Sreyasi Nag Chowdhury, Niket Tandon, Gerhard Weikum · Sreyasi Nag Chowdhury, Niket Tandon, Gerhard...
Transcript of Sreyasi Nag Chowdhury, Niket Tandon, Gerhard Weikum · Sreyasi Nag Chowdhury, Niket Tandon, Gerhard...
Sreyasi Nag Chowdhury, Niket Tandon, Gerhard Weikum
Max Planck Institute for Informatics, Saarbrücken, Germany
17/06/2016 1Sreyasi Nag Chowdhury, AKBC 2016
User Query
Concrete Abstract
Q: bicycle in street Q: environment friendly traffic
17/06/2016 1Sreyasi Nag Chowdhury, AKBC 2016
User Query
Concrete Abstract
“Wow! Double-decker buses still run!”
Q: bicycle in street Q: environment friendly traffic
17/06/2016 1Sreyasi Nag Chowdhury, AKBC 2016
User Query
Concrete Abstract
“Wow! Double-decker buses still run!”
Text-only
Q: bicycle in street Q: environment friendly traffic
17/06/2016 1Sreyasi Nag Chowdhury, AKBC 2016
User Query
Concrete Abstract
“Wow! Double-decker buses still run!”
Visual objects: bicycle, bus, car
Text-only
Text + visual
Q: bicycle in street Q: environment friendly traffic
17/06/2016 1Sreyasi Nag Chowdhury, AKBC 2016
User Query
Concrete Abstract
“Wow! Double-decker buses still run!”
Visual objects: bicycle, bus, car
Text-only
Text + visual
Q: bicycle in street Q: environment friendly traffic
“Biking by the river”
Visual objects: train, piano
17/06/2016 1Sreyasi Nag Chowdhury, AKBC 2016
User Query
Concrete Abstract
“Wow! Double-decker buses still run!”
Visual objects: bicycle, bus, car
Text-only
Text + visual
Q: bicycle in street Q: environment friendly traffic
“Biking by the river”
Visual objects: train, piano
Text-only
Text + visual
17/06/2016 1Sreyasi Nag Chowdhury, AKBC 2016
User Query
Concrete Abstract
“Wow! Double-decker buses still run!”
Visual objects: bicycle, bus, car
Text-only
Text + visual
“Riding for a cause.”
Visual objects: person, bicycle
CSK: (riding bicycle, be, environment friendly)
Q: bicycle in street Q: environment friendly traffic
“Biking by the river”
Visual objects: train, piano
Text-only
Text + visual
17/06/2016 1Sreyasi Nag Chowdhury, AKBC 2016
User Query
Concrete Abstract
“Wow! Double-decker buses still run!”
Visual objects: bicycle, bus, car
Text-only
Text + visual
“Riding for a cause.”
Visual objects: person, bicycle
Text/visual
Q: bicycle in street Q: environment friendly traffic
“Biking by the river”
Visual objects: train, piano
Text-only
Text + visual
17/06/2016 1Sreyasi Nag Chowdhury, AKBC 2016
User Query
Concrete Abstract
“Wow! Double-decker buses still run!”
Visual objects: bicycle, bus, car
Text-only
Text + visual
“Riding for a cause.”
Visual objects: person, bicycle
CSK: (riding bicycle, be, environment friendly)
Text/visual
Text + visual + CSK
Q: bicycle in street Q: environment friendly traffic
“Biking by the river”
Visual objects: train, piano
Text-only
Text + visual
17/06/2016 1Sreyasi Nag Chowdhury, AKBC 2016
User Query
Concrete Abstract
“Wow! Double-decker buses still run!”
Visual objects: bicycle, bus, car
Text-only
Text + visual
“Riding for a cause.”
Visual objects: person, bicycle
CSK: (riding bicycle, be, environment friendly)
Text/visual
Text + visual + CSK
Q: bicycle in street Q: environment friendly traffic
“Biking by the river”
Visual objects: train, piano
Text-only
Text + visual
Our contribution
CSK: Where do we get it from?
CSK: How do we use it?
CSK: How to combine noisy signals?
