Integrating AI and Large Language Models for Automated Data Quality Enhancement in Data Integration Systems

You’re listening to “Integrating AI and Large Language Models for Automated Data Quality Enhancement in Data Integration Systems,” by N.K. Ponon and M.A.R. Basani. Published in 2026.

Abstract.

This article introduces an AI and LLM-based framework to automate data quality improvement in complex data systems. Traditional methods struggle with semantic inconsistencies and evolving schemas, degrading quality as data scales. The framework incorporates Real-Time Semantic Annotation (RTSA), adaptive ontology reinforcement, contextual similarity for duplicate detection, and continuous auto-healing. Explainability is ensured via SHAP-based alignment for transparency. Evaluated on the GOBY Benchmark dataset, it achieved 89.4% semantic annotation accuracy, outperforming the strongest baseline by 3%. The duplicate reduction rate was 64.5%, and the quality score averaged 83.2%, validating the auto-healing loop’s effectiveness. It adapts to evolving data without retraining, confirmed by robust performance under semantic drift.

The explainability analysis showed a low SHAP divergence of 0.11, aligning model predictions with feature importance. An ablation study confirmed that the semantic feedback loop and ontology reinforcement significantly contribute to stability. This research offers a scalable solution for enhancing data quality by integrating AI-driven adaptability and LLM-based semantic understanding with built-in explainability for enterprise-grade data pipelines.

Introduction.

The rising reliance on data-driven systems across indus-tries has intensified the demand for robust data integration pipelines capable of handling diverse, large-scale datasets. As organizations accumulate data from web scrapers, IoT sensors, APIs, and distributed databases, the influx of het-erogeneous, inconsistent, and often incomplete data becomes a persistent challenge. Issues such as semantic inconsis-tencies, duplication, schema mismatches, and noisy attributes routinely compromise data integrity, leading to inaccurate analyses, unreliable models, and flawed decision-making. Traditional methods often require intensive human supervi-sion to manually cleanse and curate datasets before integra-tion, increasing both operational cost and time-to-insight.

This work applies and evaluates an integrated AI-driven framework for automated Data Quality Enhancement (DQE) on large-scale, real-world data integration systems.

Despite advancements in machine learning and data in-tegration tools, existing systems typically operate in batch mode, relying on static rules or pre-trained models that are incapable of adapting to dynamic changes in data streams. These methods often fall short in addressing the real-world complexities of continuous data ingestion, where semantic drift, evolving schemas, and latent duplications emerge over time. The lack of adaptive mechanisms to detect, respond, and correct quality issues in real time poses significant risks in applications where data accuracy directly impacts business-outcomes.

This research is motivated by the need to advance from static data cleaning to adaptive frameworks for real-time quality assessment and correction. LLMs show promise in understanding and annotating complex data. However, their use in data integration is limited, often seen as static predictors instead of dynamic learners providing transparent explanations.

This research tackles the lack of a unified framework that integrates Real-Time Semantic Annotation (RTSA), adaptive quality control, and explainable AI into a self-regulating sys-tem. Current literature often addresses separate tasks like duplicate detection or semantic labeling, but fails to in-tegrate them into an architecture capable of dynamic data quality management. Additionally, many approaches overlook explainability, making outputs unclear and inap-propriate for settings requiring auditability and trust. Unlike prior studies that address these components in isola-tion, this work unifies RTS, adaptive ontology reinforcement, continuous quality monitoring, and explainability into a single self-regulating data integration framework evaluated under dynamic conditions.

Several recent works have attempted to utilize AI and Deep Learning (DL) models for data integration tasks, including using pre-trained language models for entity matching or schema alignment. While these approaches demonstrate potential, they often overlook the importance of continuous quality evaluation, fail to handle emerging data inconsis-tencies, and provide little to no insight into how decisions are made. Other works emphasize rule-based systems or heuristic-driven cleaning pipelines, which, while inter-pretable, lack scalability and adaptability to complex, evolv-ing datasets. Several recent works have attempted to utilize AI and DLmodels for data integration tasks, includ-ing using pre-trained language models for entity matching or schema alignment.

While these approaches demonstrate potential, they often overlook the importance of continuous quality evaluation, fail to handle emerging data inconsis-tencies, and provide little to no insight into how decisions are made. Other works emphasize rule-based systems or heuristic-driven cleaning pipelines, which, while inter-pretable, lack scalability and adaptability to complex, evolv-ing datasets.

This research proposes an AI-driven framework with LLM integration to improve data quality in enterprise systems. It features a feedback loop for real-time consistency checks, duplicate resolution, and semantic accuracy monitoring. An auto-healing mechanism enables model retraining and data correction when needed. SHAP-based alignment ensures model interpretability by synchronizing attention and feature importance.

The primary aim of this research is to design and implement an automated DQE framework that combines RTS, dynamic correction mechanisms, and explainable AI principles to im-prove the reliability of integrated datasets in large-scale data environments. The novel contributions of this research, which distinguish it from prior works, are summarized as follows: r Introduction of a real-time, continuous data quality monitoring system tightly integrated with LLM-driven semantic annotation for large-scale, heterogeneous datasets.

r Development of an adaptive auto-healing loop capable of performing local model retraining and dynamic cor-rection cycles based on live quality metrics and conflict detection.

r Embedding of SHAP-driven explainability alignment that enforces transparency by synchronizing model at-tention scores with external feature importance calcula-tions, addressing the growing demand for interpretable AI systems in enterprise applications.

