ENTRep: Multi-Domain Research Label

• ENTRep is a multi-domain research label that encompasses reputation-based expert ranking, ENT endoscopy benchmarks, and enterprise management systems.
• In enterprise analytics, ENTRep employs a Bayesian reputation model to rank tacit experts using employee interactions with improved precision and reduced error.
• In medical multimedia and ERP contexts, ENTRep integrates fine-grained classification and natural language interfaces to support decision-making under operational constraints.

Searching arXiv for papers relevant to “ENTRep” and closely related usages. Tool call: arXiv search for "ENTRep" ENTRep is a label used in several distinct research contexts. In enterprise analytics, it denotes a reputation-based expert-finding framework derived from EER, “Enterprise Expert Ranking using Employee Reputation,” for ranking tacit experts from positive and negative employee interactions (Mahmood et al., 2020). In medical multimedia, it names the ACM Multimedia 2025 Grand Challenge on ENT endoscopy analysis, which unifies fine-grained anatomical classification with image-to-image and text-to-image retrieval under bilingual clinical supervision (Nguyen et al., 6 Aug 2025). In enterprise systems, ENTRep also appears as “Intellectual Management of Enterprise,” a thin-client natural-language–to–decision-engine platform layered on top of ERP data (Ostapov, 2012). The term is therefore best understood as a context-dependent research label rather than a single technical framework.

1. Terminological scope and adjacent usages

The main documented uses of ENTRep in the supplied literature are summarized below.

Usage	Domain	Core function
EER / ENTRep	Enterprise analytics	Expert ranking using employee reputation
ENTRep Grand Challenge	ENT endoscopy analysis	Classification and retrieval benchmark
ENTRep	Enterprise management systems	Natural-language interface, optimization, adaptive control

A common source of ambiguity is the proximity of ENTRep to “ENTREP” in entrepreneurship-education literature. In that adjacent usage, the relevant construct is the Entrepreneurship Competences Framework, defined as “the capacity to act upon opportunities and ideas and transform them into value for others,” with three key areas and 15 sub-competences; a STEM study operationalized these through 22 items and derived a five-factor structure rather than discussing expert ranking, ENT endoscopy, or ERP-centered control (Cristina et al., 27 Aug 2025). This suggests that acronym-level similarity should not be taken as conceptual identity.

2. Reputation-based expert ranking in enterprises

In the enterprise-ranking usage, ENTRep refers to a reputation-based expert-finding framework for large, geographically distributed enterprises in which employees’ tacit expertise may not be documented or self-disclosed. The system continuously harvests interactions among employees, including direct ratings, social-network events such as follows, unfollows, likes, and dislikes, and textual messages, posts, or comments that can be sentiment-scored. Each raw event is timestamped and attached to a rater–ratee pair. Recorded events are then classified as positive $(\alpha)$ or negative $(\beta)$, with numerical signals binarized via a threshold and free text passed through a sentiment analyzer. For each candidate expert, the sequence of positive and negative interactions is fed into a Beta-distribution–based Bayesian reputation model, and the posterior expected value is treated as the expertise score. At any time $t$, employees are sorted by computed reputation and the top-$K$ are presented as experts in a given domain or in the enterprise as a whole (Mahmood et al., 2020).

This formulation directly addresses the expert-identification problem in online enterprises spread across continents. The underlying motivation is the intention to find tacit expertise and knowledge of an employee that is not documented or self-disclosed. The approach is explicitly positioned against existing reputation-based expertise-ranking methods using PageRank, normal distribution, and hidden Markov models, which are described as suffering from negative referral, collusion, reputation inflation, and dynamism (Mahmood et al., 2020).

3. Bayesian scoring, temporal adaptation, and empirical behavior

The core scoring model adapts the Beta reputation system. For employee $i$ at time $t$, let $p_i(t)$ be the number of positive events and $n_i(t)$ the number of negative events. With prior pseudocounts $\alpha_0=\beta_0=1$ by default, the posterior parameters are

$$\alpha_i(t)=\alpha_0+p_i(t), \qquad \beta_i(t)=\beta_0+n_i(t),$$

and the expected reputation is

$(\beta)$0

In the simplest case,

$(\beta)$1

ENTRep can operate in full-history mode with $(\beta)$2 or in sliding-window mode with $(\beta)$3, where only the last $(\beta)$4 interactions are used. By explicitly modeling positive versus negative events, the method is described as naturally resisting negative referrals, tempering collusion, and controlling reputation inflation through the denominator $(\beta)$5 (Mahmood et al., 2020).

