SRL and dependency tree english version
- 2. • Introduce • Three Linguistics issues • Syntactic & Semantics • Constraint & Dependency • Tools • Samples • Appendix
- 3. Three issues Linguistics Issues: Sound, Syntactic and Semantics *the most issue we want to solve: Semantics
- 4. Syntactic & Semantics • example: 『我們打敗敵人』(we have defeated the enemy) 和『敵人被我們打敗』 (The enemy have been defeated by us) • different Syntactic • same Semantics: defeat(我們、敵人)
- 5. Syntactic & Semantics • Syntactic • the way words put together • tools: POS tagging (Part-of-Speech tagging) • Semantics • meaning of text • analysis of word-relation : Dependency, Constraint • Constraint including: SRL & SDP
- 6. Constraint & Dependency • Constraint • Hierarchical format (tree) • Dependency • Sequencial format
- 7. Tools • POS tools • Stanford corenlp • (+Semantics) CKIP Parser
- 8. Tools • Constraint • CKIP Parser • S(theme:NP(Head:Nb:柯文哲)|theme:PP(Head:P21:在|DUMMY:NP(property:VP‧的 (head:VP(time:Ndabd:國慶日|theme:VP(Head:VC2:穿|aspect:Di:了 |goal:NP(quantifier:DM:一件|Head:Ncc:胸口))|Head:V_2:有|range:NP(property:Nab: 國旗|Head:Nac:圖案))|Head:DE:的)|Head:Nb:Ｔ))|Head:VJ1:恤)#， • relation of word should be decided by tagging and tree structure • Dependency • Semantics Role Labeling (SRL) • only Semantics Role tagging • example：魯迅寫的故鄉是篇好文章。 魯迅/Agent • Semantics Dependency Parsing(SDP) • not only Semantics tagging but relation • example：寫/re-ContentProduct->故鄉
- 10. 我们_0 打败_1 Agt 打败_1 -1 Root 敌人_2 打败_1 Pat 我们打败敌人 we have defeated the enemy 敌人_0 打败_3 Pat 被_1 我们_2 mPrep 我们_2 打败_3 Agt 打败_3 -1 Root 敌人被我们打败 Enemy have been defeated by us same
- 11. 游客_0 破坏_1 Agt 破坏_1 -1 Root 国家_2 公园_3 Poss 公园_3 破坏_1 Pat 游客破坏国家公园 Tourist damage National Park 国家_0 公园_1 Poss 公园_1 遭到_2 Exp 遭到_2 -1 Root 游客_3 破坏_4 Agt 破坏_4 遭到_2 dCont 国家公园遭到游客破坏 National Park is been damaged by tourist
- 12. Compare with SDP & Constraint
- 13. 小影_0 感到_3 Aft 为_1 你_2 mPrep 你_2 感到_3 Datv 感到_3 -1 Root 难过_4 感到_3 Cont 小影为你感到难过 I feel sad for you 小影_0 感到_3 Aft 在_1 美国_2 mPrep 美国_2 感到_3 Loc 感到_3 -1 Root 难过_4 感到_3 Cont 小影在美國感到难过 I feel sad in US 小影為你感到難過 I feel sad for you 小影在美國感到難過 I feel sad in US
- 15. 我_0 想_1 Aft 想_1 -1 Root 她_2 想_1 Datv 。_3 想_1 mPunc 忘_0 -1 Root 不_1 了_2 mNeg 了_2 忘_0 mTime 她_3 忘_0 Datv 。_4 忘_0 mPunc 我想她。忘不了她。 I miss her. Can’t get over her. <coreference> <coreference> <mention representative="true"> <sentence>1</sentence> <start>3</start> <end>4</end> <head>3</head> <text>她</text> </mention> <mention> <sentence>2</sentence> <start>4</start> <end>5</end> <head>4</head> <text>她</text> </mention> </coreference> </coreference> Co-reference
- 16. Appendix
- 17. Appendix I • Method of Semantics Parsing • word + argument 例如： 『認同』有客體(THEME)、目標(Goal)兩個論元， 國人 很認同 傳統家電 ｜ ｜ THEME GOAL | | 國人 對傳統家電的 認同已經相當深 『違規』有主事者(AGENT)這個論元，在述詞前接一個主事者，語意即完整。 句型：AGENT <*
- 18. Appendix II • 語在中文文字使用的時候，母語人士習慣『會意』 ，而英文使用時，母語人士習慣用『形意』。 (Chinese native speaker used to catch the meaning of text, however, English native speaker used to catch the structure of text.) • 可見英文是比較偏重句法結構的語言，而中文則比較 偏重意會理解。 (Chinese focus on the meaning, but English one the structure. )
- 19. Special Thanks Mei-Yu for concept clarification :)