💬 NLP

Graph-Augmented Retrieval for Cross-Entity Financial Sentiment Analysis: A Comparative Study

This paper demonstrates that a novel two-hop Graph-RAG architecture significantly outperforms standard vector-based RAG in cross-entity financial sentiment analysis by improving entity recall and answer relevance for complex relational queries without compromising overall answer quality, while identifying an optimal graph traversal intensity threshold to balance precision and coverage.

Original authors: Rajan Bastakoti, Sagar Bhetwal, Nirajan Acharya, Gaurav Kumar Gupta

Published 2026-06-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Rajan Bastakoti, Sagar Bhetwal, Nirajan Acharya, Gaurav Kumar Gupta

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 are a financial analyst trying to understand the stock market. You have a massive library of news articles, and you want to ask a smart computer (an AI) questions like, "How might a bad day for Apple affect Microsoft?"

This paper compares two different ways to build that smart computer.

The Two Competitors

1. The "Vector-Only" System (The Standard Approach)
Think of this system as a very fast librarian who only looks at the book titles and the first few sentences.

How it works: When you ask a question, it scans the library for words that sound similar to your question. If you ask about "Apple," it finds articles with the word "Apple."
The Flaw: It treats every article as an isolated island. It doesn't know that Apple and Microsoft are neighbors in the business world. If you ask about how one affects the other, this librarian might miss the connection because the specific words linking them aren't in the exact same sentence. It's like trying to understand a family feud by only reading one person's diary without ever talking to the other family members.

2. The "Graph-RAG" System (The New Approach)
Think of this system as a detective with a giant, glowing web of connections.

How it works: Before answering, the detective builds a map (a "Knowledge Graph") of 59 different companies. It draws lines between them showing who influences whom.
The Strategy:
- Hop 1: It finds the main article about your question (just like the librarian).
- Hop 2: It follows the glowing lines on the map to see who that company is connected to. If the news is about Apple, the detective immediately checks the "influence lines" to see what's happening at Microsoft or Google, even if the article didn't explicitly mention them together.
The Result: It gathers a "ripple effect" of information, showing how sentiment (good or bad feelings) spreads from one company to another.

The Big Experiment

The researchers tested both systems with 100 specific questions about 10 major tech stocks (like Apple, Google, and Tesla). They asked two types of questions:

Direct Questions: "What was Apple's revenue?" (Simple facts).
Relational Questions: "How did a supply chain issue at one company impact its competitors?" (Complex connections).

What They Found

1. The "Relational" Win
The Graph-RAG system was a huge winner for complex questions.

The Analogy: If you ask the librarian about a family feud, they might miss the drama. But the detective with the map sees the whole picture.
The Numbers: For questions about how companies affect each other, the Graph-RAG system was 16% better at giving relevant answers. It found important connections that the standard system completely missed.

2. The "Direct" Tie
For simple questions like "What is the stock price?", both systems performed almost exactly the same.

The Takeaway: The fancy detective map doesn't hurt you on simple tasks, but it doesn't add much value there either. The librarian is fast and accurate for simple facts.

3. The Cost of Being Smart
There is a trade-off.

Speed: The Graph-RAG system was about 22% slower.
- Why? The librarian grabs a book instantly. The detective has to walk over to the map, trace the lines, check the neighbors, and then grab those books too.
Consistency: However, the Graph-RAG system was much more consistent. The librarian's speed varied wildly (sometimes fast, sometimes slow), while the detective's speed was steady and predictable.

4. The "Goldilocks" Setting
The researchers tested how "strict" the detective should be when following the map lines.

If the detective followed every weak connection, they got too much noise (irrelevant info).
If they were too strict, they missed good info.
The Sweet Spot: They found a "Goldilocks" setting (a specific threshold of 0.5) where the detective found the perfect balance of useful connections without getting distracted.

The Bottom Line

This paper proves that for financial analysis, structure matters.

If you just want to look up a fact, a standard search is fine. But if you want to understand the complex web of how one company's news ripples out to affect others, you need a system that can "walk the map" (Graph-RAG). It finds the hidden connections that simple word-matching misses, giving you a much clearer picture of the market, even if it takes a little longer to get the answer.

Technical Summary: Graph-Augmented Retrieval for Cross-Entity Financial Sentiment Analysis

Problem Statement

The rapid digitization of financial markets has generated an unprecedented volume of unstructured textual data, creating a challenge for real-time equity synthesis. While Retrieval-Augmented Generation (RAG) frameworks have become foundational for grounding Large Language Models (LLMs) in domain-specific corpora, conventional "vector-only" RAG architectures face a "contextual plateau" in specialized financial analysis.

The paper identifies three critical limitations of standard vector-based retrieval in this domain:

Lack of Structural Awareness: Vector retrieval treats text as isolated chunks, failing to capture explicit structured relationships (e.g., supply-chain cascades, multi-hop influence networks) necessary for systemic risk analysis.
Contextual Dilution: Concatenating multiple semantically similar snippets often leads to the "lost in the middle" phenomenon, where crucial details are obscured within redundant, high-token contexts.
Semantic Noise: Financial terminology is highly susceptible to noise; without a structured graph to logically separate entities, semantic overlap in dense retrieval can result in factual errors.