II. LITERATURE REVIEW.

As diverse data sources grow, automated data quality in in-tegration systems becomes crucial. Traditional methods like rule-based systems, static schema matching, and manual interventions can’t keep up with data complexity. AI and LLM advancements offer potential for automating semantic annotation and quality evaluation, but real-time adaptability, explainability, and robustness for complex, evolving datasets remain challenging.

Kayali et al. developed the GOBY Benchmark, a large-scale dataset for evaluating semantic typing and data integration models. GOBY consists of 1 187 relational ta-bles with 4.04 million rows and 23 203 columns from 1 200 web scrapers of event promotion sites. The dataset’s se-mantic diversity includes nested attributes, long-tail class distributions, schema drift, and latent duplications. Models like Sherlock, Sato, and EmbNum were tested on GOBY. Sherlock had 74.1% accuracy, Sato improved to 77.8%, and EmbNum saw 82.4% but struggled with overlapping classes and unseen hierarchies. The models lacked dynamic adapta-tion to semantic drifts and explainability, needing retraining on new data. This highlights a gap in developing real-time, explainable, adaptive systems for semantic data quality.

Jansen et al. presented the MERGEN framework, which uses LLM to automate data analysis workflows from natural language instructions. Their system handled biolog-ical omics datasets and generated R and Python scripts for analysis. Although it sped up processes, the pipelines required human intervention due to a 23% error rate in syntax and semantics. Rasheed et al. examined LLMs in qualita-tive data analysis, creating a framework where GPT-3 agents handled tasks like thematic coding. Using large datasets, the system sped up analysis by 60% over manual methods but achieved only 70% accuracy compared to human-coded themes, often due to subtle context errors. This highlights the necessity for explainability and human oversight in LLM data annotation.

Arasteh et al. introduced ChatGPT ADA, an LLM-driven analyst, analyzing 50 clinical studies to predict cancer, achieving an 87.2% F1 score. Despite its accuracy, ex-plainability issues limit its healthcare use. Sun et al. found LLM data agents reduce development time by 40% but struggle with new schemas and semantic drifts. Lou et al. created an LLM literature review pipeline with RAG techniques, reaching 84% precision but misinterpreting 16% of terms, noting a need for better semantic alignment. Shankar et al. developed SPADe to improve data quality in LLM pipelines with automatic assertion synthesis, reducing runtime errors by 30% but focusing on syntactic rather than deep semantic issues.

AI-driven DQE is gaining traction. Informatica’s CLAIRE GPT embeds LLMs into metadata management for data discovery and profiling. Despite claimed productivity gains, it lacks public benchmark results, real-time adaptability, and explainability. Microsoft’s Azure Data Factory and IBM Watson use AI-driven quality checks, but mainly operate in batch mode, limiting real-time stream applicability. Vungarala et al. used chip design datasets with fine-tuned Llama2 models for automating data augmentation and Verilog code generation. Their system exceeded GPT-3.5 in structural correctness but struggled with rare logic paths, highlighting the challenge LLMs face in ensuring semantic quality in complex domains.

Studies have tried to add explainability to LLM-powered systems. Khraisha et al. assessed GPT-4’s 80% accuracy in medical data extraction but noted it struggled with domain-specific semantics. The lack of SHAP explainability reduced trust in results, highlighting a common issue of achieving accuracy with low transparency.

Across these studies, a consistent pattern emerges: while AI and LLMs demonstrate powerful capabilities in automating data quality tasks, including semantic annotation, duplicate detection, and pipeline generation, they consistently strug-gle with semantic drift handling, real-time adaptability, and explainability. Most systems require human intervention, re-training, or operate in batch mode, failing to meet the demands of dynamic data integration environments. Recent work has increasingly focused on automated data quality enhancement in data integration systems, emphasizing adaptive monitoring, semantic consistency, and scalability in enterprise environ-ments. Parallel research in explainable AI highlights the importance of transparent and auditable decision-making in data-driven pipelines, particularly where automated inte-gration impacts downstream analytics and governance.

These studies reinforce the need for unified frameworks that jointly address data quality, adaptability, and explainability.

III. METHODOLOGY.

This section introduces a real-time, automated methodology for enhancing data quality in integration systems using the GOBY Benchmark. It comprises distinct phases: Data Model Formulation, RTS with Adaptive Feedback, Dynamic Du-plicate Resolution, Ontology-Driven Quality Reinforcement, and Continuous Data Quality Metrics Evaluation. Each phase refines data quality through LLM capabilities, hierarchical modeling, and optimization-based decisions. Fig. 1 presents a high-level view of the proposed framework, showing how data flows through semantic annotation, duplicate handling, ontology adaptation, and continuous quality monitoring with an auto-healing feedback loop.

A. DATA MODEL FORMULATION FOR ENTERPRISE DATA STREAMS

Let G = {Ti }N i=1 represent the collection of N tables in the GOBY dataset, where each Ti is defined by Ti = (Ci, Ri). Each column c j ∈ Ci contains values v (k) corresponding to j row rk ∈ Ri. A streaming abstraction is introduced where in-coming data fragments Tt arrive over time t.

Xt = E (v (k) j) + λ1T (rk) + λ2 S (c j) where E (·) is the LLM semantic encoder, T (·) encodes timestamp-driven recency, and S (·) captures schema-driven priors. λ1 and λ2 are tunable weights learned during deployment.

