Presentation summarization, alignment and generation

1
Presentation summarization, alignment and
generation
2
Re-cap

• Automatic Alignment
• Slide 1 --gt Abstract
• Slide 2 --gt Intro
• Slide 3 --gt some other para

3
Re-cap

• Generation
• Using the model ? from alignment, generate a
  slide show.
• ? will tell which section is important.
• Summarize the section into a slide.
• (Dual of alignment)
• Slide 1 lt-- Abstract
• Slide 2 lt-- Intro
• Slide 3 lt-- some other para

4
Baseline Method

• Cosine similarity
• Treat a slide as a query, and a para as a doc.
• Return the para with the highest cos measure

5
Evaluation Results

• Binary Evaluation (1-0)
• Assign 1 if slide and para related or slide is
  not in paper e.g Thank You slide, Content slide.
• 0 otherwise.
• Recall, precision F1 is 0.7 over a subset of
  the corpus of 10 ltslide, papergt pairs. Std Dev is
  0.2.

6
Weakness

• Equal weight of terms
• Sparse text slide shows dont perform well
• Not able to identify which slide is not in paper
• No learning involved

7
Gene Sequence Alignment

• Similar to Edit Distance
• X GATTACA
• Y GAGA
• GA- -A- -
• GA- A
• S1 , S2.
• P1 , P2..

8
Pair HMM (3 state)
M match X slide Y para e affine penalty
(lower penalty for further deletion) d penalty
9
Pair HMM (5 states)
Full probabilistic version of 3 state
10
Profile HMM

• Simply multiple instances of Pair HMM