SlideShare a Scribd company logo
1 of 16
Yanghoon Kim, Hwanhee Lee, Joongbo Shin and Kyomin Jung
Improving Neural Question Generation
using Answer Separation
김양훈
Background
Neural question generation (NQG)
- Generating a question from a given text passage with deep neural networks.
Importance of NQG
- Generating questions for educational materials.
- Generating questions for improving QA systems.
NQG with target Answer
- Generating a question from a given text passage and a given target
answer(question target) with deep neural networks.
2
Original passage: John Francis O’hara was elected president of Notre Dame in 1934.
Generated question 1: Who was elected president of Notre Dame in 1934?
Generated question 2: When was John Francis O’hara elected president of Notre Dame?
Problem
Previous NQG systems suffer from a critical problem
- Some models don’t take the target answer into account.
- RNNs often follow a shallow generation process.
- Some models can’t well grasp the target answer(question target) .
- A sophisticated proportion of generated questions include word in the target
answer.
3
Original passage: John Francis O’Hara was elected president of Notre Dame in 1934.
Given target answer: John Francis O’hara
Correctly Generated question: Who was elected president of Notre Dame in 1934?
Incorrectly generated question: Who was elected John Francis?
Contribution
We propose answer-separated seq2seq
- Treats the target answer(question target) and the passage separately.
- Prevent the generated question from including words in the target answer.
- Better capture the information from both the target answer and the passage
- We propose keyword-net
- Model is consistently aware of the target answer.
- Extract the key information in the target answer.
- We use retrieval style word generator
- Take the word meaning into account when generating words.
4
Task Definition
5
Model
Base model
- We use RNN encoder-decoder with attention
Answer-separated seq2seq consist of
- Answer-separated passage encoder
- Target answer encoder
- Answer-separated decoder
- keyword-net
- Retrieval style word generator
6
Model
Answer-separated passage encoder
- A simple preprocessing of the input passage
- Original passage: Steve Jobs is the founder of Apple.
- Masked passage: Steve Jobs is the <a> .
- A one-layer bi-LSTM
Answer encoder
- A one-layer bi-LSTM
7
Model
Answer-separated decoder
- A one-layer LSTM
- keyword-net
- Let the model consistently be aware of the target answer.
- Extract key information.
- Passage: Steve Jobs is the founder of Apple
- Target answer: founder of Apple
8
Model
Answer-separated decoder
- Retrieval style word generator by (Ma et al. 2018)*
- seq2seq has tendency to memorize the sequence pattern rather than
reflecting word meanings
- The word generator produces words by querying the distributed word
representations.
*Query and Output: Generating Words by Querying Distributed Word Representations for Paraphrase Generation 9
Experiment
Data
- Processed version of SQuAD 1.1
- Data split 1: 70,484/10,570/11,877 (train/dev/test)
- Data split 2: 86,635/8,965/8,964
Evaluation
Our model(Ass2s) outperform the previous state-of-the-art model
10
Experiment
Impact of answer separation
- Ability to capture target answer
- We checked if the target answer is included in the generated question
- AP : Answer position Feature (BIO scheme)
- (Song et al. 2018) used the copy mechanism.
11
Our model has better ability to generate the right question given the target answer
Experiment
Impact of answer separation
- Interrogative word prediction
- “What” takes up more than half of the whole training set
- “Which” : “Which year” can be represented as “When”
- “why”, “yes/no” : only takes up 1.5% and 1.2% of the training set.
12
Our model has better ability to predict the question type for the given target answer
Experiment
Impact of answer separation
- Attention from <answer>
- (a) is the attention matrix from our model
- (b) is the attention matrix from seq2seq + AP
- <a> token gives the highest attention weights to the interrogative word “who” in (a)
13
Experiment
Question generation for machine comprehension
- Use named entities as target answers, generate synthetic data for machine
comprehension system(QA net by Google).
- ALL : Evaluation result of SQuAD dev set(10k)
- NER : Evaluation result of partial SQuAD dev set(4k)
- answers of single named entity
14
Conclusion
We propose Answer-separated seq2seq for NQG
- Separate utilization of target answer and the passage(without target answer)
- By masking the target answer inside the passage
- By using keyword-net to extract key feature from target answer
- By using retrieval style word generator to capture word meaning information
- Our model can
- Reduce the probability that the target answer is included by the generated question
- Generate fluent and right question for the given passage and the target answer
- Better inference the type of question
Thank you for listening!
Code, paper: https://yanghoonkim.github.io
Questions: ad26kr@snu.ac.kr

More Related Content

Similar to Improving Neural Question Generation using Answer Separation.