CSK: Does it help?
17/06/2016Sreyasi Nag Chowdhury, AKBC 2016 2
Existing CSK knowledge bases: WordNet, ConceptNet, WebChild, Knowlywood
CSK: WHERE DO WE GET IT FROM?
317/06/2016Sreyasi Nag Chowdhury, AKBC 2016
Existing CSK knowledge bases: WordNet, ConceptNet, WebChild, Knowlywood
Our corpus: Wiki articles from domain ‘tourism’
CSK: WHERE DO WE GET IT FROM?
317/06/2016Sreyasi Nag Chowdhury, AKBC 2016
Existing CSK knowledge bases: WordNet, ConceptNet, WebChild, Knowlywood
Our corpus: Wiki articles from domain ‘tourism’
Pruned by Jaccard Similarity
CSK: WHERE DO WE GET IT FROM?
317/06/2016Sreyasi Nag Chowdhury, AKBC 2016
Existing CSK knowledge bases: WordNet, ConceptNet, WebChild, Knowlywood
Our corpus: Wiki articles from domain ‘tourism’
Pruned by Jaccard Similarity
~22,000 CSK triples
“tourism” “be travel for” “recreation, leisure, family, business purposes”
“people” “fall in” “love”
---------------------------------------------------------------------------------------------------
“the bloody hell” “be” “you”
CSK: WHERE DO WE GET IT FROM?
3
Domain-specific
ReVerb triples
17/06/2016Sreyasi Nag Chowdhury, AKBC 2016
CSK: HOW DO WE USE IT?
Query string: travel with backpack
CSK to expand query
• t1: (tourists, use, travel maps)
• t2: (tourists, carry, backpack)
• t3: (backpack, is a type of, bag)
417/06/2016Sreyasi Nag Chowdhury, AKBC 2016
Query string: travel with backpack
CSK to expand query
• t1: (tourists, use, travel maps)
• t2: (tourists, carry, backpack)
• t3: (backpack, is a type of, bag)
Document x with features
• Textual: “A tourist reading a map by the road”
• Visual: person, bag, bottle, bus
Text-only systems
Text + visual + CSK systems
417/06/2016Sreyasi Nag Chowdhury, AKBC 2016
CSK: HOW DO WE USE IT?
Query string: travel with backpack
CSK to expand query
• t1: (tourists, use, travel maps)
• t2: (tourists, carry, backpack)
• t3: (backpack, is a type of, bag)
Document x with features
• Textual: “A tourist reading a map by the road”
• Visual: person, bag, bottle, bus
Text-only systems
Text + visual + CSK systems
4
CSK bridge vocabulary gap
between query and document
CSK establish relations
between concepts
CSK diminish noise from
modalities – ensemble effect
17/06/2016Sreyasi Nag Chowdhury, AKBC 2016
CSK: HOW DO WE USE IT?
5
A tour group is standing on the grass with ruins in the background.
Group of people standing in front of a stone structure.
Document x
17/06/2016Sreyasi Nag Chowdhury, AKBC 2016
CSK: HOW DO WE USE IT?
5
Document x
Textual features xx
17/06/2016Sreyasi Nag Chowdhury, AKBC 2016
CSK: HOW DO WE USE IT?
A tour group is standing on the grass with ruins in the background.
Group of people standing in front of a stone structure.
5
Document x
Textual features xx
17/06/2016Sreyasi Nag Chowdhury, AKBC 2016
CSK: HOW DO WE USE IT?
A tour group is standing on the grass with ruins in the background.
Group of people standing in front of a stone structure.
Visual features xv : backpack, person
5
Document x
Textual features xx
17/06/2016Sreyasi Nag Chowdhury, AKBC 2016
CSK: HOW DO WE USE IT?
A tour group is standing on the grass with ruins in the background.
Group of people standing in front of a stone structure.
Visual features xv : backpack, bag, container
person, casual agent, organism
5
Document x
Textual features xx
17/06/2016Sreyasi Nag Chowdhury, AKBC 2016
CSK: HOW DO WE USE IT?
A tour group is standing on the grass with ruins in the background.
