Faisal ahmed talk w4a2012

23
Why Read if You Can Skim: Towards Enabling Faster Screen Reading Faisal Ahmed, Yevgen Borodin, Yury Puzis, I.V. Ramakrishnan Stony Brook University, USA. 2012 1

Transcript of Faisal ahmed talk w4a2012

Page 1: Faisal ahmed talk w4a2012

1

Why Read if You Can Skim: Towards Enabling Faster Screen Reading

Faisal Ahmed, Yevgen Borodin, Yury Puzis, I.V. RamakrishnanStony Brook University, USA. 2012

Page 2: Faisal ahmed talk w4a2012

2

Motivation

• Screen Reader – No Support • Ad hoc techniques – Inefficient

Sighted people: Skimming

Sift quicklyBasic meaning

Page 3: Faisal ahmed talk w4a2012

3

Contribution

• Identifying the types of skimming

• Proposing a usable interface

• Identifying Use scenarios

• Providing Insight: automated skimming

Page 4: Faisal ahmed talk w4a2012

4

Skimming in Accessibility • JAWS – 1st sentence/line of each paragraph– Rules (regular expressions) for skimming

• VoiceOver– Drag a finger & listen

• Ad hoc– HighSpeechRate/NavigationShortcuts/FastKeyPress– Often read entire content

Page 5: Faisal ahmed talk w4a2012

5

Summarization

• Existing summarization tools–Provide summaries, Preview Original–No option for• Skim/switch – Summary <=> Original

Page 6: Faisal ahmed talk w4a2012

6

Approach• Generated gold-standard summaries– Experiment – 12 sighted, 6 articles

• Developed Screen-reading interface – Switch seamlessly article <=> summary.

• Conducted User Study - 20 blind subjects –– Listening-and-comprehension – Search scenario

Page 7: Faisal ahmed talk w4a2012

8

Generating Gold Standard Summaries

Loose Guidelines• Summarize each sentence separately • Include words and phrases from original only• Preserve original word order• Keep it no more than 1/3 of the original article • Make it as informative as possible

Page 8: Faisal ahmed talk w4a2012

9

Analysis: Gold Standard Summaries

• Prominent POS• Punctuation preserved

noun

verb

adj

adv

preposition

other

Content Types

Description

C Gold Summaries (Combined 4 versions using voting)B All nouns along with prepositions of C

A All nouns of C D Original Article

Page 9: Faisal ahmed talk w4a2012

10

Example: SummariesD: Original paragraph (Comp. Ratio 100%)Twitter, which was created by a 10 person startup in San Francisco was called Obvious. It is a heady mixture of messaging, social networking, ’microblogging’ and something called ’presence.’ It's shorthand for the idea that people should enjoy an ’always on’ virtual omnipresence. Twitter's rapid growth made it the object of intense interest. The object of fair amount of ridicule, as it was derided as high tech trivia or the latest in time-wasting devices. But its use in Iran in the wake of the disputed presidential election of June 2009 brought it new respect. It was used to organize protests and disseminate information in the face of a news media crackdown.C: A Gold-standard summary (Comp. Ratio 25%)Twitter, 10 person startup San Francisco, called Obvious. mixture of social networking microblogging. virtual omnipresence. rapid growth. time-wasting devices. use in Iran disputed election. used to organize protests.

Page 10: Faisal ahmed talk w4a2012

11

Example: Summaries

C: A Gold-standard summary (Comp. Ratio 25%)Twitter, 10 person startup San Francisco, called Obvious. mixture of social networking microblogging. virtual omnipresence. rapid growth. time-wasting devices. use in Iran disputed election. used to organize protests.B: Gold Summary with nouns and prep only (Comp. Ratio 17%)Twitter, 10 person startup San Francisco, obvious. mixture of networking microblogging. omnipresence. growth. devices. in Iran election. to protests.

Page 11: Faisal ahmed talk w4a2012

12

Example: Summaries

B: Gold Summary with nouns and prep only (Comp. Ratio 17%)Twitter, 10 person startup San Francisco, obvious. mixture of networking microblogging. omnipresence. growth. devices. in Iran election. to protests. A: Gold summary with nouns only (Comp. Ratio 14%)Twitter, 10 person startup San Francisco, obvious. mixture networking microblogging. omnipresence. growth. devices. iran election. protests.

