feiyang-k/curation-bench/contrastive-language-image-pre-training-with-knowledge-graphs-arxiv-2210-08901v2

Contrastive Language-Image Pre-training with Knowledge Graphs

One-line decision

Use this skill when you want to augment CLIP training data with knowledge graph information for improved conceptual understanding. Avoid it when you do not have access to knowledge graphs or standard CLIP training is sufficient.

Skill metadata

• Skill type: knowledge-augmented-pretraining
• Paper kind: operational-method
• Actionability: medium
• Evidence quality: full_paper

Goal

Augment CLIP training with knowledge graph information, enriching image-text pairs with structured knowledge about entities, relationships, and hierarchies for improved conceptual understanding.

Problem signature

• Modality: image-text pairs augmented with knowledge graph entity and relationship information.
• Data state: standard image-text pairs enriched with knowledge graph annotations.
• Scale regime: standard CLIP-scale data augmented with knowledge.
• Model requirement: CLIP architecture with knowledge graph integration.

Use when

• You want to improve CLIP's conceptual and relational understanding.
• You have access to knowledge graphs (Wikidata, ConceptNet, etc.).
• You want richer semantic representations.

Do not use when

• Standard CLIP training is sufficient.
• You lack access to knowledge graphs.
• Simple contrastive learning meets your needs.

Required inputs

• image_text_pairs: Standard image-text pairs for CLIP training.
• knowledge_graph: Structured knowledge graph (Wikidata, ConceptNet).
• entity_linker: Tool for linking text mentions to knowledge graph entities.

Optional inputs

• relationship_embeddings: Pre-computed knowledge graph embeddings.

Outputs

• knowledge_augmented_data: Image-text pairs enriched with knowledge graph information.
• knowledge_clip: CLIP model with improved conceptual understanding.

Assumptions and prerequisites

• Knowledge graphs provide structured information not in raw text.
• Entity linking can connect text mentions to knowledge graph entries.
• Knowledge augmentation improves conceptual understanding.

Procedure

1. Link text to knowledge graph

Action: Use entity linking to connect caption mentions to KG entities. Why: Linking enables knowledge injection. Note: See paper for details.

1. Extract relevant knowledge

Action: Retrieve relationships and properties for linked entities. Why: Knowledge enriches the training signal. Note: See paper for details.

1. Augment training data

Action: Add knowledge graph information to the training pairs. Why: Augmented data teaches structured relationships. Note: See paper for details.

1. Train knowledge-augmented CLIP

Action: Train CLIP with the enriched training data. Why: Knowledge-augmented training improves conceptual understanding. Note: See paper for details.

Parameters to set

• knowledge_source — Role: Knowledge graph to use. How to set: Wikidata for broad coverage, ConceptNet for commonsense. Default/range: Wikidata or ConceptNet. Effect: Different KGs provide different knowledge types.
• augmentation_method — Role: How knowledge is integrated. How to set: Add KG entities to text or use separate KG embeddings. Default/range: Text augmentation. Effect: Integration method affects learning.

Validation checks

• Knowledge-augmented CLIP should improve on knowledge-intensive tasks.
• Conceptual understanding should be measurably better.
• Standard vision-language tasks should not degrade.

Failure modes

• Entity linking errors introduce noise.
• Knowledge graph coverage may be incomplete.
• The augmentation may add irrelevant information.

Adaptation notes for VLM training

• Knowledge augmentation can improve VLM understanding of entities and relationships.
• Apply to VLM pretraining data for richer semantic content.
• Combine with standard caption augmentation.

Implementation notes

• Use efficient entity linking tools (spaCy, GENRE).
• Cache knowledge graph lookups for efficiency.
• Monitor the quality of entity linking.

Evidence from the paper

• Knowledge graph augmentation improves CLIP's conceptual understanding.
• Entity linking connects text to structured knowledge.
• The approach improves zero-shot classification on knowledge-intensive categories.
• Knowledge-augmented CLIP better understands relationships and hierarchies.

Source paper

• Title: Contrastive Language-Image Pre-training with Knowledge Graphs
• Year: 2022
• Venue: NeurIPS
• Paper ID: arxiv-2210.08901v2
• URL: http://arxiv.org/abs/2210.08901v2
• arXiv ID: 2210.08901v2