Group of people standing in front of a stone structure.
Visual features xv : backpack, bag, container
person, casual agent, organism
Query: “group excursion”
Query expansion: (an excursion, be trip by, a group of people)
(organized excursions, book through, a tour company)
CSK features
A tour group is standing on the grass with ruins in the background.
Group of people standing in front of a stone structure.
517/06/2016Sreyasi Nag Chowdhury, AKBC 2016
CSK: HOW DO WE USE IT?
Visual features xv : backpack, bag, container
person, casual agent, organism
Query: “group excursion”
Query expansion: (an excursion, be trip by, a group of people)
(organized excursions, book through, a tour company)
CSK: HOW TO COMBINE NOISY SIGNALS?
6
Mixture LM:
Commonsense-aware LM:
Smoothed LM: , where
Basic LM: , where
17/06/2016Sreyasi Nag Chowdhury, AKBC 2016
CSK: HOW TO COMBINE NOISY SIGNALS?
6
Mixture LM:
Commonsense-aware LM:
Smoothed LM: , where
Basic LM: , where
17/06/2016Sreyasi Nag Chowdhury, AKBC 2016
6
Mixture LM:
Commonsense-aware LM:
Smoothed LM: , where
Basic LM: , where
17/06/2016Sreyasi Nag Chowdhury, AKBC 2016
CSK triple
CSK: HOW TO COMBINE NOISY SIGNALS?
6
Mixture LM:
Commonsense-aware LM:
Smoothed LM: , where
Basic LM: , where
17/06/2016Sreyasi Nag Chowdhury, AKBC 2016
Probabilities based on
word-wise overlaps
CSK: HOW TO COMBINE NOISY SIGNALS?
6
Mixture LM:
Commonsense-aware LM:
Smoothed LM: , where
Basic LM: , where
17/06/2016Sreyasi Nag Chowdhury, AKBC 2016
Background corpus –
Co-occurring Flickr tags
CSK: HOW TO COMBINE NOISY SIGNALS?
6
Mixture LM:
Commonsense-aware LM:
Smoothed LM: , where
Basic LM: , where
17/06/2016Sreyasi Nag Chowdhury, AKBC 2016
Textual and visual features
CSK: HOW TO COMBINE NOISY SIGNALS?
717/06/2016Sreyasi Nag Chowdhury, AKBC 2016
CSK: DOES IT HELP?
Image Dataset
o Flickr30k
o MS COCO captioned dataset
o Pascal Sentence Dataset
o SBU captioned dataset
717/06/2016Sreyasi Nag Chowdhury, AKBC 2016
CSK: DOES IT HELP?
Image Dataset
o Flickr30k
o MS COCO captioned dataset
o Pascal Sentence Dataset
o SBU captioned dataset
Boat trip to see the mythical pink dolphins...
this is John checking in with the office for that day.
717/06/2016Sreyasi Nag Chowdhury, AKBC 2016
CSK: DOES IT HELP?
Image Dataset
o Flickr30k
o MS COCO captioned dataset
o Pascal Sentence Dataset
o SBU captioned dataset
A group of tourists is crossing a bridge that connects a walking path to a trail of nature.
Many people cross a very tall footbridge with a tree-covered hill in the background.
This shows a group of people walking over an arched red bridge.
People cross a large bridge to get over the body of water.
People walking over a white and red bridge over a pond.
Boat trip to see the mythical pink dolphins...
this is John checking in with the office for that day.
717/06/2016Sreyasi Nag Chowdhury, AKBC 2016
CSK: DOES IT HELP?
Image Dataset
o Flickr30k
o MS COCO captioned dataset
o Pascal Sentence Dataset
o SBU captioned dataset
A group of tourists is crossing a bridge that connects a walking path to a trail of nature.
Many people cross a very tall footbridge with a tree-covered hill in the background.
This shows a group of people walking over an arched red bridge.
People cross a large bridge to get over the body of water.
People walking over a white and red bridge over a pond.
Boat trip to see the mythical pink dolphins...
this is John checking in with the office for that day.
social media post
blog post
717/06/2016Sreyasi Nag Chowdhury, AKBC 2016
CSK: DOES IT HELP?
