12.07.2015 Views

file - ChaSen - 奈良先端科学技術大学院大学

file - ChaSen - 奈良先端科学技術大学院大学

file - ChaSen - 奈良先端科学技術大学院大学

SHOW MORE
SHOW LESS
  • No tags were found...

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

R = |Rc||Ra|(4.10)Varying combination of part-of-speeches, the feature set mainly consistingof functional words shows good result, 0.83 in F-measure when removing noun,collocation, adjective, nominal adjective, and verb from features.Seeing withrespect to each the description types, in same combination of POS above, theProcess and the Order of time achieved 0.88 and 0.76 respectively. The Definitiontype resulted in 0.85 when additionally removing auxiliaries.For rules for categorization, the highest performance of F-measure is 0.79 resultedin the combination of word rules and phrase rules. For each the Descriptiontypes, 0.87 for the Process type and 0.71 for the Definition using word and phraserules, 0.72 for order of time only using phrase rule.When only using sentence rules, the accuracies of categorizations for the Processtype remain in high level, more than 0.6, however that for the Definition andthe Order of time were declined drastically.For observations of acquired rules inthese experiments, for the Process type, combinations of words such as the particlesuch as “ (node)” describing reason and the auxiliary such as “(te kudasai)” descriving requests, expressions at terminals of clauses or sentencesoften appeared. For the Definition types, a particle of topic marker such as “(ha),” brackets and blank characters were obtained. For the Order-of-time, endingsof conjugation marking passed tense such as “ (ta),” conjunctive particlesuch as “ (toiu),” and conjunctions such as “ (shikashi).”4.5 Discussion and concluding remarksProposed question-answering system in this thesis is based on answer extractionsusing their description types, therefore the scope of applicaation is restricted toanswers that are preferentially used certain description types such that identifiedin surface features. For instance, it is difficult to learn from annotated corpus indescription types in that inter-annotators agreements were low, such as Analysis,Fact, Instance and Cause-Result. When answers to a question appear with variousdescription types in source documents, such as free-formatted essay, proposedapproach should be not work effectively to such question. On the other hand,63

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!