Improving neural question generation using answer separation
Improving neural question generation using answer separationImproving neural question generation using answer separation
Improving neural question generation using answer separationNAVER Engineering
 
Attention-based Models (DLAI D8L 2017 UPC Deep Learning for Artificial Intell...
Attention-based Models (DLAI D8L 2017 UPC Deep Learning for Artificial Intell...Attention-based Models (DLAI D8L 2017 UPC Deep Learning for Artificial Intell...
Attention-based Models (DLAI D8L 2017 UPC Deep Learning for Artificial Intell...Universitat Politècnica de Catalunya
 
Natural Language Generation / Stanford cs224n 2019w lecture 15 Review
Natural Language Generation / Stanford cs224n 2019w lecture 15 ReviewNatural Language Generation / Stanford cs224n 2019w lecture 15 Review
Natural Language Generation / Stanford cs224n 2019w lecture 15 Reviewchangedaeoh
 
Deep Learning Enabled Question Answering System to Automate Corporate Helpdesk
Deep Learning Enabled Question Answering System to Automate Corporate HelpdeskDeep Learning Enabled Question Answering System to Automate Corporate Helpdesk
Deep Learning Enabled Question Answering System to Automate Corporate HelpdeskSaurabh Saxena
 
Wecp all-india-test-series-program-brochure
Wecp all-india-test-series-program-brochureWecp all-india-test-series-program-brochure
Wecp all-india-test-series-program-brochureBIPIN KAUSHIK
 
WISS QA Do it yourself Question answering over Linked Data
WISS QA Do it yourself Question answering over Linked DataWISS QA Do it yourself Question answering over Linked Data
WISS QA Do it yourself Question answering over Linked DataAndre Freitas
 
Python week 2 2019 2020 for g10 by eng.osama ghandour
Python week 2 2019 2020 for g10 by eng.osama ghandourPython week 2 2019 2020 for g10 by eng.osama ghandour
Python week 2 2019 2020 for g10 by eng.osama ghandourOsama Ghandour Geris
 
Building a Meta-search Engine
Building a Meta-search EngineBuilding a Meta-search Engine
Building a Meta-search EngineAyan Chandra
 
The Frontier of Deep Learning in 2020 and Beyond
The Frontier of Deep Learning in 2020 and BeyondThe Frontier of Deep Learning in 2020 and Beyond
The Frontier of Deep Learning in 2020 and BeyondNUS-ISS
 
Adversarial learning for neural dialogue generation
Adversarial learning for neural dialogue generationAdversarial learning for neural dialogue generation
Adversarial learning for neural dialogue generationKeon Kim
 
Automatic Test Suite Generation for Key-Points Detection DNNs using Many-Obje...
Automatic Test Suite Generation for Key-Points Detection DNNs using Many-Obje...Automatic Test Suite Generation for Key-Points Detection DNNs using Many-Obje...
Automatic Test Suite Generation for Key-Points Detection DNNs using Many-Obje...Lionel Briand
 
EVALUATION OF SINGLE-SPAN MODELS ON EXTRACTIVE MULTI-SPAN QUESTION-ANSWERING
EVALUATION OF SINGLE-SPAN MODELS ON EXTRACTIVE MULTI-SPAN QUESTION-ANSWERINGEVALUATION OF SINGLE-SPAN MODELS ON EXTRACTIVE MULTI-SPAN QUESTION-ANSWERING
EVALUATION OF SINGLE-SPAN MODELS ON EXTRACTIVE MULTI-SPAN QUESTION-ANSWERINGIJwest
 