B. REAL-TIME SEMANTIC ANNOTATION WITH ADAPTIVE FEEDBACK LOOP

The RTS process is designed to dynamically predict and refine the semantic label l j for each incoming column c j in the streaming window Wt. The predictive mechanism operates under continuous feedback control to maintain high semantic accuracy while adapting to changes in data distribution. The LLM, denoted as L, performs token-level probability estima-tion conditioned on the rich feature embedding tensor Xt and an evolving enterprise knowledge graph K extracted from the GOBY benchmark’s ontology O.

The model generates initial logits z j for every possible semantic label in the label space L, computed as: z j = L (Xt, K) = WL · h j + bL where h j is the hidden state of the LLM encoder for column c j, WL is the task-specific projection weight matrix, and bL is the bias term. This raw output captures the LLM’s confidence over all possible labels.

Key Mechanism: Adaptive Semantic Feedback Control. This component explains how semantic labels are continu-ously predicted and adjusted in response to streaming data quality feedback.

To ensure the model utilizes prior knowledge while adapt-ing to streaming data conditions, an adaptive feedback coef-ficient αt ∈ is applied. The semantic prediction proba-bility P(l j |Xt, K) is formulated as: This component explains how semantic labels are continuously predicted and adjusted in response to streaming data quality feedback.

P(l j |Xt, K) = σ αt · z j   +(1 − αt) · log P prev (l j) + ε

Here, σ (·) is the softmax function that makes sure valid probability outputs. The log(·) term is used to convert the previous prediction probability P prev (l j) into logit space, en-abling smooth combination with the current logits. A small constant ε is added for numerical stability.

The dynamic weight αt is calculated based on the real-time quality performance of the model, allowing immediate rein-forcement or penalization. Specifically, αt is modulated by the differential between the observed quality Qt and a pre-defined target quality Qtarget using a smooth sigmoid function:

β αt = 1 + e−γ (Qt −Qtarget) where β controls the maximum impact of feedback, and γ reg-ulates sensitivity. When Qt surpasses Qtarget, αt approaches β, granting higher trust to the model’s new predictions. Con-versely, when Qt underperforms, αt shrinks, reinforcing the historical confidence P prev (l j) to prevent semantic drift.

The instantaneous quality score Qt is recalculated per batch or sliding window by evaluating consistency, accuracy, and prediction confidence:

Further, the semantic prediction pipeline incorporates a noise rejection mechanism where low-confidence predictions below a threshold θcon f trigger automatic re-querying of the LLM:

⎪⎨ arg maxl P(l|Xt, K), if ˆl j = maxl P(l|Xt, K) ≥ θcon f ⎪⎩ re-query LLM, otherwise

Moreover, αt serves as a trigger for local online learning in edge deployment scenarios. If αt dips below a threshold αmin, an immediate fine-tuning step is invoked using the last K samples, creating a local weight adjustment WL minimizing the negative log-likelihood:

K  WL = −η∇WL log P(lk |Xk, K) k=1 where η is the learning rate controlling the extent of local adaptation.

This continuous adaptive feedback loop enables the se-mantic annotator to remain robust in the face of dynamic data shifts, schema changes, or concept drift common in enterprise datasets like GOBY. By fusing prior knowledge, current inference, and live feedback, the system make sure self-correcting behavior essential for real-time, production-grade deployments.

C. DYNAMIC DUPLICATE RESOLUTION VIA CONTEXTUAL SIMILARITY NETWORKS

The GOBY benchmark involves a high volume of similar but syntactically different data from varied sources, API in-consistencies, and wrapper flaws. Resolving these is crucial for data quality and preventing analytical errors. We created a contextual graph-based similarity network Gs = (V, E) for real-time duplicate detection and resolution in the integration pipeline.

Key Mechanism: Hybrid Similarity Score for Duplicate De-tection.

The graph Gs is incrementally constructed in each stream-ing window Wt, where every vertex v p ∈ V represents a data instance (typically a row r p from any table Ti ∈ G). Edges e pq ∈ E connect node pairs (v p, vq) with a dynamically com-puted weight ω pq, representing the hybrid similarity score between instances r p and rq. This similarity score ω pq is formulated as a convex combination of content-based seman-tic similarity and structural schema similarity, enabling both text-value analysis and columnar alignment:

The goal of this module is to identify and resolve se-mantically equivalent records by combining content-level embeddings with structural schema information.

Key Mechanism: Hybrid Similarity Score for Duplicate De-tection.

X(p) · X(q) t t ω pq = + η · Simschema (c p, cq) ∥X(p) ∥∥X(q) ∥ t t

Here, X(p) and X(q) are the high-dimensional contextual t t embeddings of instances r p and rq computed from semantic tokenization, position encoding, and temporal tagging. The dot product numerator captures the alignment of these em-beddings, normalized by their respective norms to compute cosine similarity, scale invariance.

The scalar η modulates the influence of schema-level alignment, where Simschema (c p, cq) evaluates the similarity between the corresponding column structures of r p and rq. This term is vital for differentiating false-positive matches arising from coincidental content similarity but structurally different contexts, such as “Apple” (fruit) and “Apple” (com-pany).

The graph Gs is fully constructed by calculating ω pq for every relevant pair (r p, rq) within the temporal window Wt. To manage computational complexity, an approximate nearest neighbor (ANN) search limits the number of pairwise calcu-lations to the top-K most similar pairs for each node.

Once the similarity graph is built, duplicate detection is formulated as a community detection problem, identifying densely connected subgraphs Dk representing clusters of po-tential duplicates. A node r p is considered part of duplicate cluster Dk if the cumulative similarity with its neighborhood surpasses a pre-defined threshold τdup:

⎨ ⎬  Dk = ⎩r p ∈ Gs | ω pq > τdup ⎭ rq ∈Gs

The threshold τdup is not static but adaptively updated based on the current window’s data density and drift metrics, allow-ing the system to become stricter or more lenient in noisy environments.

