💬 NLP

DKCD: Domain Knowledge-Enhanced Causal Discovery from Unstructured Data

This paper introduces DKCD, a novel framework that enhances causal discovery from unstructured data in high-expertise domains by integrating domain knowledge to improve the identification of latent factors and the accuracy of causal graph construction.

Original authors: Xin Li, Jin Li, Shoujin Wang, Kun Yu, Fang Chen

Published 2026-07-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xin Li, Jin Li, Shoujin Wang, Kun Yu, Fang Chen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you're trying to solve a massive, messy mystery: figuring out what causes what in the real world, like why some people get diabetes or why others struggle with breathing issues. Usually, detectives (scientists) need neat, organized clue cards (structured data) to solve these cases. But in the real world, clues are often hidden inside long, messy paragraphs of text (unstructured data), like a doctor's handwritten notes or a patient's story.

For a while, we've been using super-smart AI robots called Large Language Models (LLMs) to read these messy stories and guess the clues. Think of these robots as brilliant generalists who know a little bit about everything. But here's the catch: when the mystery gets really specific—like deep in the medical field—these generalist robots hit a wall. They miss hidden clues that only a true expert would know, and sometimes they guess the wrong connections between clues because they lack that deep, specialized background.

The paper introduces a new detective team called DKCD (Domain Knowledge-Enhanced Causal Discovery). Instead of letting the robot guess alone, DKCD gives the robot a "cheat sheet" made of a Domain Knowledge Graph. Imagine this graph as a giant, interconnected map of expert facts specific to that field (like a map of how kidneys, genes, and sugar levels are all linked in diabetes).

Here is how DKCD solves the mystery in three fun steps:

The Knowledge Hunt: First, the robot reads the messy text and pulls out the obvious clues it can see (like "smoking" or "age"). Then, it uses the expert map to find related hidden clues. It's like finding a footprint in the mud and then checking a map to realize, "Oh, this footprint is near a river, so maybe the culprit was fishing!" This step finds "latent factors"—hidden causes the robot would have missed on its own, like "kidney disease" or "genetic risk," which aren't always written down explicitly but are implied by other symptoms.
The Reasoning Session: Next, the robot uses the expert map to figure out how these clues connect. It doesn't just guess; it reasons based on the map. This helps it write down the clues more accurately, avoiding mistakes where it might think "smoking" causes "diabetes" directly, when actually, the map shows it's more complex. This step creates "causal clues"—little hints that guide the robot to score the data correctly.
The Final Map: Finally, with a complete list of clues and accurate scores, the team uses a standard statistical tool to draw the final "causal graph." This is the ultimate map showing exactly what causes what.

What the paper found:
The authors tested DKCD on two made-up but realistic medical datasets: one with 400 patient stories about diabetes (involving 14 different factors) and another with 400 stories about respiratory issues (involving 8 factors). They compared DKCD against other methods, including the popular "COAT" framework and robots just guessing on their own.

The results showed that DKCD was significantly better at finding the right clues and drawing the correct maps. For example, on the diabetes dataset using a top-tier robot (GPT-4o), DKCD found the right factors about 84% of the time (Node Precision), while the previous best method only got about 61%. When it came to drawing the final map of connections, DKCD made fewer mistakes, with an error score (ESHD) of 25.33, compared to 30.67 for the older method. These numbers suggest that adding the expert map really helps the robot see the whole picture, not just the obvious parts.

What the paper rules out:
The paper explicitly argues against the idea that a smart robot alone is enough for these high-expertise fields. It shows that without the external "cheat sheet" (the domain knowledge graph), robots miss crucial hidden factors and make unreliable guesses. It also notes that simply having more data isn't the answer; the quality of the reasoning, guided by expert knowledge, is what matters.

How sure are they?
The authors are confident in these results, but they are careful to say they are based on simulations and synthetic datasets. They created these 400-patient datasets carefully based on real medical knowledge and ground-truth maps to test their system. They didn't test this on real, live patients in a hospital yet. So, while the numbers look great in these controlled tests, the paper suggests this is a promising new direction rather than a solved problem for the real world. They admit that the robots can still sometimes make up facts (hallucinate) or get biased, and that building even bigger, real-world test cases is a job for the future.

In short, DKCD is like giving a brilliant detective a specialized field guide. It doesn't replace the detective's brain, but it ensures they don't miss the hidden clues that only an expert would know, leading to a much clearer picture of cause and effect.

Technical Summary: DKCD – Domain Knowledge-Enhanced Causal Discovery from Unstructured Data

Problem Definition

Causal discovery in high-expertise domains (e.g., healthcare, finance, education) faces significant challenges when applied to unstructured data, such as clinical text. Existing methods, including those leveraging Large Language Models (LLMs) like the COAT framework, are limited by two primary challenges:

Insufficient Identification of Latent Factors (CH1): Unstructured data often contains implicit causal factors that are not explicitly stated but are essential for a complete causal graph. General LLM knowledge lacks the specific domain expertise required to infer these latent variables (e.g., inferring "kidney disease" from specific lab values in a diabetes context).
Unreliable Factor Annotation (CH2): Accurate causal graph construction requires structured input (factor values). Without domain-grounded reasoning, LLMs often produce inconsistent or inaccurate annotations, leading to errors that propagate into the final causal structure, such as missing edges or incorrect directions.

