+ All Categories
Home > Documents > P4: Feedforward Neural Networks · 2021. 4. 22. · Tasks & Datasets Quizbowl QA task • Input:...

P4: Feedforward Neural Networks · 2021. 4. 22. · Tasks & Datasets Quizbowl QA task • Input:...

Date post: 01-Sep-2021
Category:
Upload: others
View: 1 times
Download: 0 times
Share this document with a friend
16
P4: Feedforward Neural Networks Spring 2021 COS 584 Advanced Natural Language Processing
Transcript
Page 1: P4: Feedforward Neural Networks · 2021. 4. 22. · Tasks & Datasets Quizbowl QA task • Input: 4-6 sentences describing anentity (authors, battles, or events) • Output:entity

P4: Feedforward Neural Networks

Spring 2021

COS 584

Advanced Natural Language Processing

Page 2: P4: Feedforward Neural Networks · 2021. 4. 22. · Tasks & Datasets Quizbowl QA task • Input: 4-6 sentences describing anentity (authors, battles, or events) • Output:entity

Annoucements

• Feedbackform: https://forms.gle/Bsgng7m21rXxWsTw5

• Reading materials

• Perusall use

• Pre-lecture questions

• Class structure (lecture + discussion)

• 584 readings will NOT be tested in midterm.

Page 3: P4: Feedforward Neural Networks · 2021. 4. 22. · Tasks & Datasets Quizbowl QA task • Input: 4-6 sentences describing anentity (authors, battles, or events) • Output:entity

ACL 2015

Page 4: P4: Feedforward Neural Networks · 2021. 4. 22. · Tasks & Datasets Quizbowl QA task • Input: 4-6 sentences describing anentity (authors, battles, or events) • Output:entity

Key takeaways

• A very simple model called Deep Averaging Networks (DAN) achieves competent performance on sentiment analysis and factoid question answering.

How can we interpret these results?

Page 5: P4: Feedforward Neural Networks · 2021. 4. 22. · Tasks & Datasets Quizbowl QA task • Input: 4-6 sentences describing anentity (authors, battles, or events) • Output:entity

A little bit of historyEMNLP 2013

NIPS 2014

Page 6: P4: Feedforward Neural Networks · 2021. 4. 22. · Tasks & Datasets Quizbowl QA task • Input: 4-6 sentences describing anentity (authors, battles, or events) • Output:entity

A little bit of historyEMNLP 2014

JMLR 2014

Page 7: P4: Feedforward Neural Networks · 2021. 4. 22. · Tasks & Datasets Quizbowl QA task • Input: 4-6 sentences describing anentity (authors, battles, or events) • Output:entity

Deep Averaging Networks (DAN)

NBOW DAN

The model doesn’t model the word order and even n-gram information!

Page 8: P4: Feedforward Neural Networks · 2021. 4. 22. · Tasks & Datasets Quizbowl QA task • Input: 4-6 sentences describing anentity (authors, battles, or events) • Output:entity

Recursive Neural Networks (RNNs)

• This model relies on a parser to provide an input tree (error-prone)• W is shared in all the composition functions

Page 9: P4: Feedforward Neural Networks · 2021. 4. 22. · Tasks & Datasets Quizbowl QA task • Input: 4-6 sentences describing anentity (authors, battles, or events) • Output:entity

Word dropout

• Drop entire word embeddings and take the average of remaining words

Page 10: P4: Feedforward Neural Networks · 2021. 4. 22. · Tasks & Datasets Quizbowl QA task • Input: 4-6 sentences describing anentity (authors, battles, or events) • Output:entity

Tasks & Datasets

Sentiment analysis

• RT: Rotten Tomatoes

• 2-class, sentence-level classification

• Stanford sentiment treebank

• 2-class or 5-class - ++, +, 0, -, —

• sentence-level classification

• IMDB

• 2-class, document-level classification

See more examples at https://nlp.stanford.edu/sentiment/treebank.html

The phrase-level labels are only used for training!

Especially crucial for tree-based models

Page 11: P4: Feedforward Neural Networks · 2021. 4. 22. · Tasks & Datasets Quizbowl QA task • Input: 4-6 sentences describing anentity (authors, battles, or events) • Output:entity

Tasks & Datasets

Quizbowl QA task

• Input: 4-6 sentences describing an entity (authors, battles, or events)

• Output: entity

• This creature has female counterparts named Penny and Gown.

• This creature appears dressed in Viking armor and carrying an ax when he is used as the mascot of PaX, a least privilege protection patch.

• This creature’s counterparts include Daemon on the Berkeley Software Distribution, or BSD.

• For ten points, name this mascot of the Linux operating system, a penguin whose name refers to formal male attire.

Example

• 3,761 questions

• Augmented with 53,234 sentence/page-title pairs from Wikipedia

Page 12: P4: Feedforward Neural Networks · 2021. 4. 22. · Tasks & Datasets Quizbowl QA task • Input: 4-6 sentences describing anentity (authors, battles, or events) • Output:entity

Experiments: sentiment analysis

• Initialization with GloVe embeddings helps

• Phrase-level labels helps

• It seems to work better on sentence-level tasks than on document-level tasks

• DAN is fast and competitive on sentiment analysis

Page 13: P4: Feedforward Neural Networks · 2021. 4. 22. · Tasks & Datasets Quizbowl QA task • Input: 4-6 sentences describing anentity (authors, battles, or events) • Output:entity

Experiments: quizbowl QA

• The gap between DAN and QANTA increases when # of sentences increases

• DAN improves with noisy data

QANTA: recursive neural networks based on dependency tree

Page 14: P4: Feedforward Neural Networks · 2021. 4. 22. · Tasks & Datasets Quizbowl QA task • Input: 4-6 sentences describing anentity (authors, battles, or events) • Output:entity

How do DANs work?

the film’s performances were awesome

Page 15: P4: Feedforward Neural Networks · 2021. 4. 22. · Tasks & Datasets Quizbowl QA task • Input: 4-6 sentences describing anentity (authors, battles, or events) • Output:entity

DANs can’t handle negations well but tree-based models can’t either

Page 16: P4: Feedforward Neural Networks · 2021. 4. 22. · Tasks & Datasets Quizbowl QA task • Input: 4-6 sentences describing anentity (authors, battles, or events) • Output:entity

Discussion

• DANs are fast and competitive on sentiment analysis and quizbowl QA tasks. Do you think these results generalize? What about other tasks?

• What are the limitations of DANs? How can we improve them?

• Do word order and compositionally matter?

• What does word dropout do?

• What do the non-linear layers do in DANs?


Recommended