For each detected duplicate cluster Dk, the challenge lies in resolving conflicts and merging records without data loss or distortion. We formulate a conflict resolution loss function Lmerge that simultaneously minimizes inter-instance variance and maximizes completeness of the aggregated instance:

{v (p) Lmerge = j } p∈Dk Var c j   −κ · Completeness Dk, c j

Here, Var({v (p) j }) computes the statistical variance across all values v (p) observed in column c j across the cluster Dk, j reflecting the conflict intensity. High variance indicates incon-sistent data entries requiring resolution.

The term Completeness(Dk, c j) measures the proportion of non-null or information-rich entries for column c j in clus-ter Dk, guiding the algorithm towards preserving fields with dense information.

The weight κ acts as a regularization hyperparameter con-trolling the trade-off between favoring consistency (low vari-ance) and information retention (high completeness). Higher κ encourages the model to merge aggressively, while lower κ prioritizes conservative merging to avoid data loss.

The optimal merged row r ∗ m for cluster Dk is generated by solving the following optimization problem: r ∗ m = arg min rm Lmerge

This optimization is implemented through iterative gradi-ent descent or rule-based majority voting, depending on data complexity. During the merge, conflicting values are resolved using semantic embeddings:

⎝  v (p) v (m) ⎠ = arg max E(v), E cos j j v∈{v(p) j } p∈Dk where E(v) is the LLM-driven embedding of value v validates semantic proximity influences the final merged value.

Post-merge, the entire Gs is pruned by removing merged nodes and updating neighboring similarities, whih makes sure graph consistency for the next iteration.

Higher values of DuplicateReductiont indicate more effective elimination of redundant records relative to the total number of instances.

D. ONTOLOGY-DRIVEN REINFORCEMENT AND REAL-TIME EXPANSION

The Ontology-Driven Reinforcement module ensures seman-tic consistency while adapting to changing enterprise data streams. The dataset requires real-time restructuring of the semantic ontology Ot.

At each streaming interval t, the system observes the dis-tribution of predicted semantic labels Pt (l) across all active columns and data points. This distribution captures the proba-bility mass assigned by the LLM to each label l in the current ontology Ot. To monitor semantic stability, the system com-putes the semantic drift δt between consecutive time windows, quantifying how much the model’s understanding of labels has shifted due to new data influx.

The drift δt is rigorously computed using the Jensen-Shannon Divergence (JSD), a symmetric and bounded mea-sure ideal for probability distributions comparison:

δt = JSD (Pt (l) ∥ Pt −1 (l)) = 1 2 DKL (Pt (l) ∥ M) + 1 2 DKL (Pt −1 (l) ∥ M) where M = 1 2 (Pt (l) + Pt −1 (l)) serves as the mixed distri-bution and DKL (·) is the Kullback-Leibler divergence. This calculation accurately captures both subtle and abrupt changes in the label space over time.

Once computed, δt is continuously compared to a pre-defined drift threshold δthr, calibrated based on historical variance and business tolerance to semantic changes. If δt ≤ δthr, the system deems the semantic model stable, and no up-dates are triggered. However, if δt > δthr, it signals potential emergence of new semantic classes, label conflicts, or shifts requiring active ontology restructuring.

Upon drift detection, the system initiates a semantic sim-ilarity scan across all label pairs (li, l j) in Ot to identify candidates for merging or new node creation. The inter-class semantic similarity S (li, l j) is calculated using the overlap of their descendant sets in the hierarchy:

This ratio measures both hierarchical proximity and shared semantics. High similarity values indicate conceptual redun-dancy or excessive granularity, suggesting that li and l j represent semantically equivalent or overlapping classes.

The ontology restructuring mechanism executes two core operations:

1) Merging Classes: If S (li, l j) > τmerge, where τmerge is.

a carefully chosen semantic redundancy threshold, the system merges li and l j into a unified class lmerged to prevent hierar-chical bloating: lmerged = li ∪ l j,

Update: Descendants(lmerged) = Descendants(li) ∪ Descendants(l j)

2) Spawning New Classes: Conversely, if δt is caused by.

the emergence of novel semantic concepts not covered by Ot, new classes are spawned by analyzing outliers in Pt (l)—those receiving high probability mass without belonging to any cur-rent l ∈ Ot. Each new node lnew is initialized with a linkage to the closest ancestor l parent determined by: l parent = arg max SemanticMatch (li, lnew) li ∈Ot

This maintains hierarchical coherence and prevents semantic drift from destabilizing the model.

1) REAL-TIME PURITY PRESERVATION AND DEPTH.

MINIMIZATION

Each modification step is governed by two optimization objec-tives. Firstly, the system preserves semantic purity, confirming that merged classes or spawned nodes do not introduce label noise or ambiguous interpretations. Purity Pl of a class l is dynamically computed as:

 ˆlk = l rk ∈C (l) I Pl = |C (l)|

Algorithm 1: Ontology-Driven Reinforcement and Real-Time Expansion.

where C (l) is the collection of records currently assigned to l and ˆlk is the predicted label.

Secondly, the model optimizes the hierarchy depth D(Ot), preventing excessive nesting, which could hinder real-time LLM performance and increase the complexity of semantic reasoning:

D(Ot) = max Depth(l) l∈Ot

If D(Ot) surpasses a system-defined threshold Dmax, a balanc-ing operation is triggered to flatten branches while preserving semantics.

