“Is depression related to cannabis?”: A knowledge-infused model for entity and relation extraction with limited Supervision

Why it matters

• Cannabis is frequently discussed as a potential mental health aid, but scientific evidence linking cannabis and depression remains inconclusive.
• Twitter contains large volumes of self-reported experiences, yet extracting structured relationships from informal text is technically difficult.
• Standard deep learning models require large annotated datasets, which are costly when domain experts are involved.
• In this study, only 3,000 of 11,000 tweets were expert-labeled (Cohen's kappa = 0.8), creating a limited supervision setting.

What we did

• We propose a knowledge-infused relation extraction model that combines Drug Abuse Ontology (315 entities, 31 relations), DSM-5 and related clinical lexicons, GPT-3 embeddings, and supervised contrastive learning with triplet loss.
• We extract and classify three relationships between cannabis and depression: Reason, Effect, and Addiction.
• The model reaches an F1 score of 75.37, reported as a +11.28 point gain over the strongest baseline.
• Ablation results show removing contrastive learning reduces F1 by 6.46 points, and removing knowledge infusion reduces F1 by 9.01 points.
• The learned representation space is then used to label the remaining ~7,000 tweets.

As shown in Table I, the knowledge-infused contrastive approach outperforms all baselines.

How it works

• Knowledge-guided phrase extraction: Map tweet n-grams to cannabis and depression entities using ontology matching (cosine similarity >= 0.75).
• Contextual embeddings: Use GPT-3 to obtain phrase representations.
• Supervised contrastive learning: Train anchor-positive-negative triplets to separate relation classes in embedding space.
• Weak supervision extension: Label the remaining ~7,000 tweets using the learned metric and clustering.

The extraction pipeline (Figure 1) shows how ontology-matched phrases are embedded before relation classification.

Key contributions

• A knowledge-infused neural model for cannabis-depression relation extraction.
• Integration of GPT-3 embeddings with supervised contrastive learning under limited supervision.
• An absolute +11.28 F1 improvement over the strongest BERT-based baseline.
• Release of an annotated dataset covering 3,000 expert-labeled tweets and model-labeled data for the full 11,000 tweet set.