EVALUATION OF SINGLE-SPAN MODELS ON EXTRACTIVE MULTI-SPAN QUESTION-ANSWERING
EVALUATION OF SINGLE-SPAN MODELS ON EXTRACTIVE MULTI-SPAN QUESTION-ANSWERINGEVALUATION OF SINGLE-SPAN MODELS ON EXTRACTIVE MULTI-SPAN QUESTION-ANSWERING
EVALUATION OF SINGLE-SPAN MODELS ON EXTRACTIVE MULTI-SPAN QUESTION-ANSWERINGdannyijwest
 

Similar to Improving Neural Question Generation using Answer Separation. (20)

Improving neural question generation using answer separation
Improving neural question generation using answer separationImproving neural question generation using answer separation
Improving neural question generation using answer separation
 
Attention-based Models (DLAI D8L 2017 UPC Deep Learning for Artificial Intell...
Attention-based Models (DLAI D8L 2017 UPC Deep Learning for Artificial Intell...Attention-based Models (DLAI D8L 2017 UPC Deep Learning for Artificial Intell...
Attention-based Models (DLAI D8L 2017 UPC Deep Learning for Artificial Intell...
 
ACL 2018 Recap
ACL 2018 RecapACL 2018 Recap
ACL 2018 Recap
 
Natural Language Generation / Stanford cs224n 2019w lecture 15 Review
Natural Language Generation / Stanford cs224n 2019w lecture 15 ReviewNatural Language Generation / Stanford cs224n 2019w lecture 15 Review
Natural Language Generation / Stanford cs224n 2019w lecture 15 Review
 
Deep Learning Enabled Question Answering System to Automate Corporate Helpdesk
Deep Learning Enabled Question Answering System to Automate Corporate HelpdeskDeep Learning Enabled Question Answering System to Automate Corporate Helpdesk
Deep Learning Enabled Question Answering System to Automate Corporate Helpdesk
 
Wecp all-india-test-series-program-brochure
Wecp all-india-test-series-program-brochureWecp all-india-test-series-program-brochure
Wecp all-india-test-series-program-brochure
 
Wecp all-india-test-series-program-brochure
Wecp all-india-test-series-program-brochureWecp all-india-test-series-program-brochure
Wecp all-india-test-series-program-brochure
 
WISS QA Do it yourself Question answering over Linked Data
WISS QA Do it yourself Question answering over Linked DataWISS QA Do it yourself Question answering over Linked Data
WISS QA Do it yourself Question answering over Linked Data
 
American sign language recognizer
American sign language recognizerAmerican sign language recognizer
American sign language recognizer
 
Seminar2017
Seminar2017Seminar2017
Seminar2017
 
Python week 2 2019 2020 for g10 by eng.osama ghandour
Python week 2 2019 2020 for g10 by eng.osama ghandourPython week 2 2019 2020 for g10 by eng.osama ghandour
Python week 2 2019 2020 for g10 by eng.osama ghandour
 
Building a Meta-search Engine
Building a Meta-search EngineBuilding a Meta-search Engine
Building a Meta-search Engine
 
The Frontier of Deep Learning in 2020 and Beyond
The Frontier of Deep Learning in 2020 and BeyondThe Frontier of Deep Learning in 2020 and Beyond
The Frontier of Deep Learning in 2020 and Beyond
 
Python week 1 2020-2021
Python week 1 2020-2021Python week 1 2020-2021
Python week 1 2020-2021
 
rlhf.pdf
rlhf.pdfrlhf.pdf
rlhf.pdf
 
Adversarial learning for neural dialogue generation
Adversarial learning for neural dialogue generationAdversarial learning for neural dialogue generation
Adversarial learning for neural dialogue generation
 
Automatic Test Suite Generation for Key-Points Detection DNNs using Many-Obje...
Automatic Test Suite Generation for Key-Points Detection DNNs using Many-Obje...Automatic Test Suite Generation for Key-Points Detection DNNs using Many-Obje...
Automatic Test Suite Generation for Key-Points Detection DNNs using Many-Obje...
 