2) ITERATIVE REAL-TIME ONTOLOGY UPDATE.

The entire ontology adjustment runs iteratively within each time window until the following convergence criterion is met:

δt < δthr D(Ot) ≤ Dmax AND l Pl ≥ Pmin AND min

This make sure the refined ontology is semantically pure, hier-archically balanced, and fully adapted to the latest enterprise data semantics.

The updated Ot is serialized and re-fed into the semantic annotation pipeline, reinforcing the learning loop and allow-ing the system to handle concept drift, vocabulary expansion, and domain adaptation in a fully autonomous, real-time manner.

E. CONTINUOUS DATA QUALITY METRICS AND AUTO-HEALING

The proposed system operates under a continuous quality monitoring regime designed to evaluate data integrity and guide real-time model adjustments. A composite data qual-ity score Qt is computed during each processing window to assess the system’s current semantic accuracy, duplication control, and structural consistency. This score not only reflects the immediate health of the data pipeline but also acts as a feedback control parameter for the self-healing mechanism.

Mathematically, the data quality score Qt is defined as a linear combination of three dynamically computed metrics, each weighted by calibrated coefficients λc, λd, and λs. These coefficients are determined based on domain priorities and can be updated during deployment through reinforcement learning strategies. The formulation is expressed as follows:

Qt = λc · Consistencyt + λd · DuplicateReductiont + λs · SemanticAccuracyt

The consistency component Consistencyt evaluates the uniformity of semantic labels and the absence of logical con-tradictions within the dataset. It is computed by measuring the cumulative number of detected conflicts across all ac-tive columns c j ∈ C during window t. Conflicts arise from scenarios such as mixed data types within a semantic class, contradictory values within duplicates, or time-inconsistent records. The normalized consistency is calculated by:

The duplicate reduction component DuplicateReductiont is computed as defined in, where |V | denotes the number of active instances (nodes) in the contextual similarity graph Gs during window t.

The semantic accuracy component SemanticAccuracyt measures the correctness of semantic label predictions by the model in the given window. It accounts for hierarchical correctness, allowing partial credit for predictions that align with parent or child nodes within the evolving ontology. The metric is formalized as:

Once Qt is computed, it is continuously evaluated against a system-defined target quality threshold Qtarget. This threshold reflects the minimum acceptable quality based on historical performance, business requirements, and downstream appli-cation needs. If the current quality score Qt remains greater than or equal to Qtarget, the system continues operation with-out interruption, signaling healthy semantic integration and structural integrity.

However, if Qt drops below Qtarget, it activates the auto-healing mechanism designed to restore model and data stability. The auto-healing process begins with resampling problematic regions of the dataset, specifically targeting data instances contributing to the high conflict rate, unresolved duplicates, or semantic misclassifications. Formally, the re-sampled batch Bresample is constructed from instances ri ∈ Gs that satisfy either the conflict threshold τcon f or fall below the model’s confidence margin:

Conflict(ri) > τcon f Bresample = ri ∈ Gs ∨ LowConf(ri)}

The local LLM model is then retrained on this targeted batch by performing a focused gradient descent step. This retraining corrects weak predictive pathways and strengthens the model’s decision-making capacity where errors were most likely. The weight update is expressed as:

WL = −η · ∇WL log P(li |Xi, K) ri ∈Bresample where η is the learning rate governing update magnitude. In parallel, the semantic feedback coefficient αt is updated to enhance reliance on corrected predictions in the subsequent window:

αnew = αt + ζ · (Qtarget − Qt) t with ζ controlling the feedback learning rate. This dynamic adjustment prevents repeated over-reliance on stale or erro-neous semantic predictions. The retrained model then re-evaluates the resampled batch, and the quality score Qt is recomputed. This iterative auto-healing loop continues until either the updated Qt surpasses Qtarget or a maximum number of iterations is reached, pre-venting infinite loops under rare pathological data conditions:

Qt ≥ Qtarget OR MaxIters reached

F. REAL-TIME EXPLAINABILITY VIA SHAP-ATTENTION ALIGNMENT

Ensuring interpretability is crucial for using LLMs in seman-tic data integration within enterprises. The framework com-bines SHapley Additive Explanations (SHAP) with attention mechanisms to assess each feature’s semantic contribution to predictions in real-time.

For every semantic prediction task involving a column c j, the model computes the Shapley value φ j, which quantifies the marginal contribution of feature c j to the overall predic-tion made by the semantic scoring function f (·). The Shapley value is formally defined based on the cooperative game the-ory principle and computed as the weighted average of the difference in model output when c j is present versus absent across all possible feature subsets S ⊆ C \ {c j }:

φ j = ES⊆C\{c j } f (S ∪{c j }) − f (S)

Here, f (S) denotes the semantic prediction probability score when the model is conditioned only on the subset S, and f (S ∪{c j }) represents the score when c j is included. This computation measures the exact additive importance of c j towards the label decision and inherently accounts for feature interactions and context.

Simultaneously, during the semantic scoring phase, the model computes the attention distribution Ai j generated by the LLM’s attention heads. Each attention score Ai j represents the normalized weight assigned by the model to the jth input token or feature while processing the ith output token or se-mantic prediction. Attention scores provide a direct view of the model’s internal focus during decision-making.

To align model explainability with model reasoning, the framework computes an explanation divergence metric explain, which quantifies the discrepancy between the ex-ternally computed SHAP values and the internal attention feature contributions, possibly due to overfitting on irrelevant tokens, noisy inputs, or semantic shifts.

