1 / 18

A High Performance Semi-Supervised Learning Method for Text Chunking

A High Performance Semi-Supervised Learning Method for Text Chunking. Authors: Rie Kubota Ando Tong Zhang. S tructural learning. Idea:

osma
Download Presentation

A High Performance Semi-Supervised Learning Method for Text Chunking

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. A High Performance Semi-Supervised Learning Method for Text Chunking Authors: RieKubota Ando Tong Zhang

  2. Structural learning. • Idea: • “What good classifiers are like” by learning from thousands of automatically generated auxiliary classification problems on unlabeled data. By doing so, the common predictive structure shared by the multiple classification problems can be discovered • Performance better than previous results

  3. Boot Strapping Methods: • Co-Training • Expectation Maximization • Goal: • Create Learning Framework

  4. Contribution of paper: • Design Novel Robust semi-supervised Method • Reporting higher performance

  5. Standard Linear Prediction Model: • f(x)=wpow(T) x • w-> weight vector • K-way Classification: • Winner takes all • One predictor per class

  6. Linear model for Structural Learning: f`((-),x)=wT`x + vT` (-)x , (-)(-)T=I (-) -> projection matrix I-> identity Matrix

  7. Alternating Structure Optimization: • Fix ((-),{v`}), and find m predictors. • Fix m predictors {u`}and find ((-),{v`} ). • Iterate until a convergence criterion is met

  8. Properties of Auxiliary Problem: • Automatic labeling • Relevancy

  9. Semi-Supervised Learning Procedure: • Create training data Z~` for each l. • Compute (-) from training data through SVD-ASO. • Minimize the empirical risk on the labeled data

  10. Auxiliary Problem Creation: • Unsupervised Strategy • Predict words • Partially-Supervised Strategy • Predict top k-choices of the classifier

  11. Extension of the SVD-ASO Algorithm: • NLP applications has natural grouping • Perform localised optimization • Sub matrxix of structure matrix (-) • Regularise the non negative components

  12. Baseline Algorithms: • Supervised classifier • Co-Training • Self-Training

  13. Named Entity Chunking Experiment:

  14. Results: • Refer Page 6 in pdf • Refer page 7 in pdf

  15. Syntactic Chunking Experiment: • Refer page 7 in pdf • Refer page 8 in pdf

  16. Conclusion: • Presented a novel semi-supervised method • Predictive low dimensional feature projection • Key is to create auxiliary problems automatically. • Risk is low and has potential gain

  17. Queries???

  18. Created a framework for carrying out possible new ideas By designing a variety of auxiliary problems SVD more info: http://en.wikipedia.org/wiki/Singular_value_decomposition ERM more info: http://en.wikipedia.org/wiki/Empirical_risk_minimization

More Related