Andrii Bidochko

• Updated: June 19, 2026
• 7 min read

Developing an Intelligent Job Recommendation System Using Semantic Retrieval and Explainable AI Techniques

Direct Answer

The paper introduces a **metadata‑driven job recommendation system** that fuses lexical TF‑IDF matching with semantic Sentence‑BERT retrieval, adds query‑aware filtering, optional Cross‑Encoder re‑ranking, and generates human‑readable explanations. This matters because it delivers high‑precision, explainable job matches without needing full job descriptions or historic user interaction data—an often‑missing piece in enterprise recruitment pipelines.

Background: Why This Problem Is Hard

Online recruitment platforms must sift through millions of postings that differ in wording, structure, and granularity. Traditional keyword search is fast and transparent, yet it collapses when the same role is described with synonyms or industry‑specific jargon. Conversely, deep‑learning recommenders that ingest full text can capture nuance but become opaque, demand extensive computational resources, and rely on rich interaction histories that many B2B HR systems lack.

Key bottlenecks include:

• Heterogeneous metadata: Titles, seniority levels, locations, and industry tags are often the only reliable signals available.
• Terminology drift: “Software Engineer” vs. “Developer” vs. “Full‑Stack Engineer” can fragment relevance.
• Explainability requirements: HR managers need to justify why a candidate sees a particular posting, especially for compliance and internal audit.
• Scalability constraints: Enterprise‑grade systems must return results in sub‑second latency across global job catalogs.

Existing solutions either sacrifice relevance for speed (pure TF‑IDF) or sacrifice transparency for performance (black‑box neural rankers). The research targets this trade‑off head‑on.

What the Researchers Propose

The authors present a **hybrid retrieval framework** that treats structured job metadata as the primary knowledge base. The system consists of five cooperating modules:

1. Lexical Retrieval (TF‑IDF): Generates a fast, interpretable score based on exact token overlap in fields such as title and company.
2. Semantic Retrieval (Sentence‑BERT): Embeds metadata into a dense vector space, enabling similarity matching across paraphrased expressions.
3. Query‑Aware Filtering: Applies rule‑based constraints (e.g., seniority, location) to prune candidates before expensive scoring.
4. Cross‑Encoder Re‑ranking (optional): A transformer that jointly encodes query and candidate metadata to produce a refined relevance score.
5. Explanation Generation: Constructs natural‑language rationales by tracing back the lexical and semantic contributions that led to a recommendation.

Each component is deliberately lightweight, allowing the pipeline to operate on a pure metadata feed while still delivering state‑of‑the‑art relevance.

How It Works in Practice

Conceptual Workflow

The end‑to‑end process can be visualized as a staged funnel:

1. Input Query Construction: An HR user types a free‑form search (e.g., “mid‑level data scientist in Berlin”). The system extracts key entities (role, seniority, location).
2. Lexical Pre‑filter: TF‑IDF scores are computed across the entire catalog; the top‑k (e.g., 500) postings are retained.
3. Semantic Expansion: Sentence‑BERT embeddings for the same k candidates are compared to the query embedding; the top‑n (e.g., 200) are forwarded.
4. Query‑Aware Filtering: Business rules enforce exact matches on mandatory fields (e.g., employment type = “full‑time”).
5. Optional Cross‑Encoder Re‑ranking: The remaining set (often < 50) is re‑scored with a cross‑encoder that captures cross‑field interactions.
6. Explanation Layer: For each final recommendation, the system highlights which lexical tokens and semantic dimensions contributed most, producing a concise sentence like “Matched because the title ‘Data Scientist’ aligns with your query and the seniority level is ‘Mid‑Level’.”

Interaction Between Components

Because each stage reduces the candidate pool, computationally intensive steps (semantic retrieval, cross‑encoder) are only invoked on a manageable subset. The lexical stage guarantees interpretability, while the semantic stage recovers relevance lost to synonymy. The optional re‑ranking acts as a “quality‑boost” knob that can be toggled based on latency budgets.

What Makes This Approach Different

• Metadata‑Only Dependency: No reliance on full job descriptions or historical click‑through data.
• Explainable Hybrid Scoring: Users receive both a numeric relevance rank and a human‑readable justification.
• Modular Design: Each retrieval block can be swapped (e.g., replace Sentence‑BERT with a newer encoder) without redesigning the whole pipeline.
• Scalable Latency: The cascade architecture keeps average query time well under one second on a 30k‑record catalog.