Text Analytics for Legal work
Text Analytics for Legal workText Analytics for Legal work
Text Analytics for Legal work
 
EVALUATION OF SINGLE-SPAN MODELS ON EXTRACTIVE MULTI-SPAN QUESTION-ANSWERING
EVALUATION OF SINGLE-SPAN MODELS ON EXTRACTIVE MULTI-SPAN QUESTION-ANSWERINGEVALUATION OF SINGLE-SPAN MODELS ON EXTRACTIVE MULTI-SPAN QUESTION-ANSWERING
EVALUATION OF SINGLE-SPAN MODELS ON EXTRACTIVE MULTI-SPAN QUESTION-ANSWERING
 
EVALUATION OF SINGLE-SPAN MODELS ON EXTRACTIVE MULTI-SPAN QUESTION-ANSWERING
EVALUATION OF SINGLE-SPAN MODELS ON EXTRACTIVE MULTI-SPAN QUESTION-ANSWERINGEVALUATION OF SINGLE-SPAN MODELS ON EXTRACTIVE MULTI-SPAN QUESTION-ANSWERING
EVALUATION OF SINGLE-SPAN MODELS ON EXTRACTIVE MULTI-SPAN QUESTION-ANSWERING
 

Recently uploaded

Nanoparticles synthesis and characterization​ ​
Nanoparticles synthesis and characterization​  ​Nanoparticles synthesis and characterization​  ​
Nanoparticles synthesis and characterization​ ​kaibalyasahoo82800
 
Call Girls in Munirka Delhi 💯Call Us 🔝8264348440🔝
Call Girls in Munirka Delhi 💯Call Us 🔝8264348440🔝Call Girls in Munirka Delhi 💯Call Us 🔝8264348440🔝
Call Girls in Munirka Delhi 💯Call Us 🔝8264348440🔝soniya singh
 
A relative description on Sonoporation.pdf
A relative description on Sonoporation.pdfA relative description on Sonoporation.pdf
A relative description on Sonoporation.pdfnehabiju2046
 
Stunning ➥8448380779▻ Call Girls In Panchshil Enclave Delhi NCR
Stunning ➥8448380779▻ Call Girls In Panchshil Enclave Delhi NCRStunning ➥8448380779▻ Call Girls In Panchshil Enclave Delhi NCR
Stunning ➥8448380779▻ Call Girls In Panchshil Enclave Delhi NCRDelhi Call girls
 
NAVSEA PEO USC - Unmanned & Small Combatants 26Oct23.pdf
NAVSEA PEO USC - Unmanned & Small Combatants 26Oct23.pdfNAVSEA PEO USC - Unmanned & Small Combatants 26Oct23.pdf
NAVSEA PEO USC - Unmanned & Small Combatants 26Oct23.pdfWadeK3
 
Call Girls in Mayapuri Delhi 💯Call Us 🔝9953322196🔝 💯Escort.
Call Girls in Mayapuri Delhi 💯Call Us 🔝9953322196🔝 💯Escort.Call Girls in Mayapuri Delhi 💯Call Us 🔝9953322196🔝 💯Escort.
Call Girls in Mayapuri Delhi 💯Call Us 🔝9953322196🔝 💯Escort.aasikanpl
 
Analytical Profile of Coleus Forskohlii | Forskolin .pdf
Analytical Profile of Coleus Forskohlii | Forskolin .pdfAnalytical Profile of Coleus Forskohlii | Forskolin .pdf
Analytical Profile of Coleus Forskohlii | Forskolin .pdfSwapnil Therkar
 
Traditional Agroforestry System in India- Shifting Cultivation, Taungya, Home...
Traditional Agroforestry System in India- Shifting Cultivation, Taungya, Home...Traditional Agroforestry System in India- Shifting Cultivation, Taungya, Home...
Traditional Agroforestry System in India- Shifting Cultivation, Taungya, Home...jana861314
 
Hubble Asteroid Hunter III. Physical properties of newly found asteroids
Hubble Asteroid Hunter III. Physical properties of newly found asteroidsHubble Asteroid Hunter III. Physical properties of newly found asteroids
Hubble Asteroid Hunter III. Physical properties of newly found asteroidsSérgio Sacani
 
GFP in rDNA Technology (Biotechnology).pptx
GFP in rDNA Technology (Biotechnology).pptxGFP in rDNA Technology (Biotechnology).pptx
GFP in rDNA Technology (Biotechnology).pptxAleenaTreesaSaji
 
Orientation, design and principles of polyhouse
Orientation, design and principles of polyhouseOrientation, design and principles of polyhouse
Orientation, design and principles of polyhousejana861314
 