Page 12: Faisal ahmed talk w4a2012

14

Scenario: Listening & Comprehension• Within subjects• Sample task to Practice• Set of 10 questions (i.e., General, POS,

Numerals)• Pilot - 1st sentence summary (JAWS) – unuseful

Measure:• # of correct answers

Page 13: Faisal ahmed talk w4a2012

15

Scenario: Ad-Hoc Searching

• Within subjects• Sample task to Practice• Questions contain different wording

Find answer using• “Accessible Skimming” VS “ad-hoc techniques”

Measures:• Time (reach/perceive), #Key Presses

Page 14: Faisal ahmed talk w4a2012

16

Interface

• Standard Screen Reader• Paragraph/Sentence/Word/Character navigation• Shortcut for toggling Skimming– Switch between summary and original– Preserve current reading position

Page 15: Faisal ahmed talk w4a2012

17

Primary Hypotheses

• A:“skimming is significantly faster than regular screen reading”

• B: “user’s comprehension does not fall off with skimming compared to regular screen reading.”

Page 16: Faisal ahmed talk w4a2012

18

Results

• Statistical significance (t-test and ANOVA) tests provided strong support of our two primary hypotheses

Page 17: Faisal ahmed talk w4a2012

19

After reading each of the summaries A, B, C, and D subjects can answer increasingly

more questions correctly.

A B C D0

102030405060708090

100

0102030405060708090100

Average Correctness Compression Ratio

Skimming using Different types of contents

% A

vera

ge C

orre

ctne

ss

% C

ompr

essi

on R

atio

(13.86) (16.73)

(14.31)(11.64)

Page 18: Faisal ahmed talk w4a2012

21

Finding information using gold summaries is faster than in ad hoc searching.

time for TTS time for subject0

20

40

60

80

100

120

140

160

180

Skimming Ad-hoc Techniques

Tim

e(se

cs)

(24.89)

(45.16)(49.54)

(37.3)

Page 19: Faisal ahmed talk w4a2012

22

Questionnaires

General Statements Avg (St Dev)I wish I could look through articles faster, than I can with a screen reader

3.80 (1.15)

I often experience difficulties looking for desired information within an article

3.40 (0.99)

Skimming made reading through articles faster 4.20 (1.00)I was able to understand articles equally well when skimming and when reading full article

2.70 (1.34)

I want to use the skimming feature in the future 4.30 (1.03)

5 point Likert Scale Values (St. Dev)[1=strongly disagree, 5=strongly agree]

Page 20: Faisal ahmed talk w4a2012

23

Summarization For Skimming

Literature review

examination of visual speed-

reading technique

experiments with the gold-standard

summaries

Main Characteristics:•Nouns, Prepositions, Verbs, Adj/Adv•Extractive Summarization•Extract words and word combinations •Summarize each sentence separately•Preserve original order•Preserve punctuations

Ideal Technique: Semantic and/or Syntactic analysisExtract salient parts

Page 21: Faisal ahmed talk w4a2012

24

Testimonials

Good• “If I don't know what I am looking for, this is definitely very

handy. I can decide whether to read further or not.”• “I think skimming would be really helpful for long articles,

books. In the skimming mode, it was like I was going over my class notes!”

Bad• “Skimming is faster, but important info is sometimes

missing.”• “Skimming breaks up information… Introduces

disorganization... There were no reference points… I don't know where I am...”

Page 22: Faisal ahmed talk w4a2012

25

Future Research

• Design of an appropriate summarization

• Skim other types of content (i.e., Images, Buttons, lists)

• Variable-speed skimming

This material is based upon work supported by the National Science Foundation - Awards: IIS-0808678 and CNS-0751083

Page 23: Faisal ahmed talk w4a2012

26

Questions ?Comments?

Suggestions?

[email protected]

This material is based upon work supported by the National Science Foundation - Awards: IIS-0808678 and CNS-0751083