linguistics & languages · Natural language processing

Cite

The paper itself, not this page. BibTeX from the library's record of the work.

@article{2023,
  title = {{基于依存句法的中文语义模型及语义提取方法}},
  author = {王佳琦 and 韩军 and 孙启童},
  journal = {中文信息学报 (Journal of Chinese Information Processing)},
  year = {2023},
  url = {http://jcip.cipsc.org.cn/article/id/zwxxxb_3607}
}

References

13 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. [ 1] 刘群,詹卫东,常宝宝,等 . 一个汉英机器翻译系统的 //智 能 计 算 机 接 口 与 应 用 进 计算模型与语言模型[ C] 展:第 二 届 中 国 计 算 机 智 能 接 口 与 智 能 应 用 学 术 会 议 论文集 . 北京:电子工业出版社,1997:253 - 258. find
  2. [ 2] FINKSR.Aspe c t so fapedagog i c a lg r amma rba s edon c a s eg r amma randva l enc et he o r emeye r y[M].Max Ni Ve r l ag,2011. find
  3. [ 现代汉语配价语法研究综述[ 今 古 文 创, 3] 徐凯 . J]. 2021, 1( 31):108 - 109. find
  4. [ 4] HARRI S M.Somep r ob l emsf o rac a s eg r amma ro f La t i nande a r l e[ J].J ou r na lo fL i ngu i s t i c s, y Romanc 1975,11( 2):183 - 194. find
  5. [ 5] BACKLEYP. I n t r oduc t i ont oe l emen tt he o r - y[M].Ed i nbu r i ve r s i t e s s,2011. ghUn yPr [ 6] NIVERJ.Dependencypa r s i ng[ J].LanguageandL i n - i s t i c sCompa s s,2010,4( 3):138 - 152. gu 王佳琦( 1996—), 硕 士 研 究 生, 主 要 研 究 领 域 为 自然语言处理。 E-ma i l:wang a. edu. cn j q@bua 孙启童( 1994—… find
  6. [ 8] LIU T,MAJ,LIS.Bu i l d i ngadependencyt r e ebank f o rimp r ov i ng Ch i ne s epa r s e r[ J].J ou r na lo fCh i ne s e LanguageandCompu t i ng,2006,16( 4):207 - 224 find
  7. [ 9] BLOOMFIELD L.L i ngu i s t i c sandr e ad i ng[ J].Lan - , , ( ) : a r n i ng 1955 5 3 - 4 94 - 107. guageLe [ 10] 朱德熙 . 现代 汉 语 形 容 词 研 究 [ J]. 语 言 研 究,1956, 1( 1):1 - 37. find
  8. [ 11] 邵敬敏,周芍 . 语义特征的界定与提取方法[ J]. 外 语 教学与研究( 外国语文双月刊),2005,37( 1):21 - 28. find
  9. [ 12] 荀恩东,饶高琦,肖晓悦,等 . 大数据背景下 BCC 语 料库的研制[ J]. 语料库语言学,2016,3( 1):93 - 118. find
  10. [ 13] CHEN Y.Convo l u t i ona lneu r a lne two r kf o rs en t enc e c l a s s i f i c a t i on[ D].Un i ve r s i t fWa t e r l oo,2015. yo [ 14] LIU P,QIU X,HUANG X.Re cu r r en tneu r a lne t - wo r kf o rt ex tc l a s s i f i c a t i on wi t h mu l t i - t a skl e a r n i ng [ //Pr C] o c e ed i… find
  11. [ 16] LAIS,XU L,LIU K,e ta l.Re cu r r en tc onvo l u t i ona l //Pr neu r a lne two r ksf o rt ex tc l a s s i f i c a t i on[ C] o c e ed - i ngso ft he29 t hAAAICon f e r enc eonAr t i f i c i a lI n t e l - l i e,2015:2267 - 2273. genc [ 17] JOULIN A,GRAVE É,BOJANOWSKIP,e ta l. //Pr Bago ft r i… find
  12. [ 19] VASWANIA,SHAZEER N,PARMAR N,e ta l. //Advanc At t en t i oni sa l lyoune ed[ C] e si nNeu r a lI n - f o rma t i onPr o c e s s i ngSy s t ems,2017:5998 - 6008. find
  13. [ 20] PULVERMÜLLER F.Neu r ob i o l og i c a l me chan i sms f o rs eman t i cf e a t u r eex t r a c t i onandc onc ep t ua lf l ex i - b i l i t J].Top i c si nCogn i t i veSc i enc e,2018,10( 3): y[ 590 - 620. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wang-2023-chinese-semantic-model/