Image Dataset
o Flickr30k
o MS COCO captioned dataset
o Pascal Sentence Dataset
o SBU captioned dataset
~ 50,000 images with captions
A group of tourists is crossing a bridge that connects a walking path to a trail of nature.
Many people cross a very tall footbridge with a tree-covered hill in the background.
This shows a group of people walking over an arched red bridge.
People cross a large bridge to get over the body of water.
People walking over a white and red bridge over a pond.
Boat trip to see the mythical pink dolphins...
this is John checking in with the office for that day.
social media post
blog post
Baselines: Text-only and Text + Visual search approaches
Evaluation metric: Average Precision @ 10
817/06/2016Sreyasi Nag Chowdhury, AKBC 2016
CSK: DOES IT HELP?
Baselines: Text-only and Text + Visual search approaches
Evaluation metric: Average Precision @ 10
817/06/2016Sreyasi Nag Chowdhury, AKBC 2016
CSK: DOES IT HELP?
47% 64% 85%
Text-only Text + Visual Text + Visual + CSK (Know2Look)
Baselines: Text-only and Text + Visual search approaches
Evaluation metric: Average Precision @ 10
Examples queries:
o Concrete – ball park, bridge road, table home, bicycle road
o Abstract – diesel transport, housing town
o Mixed – old clock, backpack travel, boat tour
817/06/2016Sreyasi Nag Chowdhury, AKBC 2016
CSK: DOES IT HELP?
47% 64% 85%
Text-only Text + Visual Text + Visual + CSK (Know2Look)
Baselines: Text-only and Text + Visual search approaches
Evaluation metric: Average Precision @ 10
Examples queries:
o Concrete – ball park, bridge road, table home, bicycle road
o Abstract – diesel transport, housing town
o Mixed – old clock, backpack travel, boat tour
817/06/2016Sreyasi Nag Chowdhury, AKBC 2016
CSK: DOES IT HELP?
47% 64% 85%
Text-only Text + Visual Text + Visual + CSK (Know2Look)
Co-occurring Flickr tags
917/06/2016Sreyasi Nag Chowdhury, AKBC 2016
Text-only Text + Visual Text + Visual + CSK (Know2Look)
Query: “group excursion”
xxj: “A small excursion boat anchored
on the beach at the resort in Mexico. "
xvj: lunar excursion module, conveyance xxj: “A group of people riding camels.“
yk: (an excursion, be trip by,
a group of people)
CSK: DOES IT HELP?
917/06/2016Sreyasi Nag Chowdhury, AKBC 2016
Text-only Text + Visual Text + Visual + CSK (Know2Look)
Query: “group excursion”
xxj: “A small excursion boat anchored
on the beach at the resort in Mexico. "
xvj: lunar excursion module, conveyance xxj: “A group of people riding camels.“
yk: (an excursion, be trip by,
a group of people)
CSK: DOES IT HELP?
Noisy OpenIE triples capture commonsense knowledge
Noisy textual cues + noisy visual object detection + noisy commonsense knowledge ensemble effect better results for multimodal document retrieval
CSK act as bridge between text and vision
1017/06/2016Sreyasi Nag Chowdhury, AKBC 2016
Noisy OpenIE triples capture commonsense knowledge
Noisy textual cues + noisy visual object detection + noisy commonsense knowledge ensemble effect better results for multimodal document retrieval
CSK act as bridge between text and vision
Do word co-occurrences or word embeddings provide similar results?
Does structured commonsense knowledge improve retrieval?
1017/06/2016Sreyasi Nag Chowdhury, AKBC 2016
Noisy OpenIE triples capture commonsense knowledge
Noisy textual cues + noisy visual object detection + noisy commonsense knowledge ensemble effect better results for multimodal document retrieval
CSK act as bridge between text and vision
Do word co-occurrences or word embeddings provide similar results?
Does structured commonsense knowledge improve retrieval?
1017/06/2016Sreyasi Nag Chowdhury, AKBC 2016
Thank you