Robust Text Classifier on Test-Time Budgets Md Rizwan Parvez, Tolga - - PowerPoint PPT Presentation

▶
robust text classifier on test time budgets
SMART_READER_LITE
LIVE PREVIEW

Robust Text Classifier on Test-Time Budgets Md Rizwan Parvez, Tolga - - PowerPoint PPT Presentation

Jan 03, 2024 245 likes •407 views

Robust Text Classifier on Test-Time Budgets Md Rizwan Parvez, Tolga Bolukbasi, Kai-Wei Chang, Venkatesh Saligrama 1 Motivation Five stars for getting what you pay for here. The food is good and the price is affordable. A great location right

slide-1
SLIDE 1

Robust Text Classifier on Test-Time Budgets

Md Rizwan Parvez, Tolga Bolukbasi, Kai-Wei Chang, Venkatesh Saligrama

1

slide-2
SLIDE 2

Motivation

2

Great location but the food is not

Test Time

Five stars for getting what you pay for here. The food is good and the price is affordable. A great location right off Wilshire, very clean inside, and family friendly. The menu has more than burgers which is great. The steak sandwiches are delicious as well, I favor the teriyaki. The ahi is pretty good too and usually seared just perfectly. Also the salads aren't that bad as well. The soda fountain is the usual at places like this. This place has two signature drinks which are pretty refreshing, cucumber and strawberry.

Performance

Kai-Wei Chang (http://kwchang.net)

slide-3
SLIDE 3

Motivation

3

Five stars for getting what you pay for here. The food is good and the price is affordable. A great location right off Wilshire, very clean inside, and family friendly. The menu has more than burgers which is great. The steak sandwiches are delicious as well, I favor the teriyaki. The ahi is pretty good too and usually seared just perfectly. Also the salads aren't that bad as well. The soda fountain is the usual at places like this. This place has two signature drinks which are pretty refreshing, cucumber and strawberry.

Performance Test Time

Kai-Wei Chang (http://kwchang.net)

slide-4
SLIDE 4

intuition

A single text classifier for various test-time budgets

4

Performance Test Time

Kai-Wei Chang (http://kwchang.net)

slide-5
SLIDE 5

Business Sci/Tech .. Reuters - Software security companies and handset makers, including Finland’s Nokia (NOK1V.HE), are ..

Intuition

5

Telecom Austria taps the Bulgarian market. Telecom Austria, Austria's largest telecoms operator, obtained ..

Document Classification

Kai-Wei Chang (http://kwchang.net)

slide-6
SLIDE 6

A Meta Selector-Classifier Framework

6

Five stars for getting what you pay for here. The food is good and the price is affordable. A great location right off Wilshire, very clean inside, and family friendly. The menu has more than burgers which is great. The steak sandwiches are delicious as well, I favor the teriyaki. The ahi is pretty good too and usually seared just perfectly. Also the salads aren't that bad as well. The soda fountain is the usual at places like this. This place has two signature drinks which are pretty refreshing, cucumber and strawberry.

Original input sentence Text Classifier Selector

Time-time Budget (selection rate)

Great location but the food is not

Kai-Wei Chang (http://kwchang.net)

slide-7
SLIDE 7

A Meta Selector-Classifier Framework

7

Five stars for getting what you pay for here. The food is good and the price is affordable. A great location right off Wilshire, very clean inside, and family friendly. The menu has more than burgers which is great. The steak sandwiches are delicious as well, I favor the teriyaki. The ahi is pretty good too and usually seared just perfectly. Also the salads aren't that bad as well. The soda fountain is the usual at places like this. This place has two signature drinks which are pretty refreshing, cucumber and strawberry.

Original input sentence Text Classifier Selector

Budget (selection rate)

Great location but the food is not

  • 1. How to build an effective/efficient selector
  • 2. How to train a robust classifier?

Kai-Wei Chang (http://kwchang.net)

slide-8
SLIDE 8

End-to-End Training for WE Selector

  • Selector has to be light-weight
  • We consider a selector based on word-embedding
  • P(word x is selected) = σ(θT x)
  • End-to-end training [Lei et al. (2016)]
  • L1-regularizer to encourage sparsity

[Lei et al., (2016)] References: Tao Lei, Regina Barzilay, and Tommi S. Jaakkola. 2016. Rationalizing neural predictions. In EMNLP

Five stars for getting what you pay for

  • here. The food is good and the price is
  • affordable. A great location right off

Wilshire, very clean inside, and family

  • friendly. The menu has more than

burgers which is great. The steak sandwiches are delicious as well, I favor the teriyaki. The ahi is pretty good too and usually seared just perfectly. Also the salads aren't that bad as well. The soda fountain is the usual at places like

  • this. This place has two signature drinks

which are pretty refreshing, cucumber and strawberry.

σ(θT x)

Great location but the food is not

Text Classifier Loss Training with Loss + L1 regularizer

8

slide-9
SLIDE 9

Key Challenges - Fractured Texts

v Text classifier is not trained to deal with fractured texts

Kai-Wei Chang (http://kwchang.net) 9

Great location but the food is not Great but not Great location but food not Great location food not Text Classifier Prediction Accuracy Test-time

slide-10
SLIDE 10

Training a Robust Classifier with Data Aggregation

Five stars for getting what you pay for

  • here. The food is good and the price is
  • affordable. A great location right off

Wilshire, very clean inside, and family

  • friendly. The menu has more than

burgers which is great. The steak sandwiches are delicious as well, I favor the teriyaki. The ahi is pretty good too and usually seared just perfectly. Also the salads aren't that bad as well. The soda fountain is the usual at places like

  • this. This place has two signature drinks

which are pretty refreshing, cucumber and strawberry.

Selector A Selector B Selector C

Great location but food is not food is Great but the not Great but but not location is food is but the is Great but the not Great location but food is not

Text Classifier

Train

Aggregated corpus

Kai-Wei Chang (http://kwchang.net) 10

slide-11
SLIDE 11

An Alternative Selector

  • L1-regularized Bag-of-Words Model

11

…

θ3 θ|v| θ2 θ1 θ4

𝜄* = argmin𝜄 ∑ log (1 + exp (−𝑧 𝜄T . 𝑦 ⃗)) +

(2,4) 5 6 ||𝜄||1

𝑐: selection rate

Kai-Wei Chang (http://kwchang.net)

slide-12
SLIDE 12

Related Work

vSkip/Skim unimportant Text

vSkip-RNN [Campos et al., 2017] vSkim-RNN [Seo et al., 2018] vOur approach is agonistic to text classifier

v Model Compression [Bucila et al., 2006]

Kai-Wei Chang (http://kwchang.net) 12

slide-13
SLIDE 13

Experiment Setting

  • Datasets: SST-2, IMDB, AGNews, and Yelp
  • Classifiers: LSTM, BCN [McCann et al., 2017]
  • Aggregate corpus from both selectors:
  • Selection rate = {50%, 60%, 70%, 80%, 90%, 100%}

13 Kai-Wei Chang (http://kwchang.net)

slide-14
SLIDE 14

Error Rate v.s. Test-Time

14

IMDB AGNews See our paper for more results

Kai-Wei Chang (http://kwchang.net)

slide-15
SLIDE 15

Conclusion

  • We present a selector-classifier meta framework for

building efficient text classifier

  • We propose efficient selectors and data aggregation

scheme to train classifiers that are compatible with fractured text selected at any rate. Source codes are available at: Thanks!!!!

15

https://github.com/uclanlp/Fast-and-Robust-Text-Classification

Kai-Wei Chang (http://kwchang.net)