|1||April 10||Introduction to this lecture.
Tagging with HMM. slides slides
- *install Python into your laptop computer.
- *learn basics of Python if you are a novice.
- read a note on HMM by Michael Collins.
- implement your HMM-based POS tagger.
|2||April 17||Text classification with naive bayes classifiers slides||
- read "a comparison of event models for naive bayes text
classification" by McCallum and Nigam.
- install MeCab, and try to find sentences that MeCab cannot analyze correctly
|3||April 24||The method of Lagrange multipliers.
Maximum likelihood estimation.
Maximum a posteriori estimation.slides
- read Sections 1 and 2 of the following tutorial on Lagrange multipliers.
tutorial by Dan Klein.
Try to give an intuitive explanation of this method when the solution space is 3-dimensional.
- implement a naive bayes classifier. Train it on this file, and test it on this file.
Each line of these files consists of a class (+1 or -1) and a segmented sentence.
Maximum likelihood estimation.
Maximum a posteriori estimation.
bag-of-words representation of document.
- MAP estimation of multinomial model of naive bayes classifiers
- derive the dual problem of the optimization problem of SVM with soft-margin
- Use an SVM tool (e.g., TinySVM) to train a model on this file, and test it on this file. You need to write a script that converts those files into input format of the tool.
- Read Section 2.1 of this tutorial.
|--||May 8||NO LECTURE|
- read Section 3 and Section 5.1 of the "CaboCha" paper":
"Japanese Dependency Analysis using Cascaded Chunking", CoNLL 2002.
and answer the following questions:
-- which static features are used?
-- which dynamic features are used?
-- are dynamic features effective? If so, in what situation?
-- which kernel function is used?
-- what benefit does the use of the kernel function above have? (not written in the paper. Think for yourself)
|6||May 22||Log-linear Model
Conditional Random Fields (CRF)
- read Sections 1, 2, and 3 of the tutorial on CRF
- read Section 6.3 of a book (in Japanese) to review CRF and try to understand the forward-backward algorithm.
|7||May 29||Forward-backward algorithm
|Read the following paper and learn how the weights on words are calculated in their work:
Yih et al., 2007
|8||June 5||text summarization slides||take a rest|
|9||June 12||k-means clustering, EM, PLSI slides||
- derive the update equations for the product model.
- Answer the following questions with the reference to Hofmann's paper.
* how is ``document'' integrated into the model?
* what is the tempered EM? What is the update equation for PLSI when the tempered EM is used?
* what is the folding-in? What kind of calculation is needed for the folding-in?
- Implement PLSI, and train it on this file, and calculate the perplexity of this file.
|implement Gibbs Sampling for LDA.
Train it on this file. Each line of this file corresponds to a document, which is represented as a set of nouns, verbs, adverbs, and adjectives that appear in the document.
|--||June 26||NO LECTURE|
Check LDA code.
|No assignment. But see the slides for details on the report submission (GRADING 1).|
Derivation of update equations for LDA's Gibbs Sampling.
slides, survey by Kaji-san
Watch this video (10 minutes).
GRADING 2: Read the submission, write the review form, and send it to me by July 23rd?
|14||July 24||NO LECTURE|