Evaluation & Results

Experimental Setup

The authors cleaned a public LinkedIn job posting dataset, ending with **31,262** structured records. Each record contained fields for title, company, location, seniority, function, employment type, and industry. They split the data into 80 % training (for encoder fine‑tuning) and 20 % test (for evaluation).

Four configurations were benchmarked:

• Pure TF‑IDF
• Pure Sentence‑BERT
• Hybrid (TF‑IDF + Sentence‑BERT) without re‑ranking
• Hybrid + Cross‑Encoder re‑ranking

Key Findings

Performance was measured with **Precision@10** (fraction of top‑10 results that are truly relevant) and **nDCG@10** (discounted cumulative gain that accounts for ranking order). The hybrid system without re‑ranking achieved a Precision@10 of **0.8032** and an nDCG@10 of **0.9496**, outperforming both pure lexical and pure semantic baselines by a wide margin.

Adding the optional Cross‑Encoder nudged Precision@10 to **0.7948** and lifted nDCG@10 to **0.9739**, demonstrating that a modest computational investment can still improve ranking quality, especially for the most critical top positions.

Beyond raw numbers, the explanation module was evaluated qualitatively: HR reviewers reported that the generated rationales reduced the time needed to validate a match by roughly **30 %**, and increased confidence in automated recommendations.

Why This Matters for AI Systems and Agents

For AI‑driven recruitment platforms, the study offers a blueprint for building **high‑precision, explainable agents** that can operate under strict data‑privacy constraints (no user‑level interaction logs). The hybrid retrieval pattern aligns with the emerging practice of **retrieval‑augmented generation (RAG)**, where a lightweight lexical filter feeds a semantic encoder, which then informs a downstream LLM or decision engine.

Practical implications include:

• Reduced Engineering Overhead: Teams can rely on readily available metadata rather than constructing costly pipelines to ingest full text.
• Compliance‑Ready Recommendations: Explainability satisfies GDPR‑style audit trails, a growing requirement for enterprise HR software.
• Agent Orchestration Flexibility: The modular stages can be orchestrated by workflow engines such as the Workflow automation studio, enabling dynamic scaling based on query load.
• Improved User Trust: When a recommendation is accompanied by a clear “why,” recruiters are more likely to adopt AI suggestions, accelerating digital transformation.

What Comes Next

While the results are compelling, several limitations remain:

• The system assumes **high‑quality, consistently populated metadata**; noisy or missing fields can degrade performance.
• Cross‑Encoder re‑ranking, though effective, still adds latency that may be prohibitive for real‑time chat‑based assistants.
• Explainability is currently limited to token‑level highlights; richer narrative explanations could further aid decision makers.

Future research directions could explore:

1. Integrating **user‑profile embeddings** when privacy permits, to personalize relevance beyond pure job attributes.
2. Applying **prompt‑based LLMs** to generate more nuanced explanations, possibly leveraging the OpenAI ChatGPT integration for natural language synthesis.
3. Extending the pipeline to **multilingual job catalogs**, using multilingual Sentence‑BERT variants.
4. Embedding the framework within a **full‑stack AI recruitment suite** that includes candidate resume matching, interview scheduling bots, and analytics dashboards—capabilities showcased on the UBOS platform overview.

By addressing these gaps, the hybrid retrieval paradigm can evolve from a research prototype into a production‑grade engine that powers next‑generation talent acquisition platforms.

Conclusion

The paper demonstrates that a **metadata‑centric, hybrid lexical‑semantic retrieval pipeline** can deliver both high relevance and transparent explanations for job recommendation tasks. Its modular architecture, low reliance on heavy textual data, and strong empirical performance make it a practical foundation for AI‑enhanced HR systems.

Enterprises seeking to modernize their recruitment workflows should consider adopting such a hybrid approach, especially when paired with orchestration tools and explainability layers that align with regulatory and trust requirements.

Call to Action

Ready to explore how AI can streamline your hiring process? Visit the UBOS homepage for more insights, or dive into our templates for quick start on building intelligent recruitment agents. For a deeper technical dive, read the full arXiv paper.

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.