When explain > τexp occurs, the system starts an inter-pretability refinement cycle with two corrective actions. The first action re-weights the LLM’s attention layers using a reg-ularization loss Lalign to minimize explain in future training.

|C|   2 Lalign = φ j − Ai j j=1

Second, the semantic scoring function f (·) is temporarily con-strained to down-weight low-contribution features identified by SHAP while emphasizing high-contribution tokens, forc-ing the model to re-learn its focus.

This dual-step refinement make sure that the model’s decision pathway aligns with logically sound, explainable rea-soning derived from SHAP analysis. Over time, this prevents degradation of interpretability, reduces the risk of semantic hallucinations, and maintains accountability within enterprise data processing pipelines.

IV. EXPERIMENTAL SETTING.

The experimental design for this study was carefully struc-tured to evaluate the proposed AI-driven DQE framework on the GOBY Benchmark dataset. The dataset itself posed significant real-world challenges due to its complex structure, consisting of 1,187 source tables comprising approximately 4.04 million rows and 23 203 columns. Each table represented diverse enterprise-level event promotion data, collected via over 1 200 production-grade web scrapers. These sources in-troduced common inconsistencies such as varying schema structures, redundant attributes, diverse numeric formats, missing values, and textual noise, which made GOBY a highly suitable environment for testing automated semantic annota-tion, ontology adaptation, and data quality restoration.

All experiments used the raw GOBY dataset without synthetic cleaning to preserve data inconsistencies. The to-kenization was executed using Byte-Pair Encoding (BPE) optimized with a vocabulary size of 32,000. This parameter was initially explored with 16 000, 32 000, and 50 000 token vocabularies, and the 32 000 size offered the best balance between sequence length and semantic resolution. Sequence lengths ranging from 256 to 1 024 tokens were tested, but a limit of 512 tokens was finalized as it provided enough context per data instance while controlling GPU memory consump-tion. Data samples with larger lengths suffered from gradient instability and latency, while lower sequence lengths missed necessary semantic context.

The Large Language Model configuration was tuned exten-sively across various parameter combinations. Model scales of 7B, 13B, and 30B were initially evaluated. The 7B model underperformed in capturing complex semantic dependen-cies across multi-column relations, particularly in identifying nested organizational hierarchies. The 30B model yielded marginal gains in semantic F1 but significantly increased training times and required memory beyond the A100 80 GB GPU configuration. The 13B parameter configuration demon-strated optimal performance in terms of semantic accuracy and resource efficiency and was used for final experiments. The learning rate η for LLM fine-tuning was thoroughly explored from 1e−5 to 1e−4. The setting of 3e−5 was fi-nalized as it ensured stable convergence while minimizing the risk of catastrophic forgetting during local retraining phases.

The semantic feedback coefficient αt was initially set to 0.5, with exploration of 0.3, 0.5, and 0.7. Lower values decreased adaptability, while higher values led to reliance on predictions. Empirical results showed αt = 0.5 effectively balanced past knowledge with current LLM inferences for stability during concept drift. The adjustment rate ζ, tested between 0.005 and 0.05, performed best at 0.01 for smooth responsiveness. The target quality threshold Qtarget, adjusted from 0.8 to 0.9, was optimally set at 0.85 to trigger auto-healing as needed, avoiding unnecessary retraining.

Drift detection sensitivity was optimized with Jensen-Shannon divergence δt thresholds from 0.1 to 0.2; 0.15 was ideal for capturing semantic shifts without false flags. Lower values caused unnecessary restructuring, while higher values missed changes. The inter-class similarity threshold τmerge was set to 0.7 for coherent merges, avoiding over-merging and reducing ontology size. The purity threshold Pmin was optimized at 0.9, maintaining label integrity and flexibility for data noise. Ontology hierarchy depth Dmax was best at 6, balancing structure and efficiency.

Duplicate detection hyperparameters were equally scruti-nized. The similarity threshold τdup in the contextual simi-larity graph was examined from 0.7 to 0.9. A threshold of 0.8 provided the best compromise between high recall in duplicate detection and false positive control. Values below 0.8 resulted in excessive merging of semantically unrelated records. The schema similarity weight η was explored in the range of 0.2 to 0.5 and setting it at 0.3 provided the best bal-ance, enabling the model to consider schema alignment while retaining a dominant influence of content similarity. Conflict resolution regularization parameter κ was tested within the range 0.4 to 0.8. A value of 0.6 worked well, prioritizing com-pleteness while effectively minimizing intra-cluster conflicts during merges.

The SHAP module was extensively evaluated. Background sample sizes of 5% to 20% of the GOBY dataset were tested for stable SHAP value estimation, with 10% providing model-agnostic, stable values without high compute cost. The alignment divergence threshold τexp was empirically set be-tween 0.1 to 0.3, with 0.2 τexp performing best by correcting only true misalignments between SHAP importance and LLM attention scores. The alignment loss weight λalign examined from 0.01 to 0.1, settled at 0.05 to balance semantic learning with explainability preservation.

Auto-healing was analyzed with iteration limits of 3, 5, and 7; 5 iterations proved optimal, balancing recovery and compute efficiency. The resampling threshold τcon f was tested from 0.2 to 0.4, with 0.25 maximizing efficiency by targeting high-conflict records. Batch gradient accumulation evaluated from 4 to 16 steps, with 8 yielding best GPU utilization and stability.