Lucknow 💋 Russian Call Girls Lucknow Finest Escorts Service 8923113531 Availa...
Lucknow 💋 Russian Call Girls Lucknow Finest Escorts Service 8923113531 Availa...Lucknow 💋 Russian Call Girls Lucknow Finest Escorts Service 8923113531 Availa...
Lucknow 💋 Russian Call Girls Lucknow Finest Escorts Service 8923113531 Availa...anilsa9823
 
Recombination DNA Technology (Nucleic Acid Hybridization )
Recombination DNA Technology (Nucleic Acid Hybridization )Recombination DNA Technology (Nucleic Acid Hybridization )
Recombination DNA Technology (Nucleic Acid Hybridization )aarthirajkumar25
 
Boyles law module in the grade 10 science
Boyles law module in the grade 10 scienceBoyles law module in the grade 10 science
Boyles law module in the grade 10 sciencefloriejanemacaya1
 
PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...
PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...
PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...Sérgio Sacani
 
VIRUSES structure and classification ppt by Dr.Prince C P
VIRUSES structure and classification ppt by Dr.Prince C PVIRUSES structure and classification ppt by Dr.Prince C P
VIRUSES structure and classification ppt by Dr.Prince C PPRINCE C P
 
Analytical Profile of Coleus Forskohlii | Forskolin .pptx
Analytical Profile of Coleus Forskohlii | Forskolin .pptxAnalytical Profile of Coleus Forskohlii | Forskolin .pptx
Analytical Profile of Coleus Forskohlii | Forskolin .pptxSwapnil Therkar
 
Is RISC-V ready for HPC workload? Maybe?
Is RISC-V ready for HPC workload? Maybe?Is RISC-V ready for HPC workload? Maybe?
Is RISC-V ready for HPC workload? Maybe?Patrick Diehl
 
Work, Energy and Power for class 10 ICSE Physics
Work, Energy and Power for class 10 ICSE PhysicsWork, Energy and Power for class 10 ICSE Physics
Work, Energy and Power for class 10 ICSE Physicsvishikhakeshava1
 
Call Us ≽ 9953322196 ≼ Call Girls In Mukherjee Nagar(Delhi) |
Call Us ≽ 9953322196 ≼ Call Girls In Mukherjee Nagar(Delhi) |Call Us ≽ 9953322196 ≼ Call Girls In Mukherjee Nagar(Delhi) |
Call Us ≽ 9953322196 ≼ Call Girls In Mukherjee Nagar(Delhi) |aasikanpl
 

Recently uploaded (20)

Nanoparticles synthesis and characterization​ ​
Nanoparticles synthesis and characterization​  ​Nanoparticles synthesis and characterization​  ​
Nanoparticles synthesis and characterization​ ​
 
Call Girls in Munirka Delhi 💯Call Us 🔝8264348440🔝
Call Girls in Munirka Delhi 💯Call Us 🔝8264348440🔝Call Girls in Munirka Delhi 💯Call Us 🔝8264348440🔝
Call Girls in Munirka Delhi 💯Call Us 🔝8264348440🔝
 
A relative description on Sonoporation.pdf
A relative description on Sonoporation.pdfA relative description on Sonoporation.pdf
A relative description on Sonoporation.pdf
 
Stunning ➥8448380779▻ Call Girls In Panchshil Enclave Delhi NCR
Stunning ➥8448380779▻ Call Girls In Panchshil Enclave Delhi NCRStunning ➥8448380779▻ Call Girls In Panchshil Enclave Delhi NCR
Stunning ➥8448380779▻ Call Girls In Panchshil Enclave Delhi NCR
 
NAVSEA PEO USC - Unmanned & Small Combatants 26Oct23.pdf
NAVSEA PEO USC - Unmanned & Small Combatants 26Oct23.pdfNAVSEA PEO USC - Unmanned & Small Combatants 26Oct23.pdf
NAVSEA PEO USC - Unmanned & Small Combatants 26Oct23.pdf
 
Call Girls in Mayapuri Delhi 💯Call Us 🔝9953322196🔝 💯Escort.
Call Girls in Mayapuri Delhi 💯Call Us 🔝9953322196🔝 💯Escort.Call Girls in Mayapuri Delhi 💯Call Us 🔝9953322196🔝 💯Escort.
Call Girls in Mayapuri Delhi 💯Call Us 🔝9953322196🔝 💯Escort.
 