Current causal discovery algorithms (e.g., PC, FCI) are designed for structured tabular data and struggle with the noise and incompleteness inherent in raw unstructured text.

Methodology: The DKCD Framework

The authors propose DKCD (Domain Knowledge-enhanced Causal Discovery), a framework that integrates domain Knowledge Graphs (KGs) with LLM reasoning to bridge the gap between unstructured text and statistical causal discovery. The framework operates through three interconnected components:

1. Knowledge Mining

This module prepares the context for reasoning by retrieving relevant domain knowledge.

Observable Factor Proposal: An LLM extracts explicitly mentioned causal factors ( $V_o$ ) from the unstructured text.
Domain KG Retrieval: Based on $V_o$ , the system queries a domain-specific Knowledge Graph to retrieve relevant subgraphs ( $G_{rel}$ ).
Semantic Matching: The retrieved triples are verbalized into natural language statements. A semantic matching process (using sentence transformers like all-MiniLM-L6-v2) aligns these statements with the observable factors, filtering for the most relevant context ( $S^{(k)}_{v_i}$ ) to support subsequent reasoning.

2. Knowledge-Guided Causal Reasoning

This core module addresses the two key challenges by leveraging the retrieved knowledge context.

Latent Factor Discovery: The LLM uses the observable factors and the retrieved knowledge context to infer latent causal factors ( $V_l$ ) that are not explicitly stated but are causally relevant. The complete factor set is defined as $V = V_o \cup V_l$ . This directly addresses CH1.
Causal Clue Generation: The LLM generates causal clues ( $H$ ), which are domain-grounded descriptions of plausible dependencies between factors. These clues serve as structured guidance for the annotation step, reducing ambiguity and improving the reliability of factor scoring. This addresses CH2.

3. Causal Structure Discovery

With a more complete factor set ( $V$ ) and a structured scoring table ( $S$ ) containing accurate annotations derived from the causal clues, a statistical causal discovery algorithm (specifically the FCI algorithm) is applied to infer the final causal graph ( $G$ ). The FCI algorithm is chosen for its ability to handle latent confounders, which is critical in specialized domains where not all factors may be fully observed.

Key Contributions

The paper outlines three primary contributions:

DKCD Framework: A novel framework that integrates domain Knowledge Graphs with LLM-based reasoning to enhance causal discovery from unstructured data in high-expertise domains.
Knowledge-Guided Reasoning Module: A specific design that utilizes retrieved domain knowledge to discover latent causal factors and generate knowledge-grounded causal clues, thereby improving annotation accuracy.
Domain-Specific Benchmarks: The construction of the first domain-specific benchmarks utilizing real-world knowledge graphs for evaluation, specifically for the diabetes and respiratory domains.

Experimental Results

The authors evaluated DKCD on two synthetic medical datasets (Diabetes and Respiratory) containing 400 patient condition descriptions each, using multiple LLM backbones (GPT-4o, Gemini 2.5 Flash, Grok-3, LLaMA 3-70B).

Performance Metrics: Evaluation included Node Precision/Recall/F1 (for factor identification) and Adjacency Precision/Recall/F1, plus Extended Structural Hamming Distance (ESHD) (for graph structure).
Comparative Results: DKCD consistently outperformed baselines, including Zero-shot LLM, META, and COAT.
- On the Diabetes dataset, DKCD achieved higher Node F1 scores (e.g., 0.77 vs. 0.41 for COAT with GPT-4o) and lower ESHD values (25.33 vs. 30.67 for COAT), indicating a closer alignment with the ground-truth causal graph.
- On the Respiratory dataset, DKCD similarly demonstrated superior factor identification (e.g., 0.74 Node F1 vs. 0.29 for COAT with GPT-4o) and structural accuracy.
Ablation Studies: Removing either the causal clues (w/o CC) or the latent factor identification (w/o LF) significantly degraded performance. Removing both resulted in severe performance drops, confirming that the two components are complementary and essential for the framework's success.

Significance and Claims

The paper claims that DKCD significantly improves both causal factor identification and causal graph construction in high-expertise domains by effectively leveraging domain knowledge graphs to guide LLM reasoning. The authors emphasize that their work addresses the critical limitation of existing methods: the inability to identify implicit, domain-specific latent factors and the unreliability of annotations without external knowledge grounding.

The authors maintain a modest stance regarding limitations:

Reliance on LLMs: The framework still depends on LLM reasoning capabilities, which may generate hallucinated or biased factors.
Benchmark Availability: The reliance on synthetic datasets is acknowledged as a necessity due to the lack of public benchmarks containing both unstructured text and aligned ground-truth causal graphs.
Scalability: While tested in medical domains, generalization to other fields (finance, education) remains underexplored.

The work is positioned as a methodological advancement for constructing causal graphs from unstructured data, intended to assist researchers in discovering potential causal structures rather than serving as a substitute for professional judgment in sensitive domains.