Experiments were conducted on a four-node NVIDIA A100 80 GB cluster using PyTorch 2.1 and Hugging Face Trans-formers 4.38, with mixed precision and distributed data par-allelism. Metrics such as semantic precision, recall, F1-score, data consistency, duplicate reduction, semantic accuracy, and explainability divergence were evaluated live per batch. A 13B model with αt = 0.5, η = 3e−5, δthr = 0.15, τdup = 0.8, and τmerge = 0.7 achieved the best accuracy, interpretability, runtime, and semantic integrity on the GOBY benchmark, demonstrating the framework’s robustness and adaptability.

The proposed framework operates under bounded latency suitable for continuous data ingestion, with end-to-end pro-cessing for each streaming window completing within a few seconds per batch on the A100 GPU cluster. Auto-healing is triggered only when the quality score Qt falls below the target threshold and is applied to a small, resampled subset of data associated with conflicts, low-confidence predictions, or unresolved duplicates. Retraining is localized, limited to a few iterations on recent samples, and introduces only mod-est overhead compared to standard inference. This controlled fine-tuning strategy avoids costly full-model retraining while enabling scalable, real-time correction in large enterprise data pipelines.

In this manuscript, real-time denotes near-real-time processing at the level of streaming data windows, where end-to-end execution completes within a few seconds per batch rather than under hard millisecond-level constraints.

V. RESULTS AND ANALYSIS.

This section presents results following the same sequence as the framework components, covering semantic annota-tion, duplicate resolution, drift handling, and explainability. The proposed AI and LLM-driven framework was exten-sively evaluated on the GOBY Benchmark dataset, testing its ability to perform RTS, duplicate detection, drift han-dling, and explainability alignment. Performance was bench-marked against Sherlock, Sato, and EmbNum models. Evalu-ation focused on semantic classification accuracy, duplicate reduction, semantic consistency, quality maintenance, and explainability.

Semantic annotation accuracy was significantly improved by integrating the adaptive ontology reinforcement module and the feedback-driven self-correction mechanism, both op-erating in real-time during batch processing. The proposed system dynamically adjusts semantic predictions for each column c j based on the evolving probability distribution P (l j |Xt, K). This adaptive behavior allowed the system to continuously recalibrate its semantic labels by comparing the live predictions against historical knowledge stored in the ontology Ot.

The annotation model was stress-tested on the GOBY dataset, where each table contained heterogeneous schema structures, missing context, and nested attributes—common in scraped web data. The system processed over 23,000 columns, predicting their semantic types while receiving continuous feedback on prediction confidence. Columns with decreasing confidence below threshold θc triggered auto-reinforcement cycles, forcing a recomputation of P(l j) conditioned on both the knowledge graph K and the updated drift coefficient δt. This mechanism directly mitigated misclassification caused by schema variations or latent contextual drifts observed dur-ing real-world table integrations.

Accuracy peaked at 89.4%, surpassing EmbNum’s 86.3%, the best-performing baseline. Notably, EmbNum’s embedding-based similarity struggled when encountering classes not well-represented in the training data’s latent space. The proposed system, by contrast, utilized the real-time αt update to balance new incoming evidence with prior beliefs. The ability to regulate αt based on the Jensen-Shannon divergence between batches allowed the model to suppress noisy updates and maintain stability, preventing oscillations commonly seen in non-adaptive systems.

Further analysis revealed that the system’s accuracy gain was primarily due to better handling of rare semantic classes, such as ‘Performer-Name’ and ’Event-Type’, which repre-sented less than 5% of the GOBY dataset distribution. Base-line models misclassified these tail classes due to their low frequency, while the proposed model dynamically expanded the ontology Ot to create new nodes for emerging semantics, preserving accuracy without retraining. Fig. 2 illustrates this steady improvement across the models, demonstrating that the proposed framework does not overfit but maintains generaliz-ability within a reasonable margin.

The plotted results confirm that accuracy improvements re-mained bounded within a realistic 3% margin over EmbNum, reflecting robustness without overestimation. Error analysis also showed that the remaining 10.6% misclassifications were predominantly due to completely ambiguous columns lacking any informative tokens—an inherent limitation in the dataset rather than the model.

The evaluation on GOBY’s diverse dataset revealed that the framework achieved a duplicate reduction rate of 64.5%, carefully surpassing EmbNum’s 62.3%. EmbNum, primar-ily relying on embeddings and Euclidean distance in la-tent space, struggled with context-dependent duplications where semantic equivalence could not be fully captured by vector similarity alone. In contrast, the proposed model’s adjustable schema similarity coefficient η allowed explicit tuning of schema-level influence, confirming that structurally re-ordered columns or nested fields were fairly evaluated for potential duplication.

The detected duplicates were clustered into communities Dk based on edge weights exceeding the threshold τ dup, fol-lowed by merging operations that minimized the conflict loss Lmerge as defined earlier. During experimentation, increasing η beyond 0.5 caused over-merging, while setting it too low degraded detection precision. The final η = 0.5 achieved a balanced detection rate, verified through conflict resolution variance analysis.

This limited yet significant performance margin of 2.2% over EmbNum is intentionally constrained, reflecting a realis-tic gain without artificial inflation. The marginal improvement directly results from the system’s ability to resolve complex scenarios—such as records differing in date formats, name or-derings, or location hierarchies—cases where baseline models either overgeneralized or under-merged.

Handling semantic drift emerged as a pivotal factor in validating the proposed framework’s adaptability to dy-namic and evolving data streams. To systematically evaluate drift resilience, controlled experiments introduced progressive schema changes within the GOBY dataset. These included attribute renaming, nesting, and injecting noisy attributes with partially overlapping semantics. This setup aimed to emulate real-world conditions where data schemas naturally evolve due to updates in data sources or integration of new, heteroge-neous datasets.