Analytical Profile of Coleus Forskohlii | Forskolin .pdf
Analytical Profile of Coleus Forskohlii | Forskolin .pdfAnalytical Profile of Coleus Forskohlii | Forskolin .pdf
Analytical Profile of Coleus Forskohlii | Forskolin .pdf
 
Traditional Agroforestry System in India- Shifting Cultivation, Taungya, Home...
Traditional Agroforestry System in India- Shifting Cultivation, Taungya, Home...Traditional Agroforestry System in India- Shifting Cultivation, Taungya, Home...
Traditional Agroforestry System in India- Shifting Cultivation, Taungya, Home...
 
Hubble Asteroid Hunter III. Physical properties of newly found asteroids
Hubble Asteroid Hunter III. Physical properties of newly found asteroidsHubble Asteroid Hunter III. Physical properties of newly found asteroids
Hubble Asteroid Hunter III. Physical properties of newly found asteroids
 
GFP in rDNA Technology (Biotechnology).pptx
GFP in rDNA Technology (Biotechnology).pptxGFP in rDNA Technology (Biotechnology).pptx
GFP in rDNA Technology (Biotechnology).pptx
 
Orientation, design and principles of polyhouse
Orientation, design and principles of polyhouseOrientation, design and principles of polyhouse
Orientation, design and principles of polyhouse
 
Lucknow 💋 Russian Call Girls Lucknow Finest Escorts Service 8923113531 Availa...
Lucknow 💋 Russian Call Girls Lucknow Finest Escorts Service 8923113531 Availa...Lucknow 💋 Russian Call Girls Lucknow Finest Escorts Service 8923113531 Availa...
Lucknow 💋 Russian Call Girls Lucknow Finest Escorts Service 8923113531 Availa...
 
Recombination DNA Technology (Nucleic Acid Hybridization )
Recombination DNA Technology (Nucleic Acid Hybridization )Recombination DNA Technology (Nucleic Acid Hybridization )
Recombination DNA Technology (Nucleic Acid Hybridization )
 
Boyles law module in the grade 10 science
Boyles law module in the grade 10 scienceBoyles law module in the grade 10 science
Boyles law module in the grade 10 science
 
PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...
PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...
PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...
 
VIRUSES structure and classification ppt by Dr.Prince C P
VIRUSES structure and classification ppt by Dr.Prince C PVIRUSES structure and classification ppt by Dr.Prince C P
VIRUSES structure and classification ppt by Dr.Prince C P
 
Analytical Profile of Coleus Forskohlii | Forskolin .pptx
Analytical Profile of Coleus Forskohlii | Forskolin .pptxAnalytical Profile of Coleus Forskohlii | Forskolin .pptx
Analytical Profile of Coleus Forskohlii | Forskolin .pptx
 
Is RISC-V ready for HPC workload? Maybe?
Is RISC-V ready for HPC workload? Maybe?Is RISC-V ready for HPC workload? Maybe?
Is RISC-V ready for HPC workload? Maybe?
 
Work, Energy and Power for class 10 ICSE Physics
Work, Energy and Power for class 10 ICSE PhysicsWork, Energy and Power for class 10 ICSE Physics
Work, Energy and Power for class 10 ICSE Physics
 
Call Us ≽ 9953322196 ≼ Call Girls In Mukherjee Nagar(Delhi) |
Call Us ≽ 9953322196 ≼ Call Girls In Mukherjee Nagar(Delhi) |Call Us ≽ 9953322196 ≼ Call Girls In Mukherjee Nagar(Delhi) |
Call Us ≽ 9953322196 ≼ Call Girls In Mukherjee Nagar(Delhi) |
 

Improving Neural Question Generation using Answer Separation.

