Semantic Role Chunking Combining Complementary Syntactic Views Sameer

Зарегистрируйтесь, чтобы просмотреть полный документ!

Semantic Role Chunking Combining Complementary Syntactic Views Sameer Pradhan, Kadri Hacioglu, Wayne Ward, James H. Martin, Daniel Jurafsky Center for Spoken Language Research Department of Computer Science University of Colorado at Boulder Department of Linguistics Stanford University

Different Syntactic Views n n Hypothesis: Different views make different errors Two views: n n Phrase structure based (Charniak, Collins) Chunk based

Constituent Views Constituents from Charniak parse tree Constituents from Collins parse tree Charniak Parse Tree Collins Parse Tree John kicked the ball.

Chunk View n n [Hacioglu & Ward 2003] Chunk using an IOB representation [Ramshaw & Marcus, 1995] Yamcha [Kudo & Matsumoto, 2001] Salomon will buy sufficient shares to cover its entire position O O O B-A 2 I-A 2 O B-V B-A 1 I-A 1 n n n Bottom up as opposed to top down Flat representation Uses flat syntactic chunks

Algorithm n n n Generate Charniak and Collins parse based features Add few features from one to the other Generate semantic IOB tags using these views Use them as features Generate the final semantic role label set using a phrase-based chunking paradigm

Architecture Charniak Collins Words Phrases IOB IOB Chunker IOB Semantic Role Labels Features

Illustration 1 2 R 1 2 B O O B B B I O O O I I I I O O O B B B I I O O B I I I B B I I I O B B O O O I I I I I Train Model H Classifier B Model O O B O B I O O B I I I

Features n n Semantic IOB tags for Charniak and Collins based semantic role labels [Pradhan et al. , 2005] Phrase level chunk features [Hacioglu et al. , 2004]

Active Learning n n n Randomly selelected 10 k examples and trained a NULL vs ARGUMENT classifier Classified remaining examples using this classifier Added misclassified examples to the seed set Iterated Final data amounted to about a third of the total

Combination Results ID + Class System ASSERTCharniak ASSERTCollins ASSERTCombined Train : Sections 02 -21 of Prop. Bank Test : Section 24 of Prop. Bank P R F 1 80 75 77 79 74 76 81 76 78

Results ID + Class Section 24 System Submitted System Bug fixed System P R F 1 80. 9 75. 4 78. 0 81. 9 75. 1 78. 3 Section 23 P R F 1 81. 9 73. 3 77. 4 82. 9 74. 7 78. 6 Brown P R F 1 73. 7 61. 5 67. 1 74. 5 63. 3 68. 4

Thank You Arda AQUAINT program contract OCG 4423 B NSF grant IS-9978025

Software n ASSERT (Automatic Statistical SEmantic Role Tagger) n n Publicly downloadable at http: //oak. colorado. edu/assert Downloaded by more than 50 research groups

Null Filtering n n Removed constituents with P(NULL) > 0. 9 Removed phrases with P(NULL) > 0. 8 after incorporating context

Analysis n n n Active learning using confidence threshold Constituent level instead of Sentence level N-Best Charniak parses

Features (Constituent)

Features (Phrase)

Representation

Features

Minipar-based Semantic Labeling n Rule-based dependency parser But analysts reckon underlying support for sterling has been eroded by the chancellor 's failure to announce any new policy measures in his Mansion House speech last Thursday

Скачать презентацию Semantic Role Chunking Combining Complementary Syntactic Views Sameer

b2fe836a9b30c2c0f4d5227ac3f71ae5.ppt

Количество слайдов: 23