The model’s drift adaptability relied heavily on the ontology-driven reinforcement mechanism, which dynami-cally computed the semantic drift coefficient δt using Jensen-Shannon Divergence. Once δt exceeded the threshold δthr, the system triggered real-time restructuring of the ontology Ot, merging or spawning nodes based on inter-class similarity scores S (li, l j). This action preserved semantic purity while minimizing hierarchy depth, allowing the model to remain aligned with evolving data semantics.

The results indicated that at increasing drift levels catego-rized as no drift, low, medium and high, the model maintained a higher accuracy compared to the baselines. Without drift, the model matched its performance in semantic annotation at 89. 4%. As drift levels increased, accuracy gracefully de-graded but retained a distinct advantage. Notably, under high drift, the model sustained an accuracy of 80.2%, while Em-bNum dropped to 77.5%, as shown in Fig. 3. This behavior validates the framework’s ability to identify and adapt to se-mantic shifts without requiring full retraining—an inherent weakness in traditional static systems.

Further analysis revealed that the drift handling capability was mainly strengthened by the real-time feedback loop that regulates αt, which determined the weight assigned to new semantic predictions versus prior knowledge. As the drift in-creased, the model dynamically reduced the dependency on previous distributions that were no longer valid, preventing error propagation. This nuanced balancing act allowed the framework to remain responsive without overfitting to tran-sient noise.

Gleichzeitig, the continuous data quality score Qt was cal-culated to capture the combined impact of semantic accuracy, duplicate resolution, and consistency between batches. The framework’s composite scoring mechanism, validated the sys-tem’s robustness by maintaining an average Qt of 83.2%, slightly above EmbNum’s 80.1%. The quality score con-sistently remained above the auto-healing trigger threshold Qtarget, requiring minimal retraining cycles.

These observations directly illustrate the impact of the framework’s self-healing loop. Whenever drift-induced in-consistencies caused local dips in semantic alignment or duplicate handling, the model initiated targeted retraining of local LLM weights. This on-the-fly correction ensured that the system never experienced prolonged quality degradation—a limitation observed in baseline models, which lacked any feedback-driven correction capability.

The explainability analysis was carried out by measuring the divergence between SHAP importance values and at-tention weights, confirming the interpretability of semantic annotations. A low divergence score of 0.11 was achieved, reinforcing that model decisions were grounded in feature importance. Baseline models lacked any comparable explain-ability mechanism. Fig. 4. Explainability alignment between SHAP feature importance values and LLM attention weights. Features shown here are those receiving the highest atten-tion during semantic prediction. The concentration of high-magnitude SHAP values on the same features emphasized by attention indicates low explainability divergence (explain = 0.11), demonstrating strong alignment between model reason-ing and feature contribution.

A detailed parameter evaluation was also conducted to mea-sure the impact of adaptive feedback (αt), ontology expansion threshold (δthr), and similarity influence (η). Table 1 reports system performance under varying configurations. Results confirmed that performance peaks when αt is dynamically updated and η is balanced to avoid overfitting to schema similarity.

An ablation study was conducted to measure the individual impact of each module by progressively disabling compo-nents. The results, shown in Table 2, confirm that the semantic feedback loop contributed the most to accuracy, while the ontology expansion improved resilience under drift.

The cumulative system comparison integrating accuracy, duplicate reduction, semantic quality, and explainability is summarized in Table 3. The framework consistently outper-formed baselines within a constrained improvement margin, enhancing realism. All ablation results were averaged over multiple runs with different random seeds, and the observed improvements were consistent across runs.

A final combined graph, presented in Fig. 5, visualizes the system’s superiority across all dimensions. The proposed model balances accuracy, quality, and explainability without overfitting or unrealistic gains.

VI. CONCLUSION.

This study introduces an AI and LLM-driven framework to automate data quality improvements in complex data inte-gration, tackling challenges in semantic annotation, duplicate detection, semantic drift, and explainability. By incorporating adaptive ontology reinforcement, contextual similarity net-works, real-time feedback, and SHAP-based explainability, it operates effectively under real-world conditions. Testing on the GOBY Benchmark revealed its superior semantic an-notation accuracy at 89.4%, surpassing the best model by 3%. It achieved a duplicate resolution rate of 64.5% and maintained an average accuracy score of 83.2%, correct-ing issues dynamically with the auto-healing loop. Ablation studies highlighted the key contributions of each module, especially the feedback-driven loop and ontology reinforce-ment.

The system’s resilience during semantic drift scenarios showcases its suitability for evolving data environments. Explainability is enhanced through SHAP alignment, with a low divergence of 0.11, supporting user trust for appli-cations in healthcare, finance, and enterprises. Integrating AI in this real-time, modular framework offers gains over static models by balancing accuracy, adaptability, and trans-parency, without needing retraining. Recent studies on digital security by design emphasize embedding resilience, trans-parency, and governance mechanisms directly into data-driven systems, particularly as AI-enabled automation becomes more prevalent.

While such work primarily addresses cyber risk and security governance, the framework presented here complements these efforts by focusing on semantic integrity, data consistency, and explainability as upstream enablers of secure digital infrastructures. By ensuring that integrated data remains semantically accurate, drift-aware, and interpretable in real time, this work contributes to the broader objective of designing AI-centric data systems that are robust, auditable, and trustworthy by construction. Future work aims to scale for multi-modal streams, integrate reinforcement learning, extend explainability to natural language, and enrich the ontology with external knowledge bases for better recognition of rare attributes.