  • 1. Yanghoon Kim, Hwanhee Lee, Joongbo Shin and Kyomin Jung Improving Neural Question Generation using Answer Separation 김양훈
  • 2. Background Neural question generation (NQG) - Generating a question from a given text passage with deep neural networks. Importance of NQG - Generating questions for educational materials. - Generating questions for improving QA systems. NQG with target Answer - Generating a question from a given text passage and a given target answer(question target) with deep neural networks. 2 Original passage: John Francis O’hara was elected president of Notre Dame in 1934. Generated question 1: Who was elected president of Notre Dame in 1934? Generated question 2: When was John Francis O’hara elected president of Notre Dame?
  • 3. Problem Previous NQG systems suffer from a critical problem - Some models don’t take the target answer into account. - RNNs often follow a shallow generation process. - Some models can’t well grasp the target answer(question target) . - A sophisticated proportion of generated questions include word in the target answer. 3 Original passage: John Francis O’Hara was elected president of Notre Dame in 1934. Given target answer: John Francis O’hara Correctly Generated question: Who was elected president of Notre Dame in 1934? Incorrectly generated question: Who was elected John Francis?
  • 4. Contribution We propose answer-separated seq2seq - Treats the target answer(question target) and the passage separately. - Prevent the generated question from including words in the target answer. - Better capture the information from both the target answer and the passage - We propose keyword-net - Model is consistently aware of the target answer. - Extract the key information in the target answer. - We use retrieval style word generator - Take the word meaning into account when generating words. 4
  • 6. Model Base model - We use RNN encoder-decoder with attention Answer-separated seq2seq consist of - Answer-separated passage encoder - Target answer encoder - Answer-separated decoder - keyword-net - Retrieval style word generator 6
  • 7. Model Answer-separated passage encoder - A simple preprocessing of the input passage - Original passage: Steve Jobs is the founder of Apple. - Masked passage: Steve Jobs is the <a> . - A one-layer bi-LSTM Answer encoder - A one-layer bi-LSTM 7
  • 8. Model Answer-separated decoder - A one-layer LSTM - keyword-net - Let the model consistently be aware of the target answer. - Extract key information. - Passage: Steve Jobs is the founder of Apple - Target answer: founder of Apple 8
  • 9. Model Answer-separated decoder - Retrieval style word generator by (Ma et al. 2018)* - seq2seq has tendency to memorize the sequence pattern rather than reflecting word meanings - The word generator produces words by querying the distributed word representations. *Query and Output: Generating Words by Querying Distributed Word Representations for Paraphrase Generation 9
  • 10. Experiment Data - Processed version of SQuAD 1.1 - Data split 1: 70,484/10,570/11,877 (train/dev/test) - Data split 2: 86,635/8,965/8,964 Evaluation Our model(Ass2s) outperform the previous state-of-the-art model 10
  • 11. Experiment Impact of answer separation - Ability to capture target answer - We checked if the target answer is included in the generated question - AP : Answer position Feature (BIO scheme) - (Song et al. 2018) used the copy mechanism. 11 Our model has better ability to generate the right question given the target answer
  • 12. Experiment Impact of answer separation - Interrogative word prediction - “What” takes up more than half of the whole training set - “Which” : “Which year” can be represented as “When” - “why”, “yes/no” : only takes up 1.5% and 1.2% of the training set. 12 Our model has better ability to predict the question type for the given target answer
  • 13. Experiment Impact of answer separation - Attention from <answer> - (a) is the attention matrix from our model - (b) is the attention matrix from seq2seq + AP - <a> token gives the highest attention weights to the interrogative word “who” in (a) 13
  • 14. Experiment Question generation for machine comprehension - Use named entities as target answers, generate synthetic data for machine comprehension system(QA net by Google). - ALL : Evaluation result of SQuAD dev set(10k) - NER : Evaluation result of partial SQuAD dev set(4k) - answers of single named entity 14
  • 15. Conclusion We propose Answer-separated seq2seq for NQG - Separate utilization of target answer and the passage(without target answer) - By masking the target answer inside the passage - By using keyword-net to extract key feature from target answer - By using retrieval style word generator to capture word meaning information - Our model can - Reduce the probability that the target answer is included by the generated question - Generate fluent and right question for the given passage and the target answer - Better inference the type of question
  • 16. Thank you for listening! Code, paper: https://yanghoonkim.github.io Questions: ad26kr@snu.ac.kr