Furthermore, purely lexical analysis often overlooks "sentiment contagion"—the propagation of sentiment across documents and entities—which is critical for understanding market dynamics.

Methodology

The authors propose and evaluate a two-hop Graph-RAG architecture designed to augment dense vector retrieval with structured graph traversal. The system is built upon a corpus of 255 news articles covering 10 major technology stocks, resulting in 1,477 distinct news chunks and a knowledge graph of 59 equity entities.

System Architecture

The methodology transitions from automated data ingestion to a comparative retrieval study involving two distinct pipelines:

Data Ingestion & Graph Construction:
- Entity Extraction: The gliner-bi-large-v2.0 model performs zero-shot recognition of entities (Stock Ticker, Equity Index, Investment Bank, Organization).
- Sentiment Weighting: FinBERT calculates sentiment polarity. The relational Intensity ( $I$ ) between entities is defined as the absolute difference between positive and negative scores ( $I = |Score_{pos} - Score_{neg}|$ ).
- Graph Topology: A Neo4j graph is constructed where Ticker nodes connect to News Chunk nodes via HAS_NEWS_CHUNK edges. Inter-ticker relationships are defined by directed INFLUENCES edges, weighted by sentiment intensity.
Baseline Architecture (General RAG):
- A standard vector-only system using FAISS for indexing.
- Retrieves the top $k=5$ chunks based on maximum cosine similarity.
- Context is a flat concatenation of retrieved text without relational metadata.
Proposed Architecture (2-Hop Graph-RAG):
- Hop 1 (Semantic Anchor): Vector search identifies top $k=5$ "seed" chunks.
- Hop 2 (Structural Expansion): From the ticker associated with each seed, the system traverses INFLUENCES edges where Intensity $> \tau$ (threshold).
- Neighbor Re-ranking: Associated news chunks for neighboring entities are retrieved and re-ranked against the query. Only the top 2 chunks per neighbor with a similarity score $> 0.2$ are added as "Network Evidence."
- Synthesis: The LLM receives a context containing primary news, connected entity names, numerical sentiment scores, and re-ranked evidence.

Evaluation Framework

The study evaluates both architectures on 100 grounded queries (30 Direct, 70 Relational) using:

Statistical Metrics: Semantic similarity (cosine similarity) and Entity Recall, validated via Wilcoxon signed-rank tests.
RAGAS Framework: Reference-free metrics including Faithfulness, Answer Relevancy, Context Precision, and Context Recall.
Ablation Study: Analysis of the intensity threshold ( $\tau$ ) on answer quality.

Key Results

1. Overall Performance

Entity Recall: Graph-RAG achieved a statistically significant improvement of +6.4% ( $p < 0.001$ ) over the baseline, confirming that structural traversal surfaces entities missed by pure vector similarity.
Answer Relevancy: Graph-RAG improved by +11.7%, suggesting that graph-augmented context helps LLMs produce more topically focused responses.
Semantic Similarity: There was no statistically significant difference in answer quality ( $\Delta = +0.001$ , $p = 0.281$ ), indicating that the additional graph context neither improved nor degraded the core generation quality.
Latency: Graph-RAG incurred a 22.6% increase in mean latency (9.03s vs. 7.37s) but demonstrated an 80% reduction in latency variance ( $\sigma = 0.96$ s vs. 4.82s), offering more deterministic response times.

2. Stratified Analysis by Question Type

The benefits of Graph-RAG are highly concentrated in Relational (cross-entity) queries:

Relational Queries: Answer Relevancy improved by +16.1%.
Direct (Single-entity) Queries: Performance was comparable between systems, with Graph-RAG showing a marginal gain in Context Precision (+8.6%) but a trade-off in Context Recall (-11.9%), as additional entities can dilute single-article evidence.

3. Ablation Study: Intensity Threshold ( $\tau$ )

The study identified an inverted-U relationship between the intensity threshold and answer quality:

$\tau = 0.3$ : Too permissive; introduces irrelevant evidence, degrading quality.
$\tau = 0.7$ : Too restrictive; misses moderately informative relationships.
$\tau = 0.5$ : Identified as the optimal balance, outperforming the production default of 0.7 by +0.5% in semantic similarity.

Significance and Claims

The paper characterizes the findings as a precision-for-coverage trade-off inherent to graph-augmented retrieval. The authors claim that Graph-RAG does not uniformly outperform vector-only retrieval but shifts the operating point to better serve complex, multi-entity reasoning tasks common in financial analysis.

Key contributions include:

Demonstrated Efficacy: Proof that integrating a specialized influence network allows systems to identify "market dissonance"—instances where textual narratives are corrected or enriched by underlying metadata.
Architectural Guidance: The identification of the inverted-U relationship for intensity thresholds provides actionable tuning guidance for practitioners, suggesting that extreme filtering values are suboptimal.
Query-Type Awareness: The results suggest that production systems should be query-type-aware, routing relational queries to Graph-RAG pipelines while reserving vector-only retrieval for simple factual lookups to optimize latency and token consumption.

The authors note limitations, including the use of LLM-generated ground truth (which may introduce a ceiling effect), the specific scale of the 10-stock dataset, and the latency overhead of cloud-based Neo4j deployments. They conclude that while the architecture is domain-agnostic, further validation is required for generalization to other domains and larger entity graphs.