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...

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

However when using IOB2/IOE2/IOBES, the performance of labeling the sentencein the inside of a chunk contrarily was declined. The number of this kindof chunks is few in our corpus, the positive examples of this case for machinelearning are considered to be insufficient.There are seventeen question segments comprising multiple sentences in testdataset. The sentence representing question or request appears at the head ofsegment in one case, at the tail of segment in nine cases and at both the headand tail of segment in six cases. One case has no sentence representing questionor request. Those question segments failed to identify the question types. Inevaluation of labeling to sentence, the best result was obtained in IOE1 labelingsuch that four sentences were correctly labeled at 34 heads and tails of sentencesof 17 question segments.This thesis proposed the chunking-based question analysis that performedconcurrently both question segmentation and question type identification, whichaimed at concurrently solving two problems in question analysis. The first problemwas a methodology that can handle more complex queries that comprisemultiple questions or question described by multiple sentences, and the secondproblem is to reduce the computational cost of previous techniques. Proposedmethods can solve these problems in theory, however the accuracies in experimentalresults have not achieved to the practical level yet.The experimental results show the opposite natures to same features in questionsegmentation and question type identification. In general, it should be difficultto reveal such two alien problems in a same computational model. Proposedmethod has not been considered in this aspect of problem. Concurrent processingof question segmentation and question type identification is effective in reductionof computational cost, that however was clarified that does not fit the conditioninvolved different properties of question segmentation and the type identification.Therefore, I am going to change the strategy to that exploiting different modelsfor question segmentation and question type identification in next step, andattempt to reduce the computational cost in such frame work.Another important observation in experimental result is that many errors ofquestion segmentation and type identification occurred in sentences comprisingmany ellipses. That process that identify ellipsis and complete it by any relevant43

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

Saved successfully!

Ooh no, something went wrong!