Empirical evaluation used three real-world advisor datasets, DS1, DS2, and DS3, with 46–77 employees each, where edge weights 0–5 or 0–6 encode how much one employee trusts or seeks advice from another. Baselines were BL1, ExpertRank; BL2, WorkerRank; and an HMM-based model. Metrics were MAE and Precision@10. Reported MAE values were DS1: EER $(\beta)$6 versus BL1 $(\beta)$7 and BL2 $(\beta)$8; DS2: EER $(\beta)$9 versus BL1 $t$0 and BL2 $t$1; DS3: EER $t$2 versus BL1 $t$3 and BL2 $t$4. Precision@10 values were DS1: EER $t$5 versus BL1 $t$6 and BL2 $t$7; DS2: EER $t$8 versus BL1 $t$9 and BL2 $K$0; DS3: EER $K$1 versus BL1 $K$2 and BL2 $K$3. On average, the method achieves roughly a 7% absolute precision gain over the best baseline, while MAE is consistently reduced by a factor of $K$4–$K$5 (Mahmood et al., 2020).

The reported robustness claims are equally specific. Performance is stated to hold across both sparse and dense graphs, making the technique independent of rating density and pattern. In contrast, PageRank is reported to fail when every node has equal degree, and WorkerRank assumes normality that may not hold. Compared with the HMM approach, the method is described as far simpler to implement and interpret while still capturing temporal reputation shifts (Mahmood et al., 2020).

4. ENTRep as an ENT endoscopy benchmark

In medical multimedia, ENTRep is the ACM Multimedia 2025 Grand Challenge on ENT endoscopy analysis. It was introduced to address automated analysis of endoscopic imagery in ear, nose, and throat care, a setting characterized by variability in devices and operators, subtle and localized findings, and fine-grained distinctions such as laterality and vocal-fold state. The benchmark integrates fine-grained anatomical classification with image-to-image and text-to-image retrieval under bilingual Vietnamese and English clinical supervision (Nguyen et al., 6 Aug 2025).

The dataset is built on real-world ENT endoscopy images acquired at Thong Nhat Hospital in Ho Chi Minh City under routine clinical conditions. Images were annotated by otolaryngology experts for fine-grained anatomical region and presence or absence of abnormal findings, and were accompanied by bilingual narrative descriptions. The benchmark description specifies a public test split of 645 images and a private test split of 646 images, while the training split size is withheld. Seven region labels are used: Ear-Right, Ear-Left, Nose-Right, Nose-Left, Throat, Vocal-Fold Open, and Vocal-Fold Closed. Each image also carries a binary “Type” label, normal or abnormal. The provided JSON schema includes “Description” for the Vietnamese clinical narrative and “DescriptionEN” for the expert-verified English translation; English queries contain 5.42 words on average (Nguyen et al., 6 Aug 2025).

Three tasks are defined. Task 1 takes a single RGB endoscopic image and outputs one of the seven anatomical labels. Task 2 takes a query RGB image and returns a ranked list of gallery images that are visually or semantically similar. Task 3 takes an English clinical description and returns a ranked list of matching images. Final rankings are determined on the private test split using server-side scoring, while the public split supports interim leaderboards (Nguyen et al., 6 Aug 2025).

Evaluation follows standard formulations. For classification, Accuracy, Precision, Recall, and F1-score are computed per class and aggregated via weighted averaging:

$K$6

For retrieval, Recall@1 and Mean Reciprocal Rank are used:

$K$7

The benchmark paper characterizes ENTRep as the first ENT endoscopy benchmark to unite fine-grained classification with both intra-modal and cross-modal retrieval under bilingual supervision (Nguyen et al., 6 Aug 2025).

5. Vision-language methods and benchmark performance

A 2025 solution paper presents a unified vision-language framework for the ENTRep Grand Challenge using a CLIP ViT-B/16 image encoder, a frozen CLIP text encoder, Low-Rank Adaptation, multi-level CLS token aggregation, and spherical feature interpolation. The starting point is pre-trained CLIP ViT-B/16 with 86.7 M parameters. LoRA is inserted into every attention projection $K$8 in each transformer block, with the low-rank update $K$9, rank $i$0, $i$1, and scaling $i$2; $i$3 is Kaiming-initialized and $i$4 at start. Only the low-rank adapters are trained in the image encoder, while the text encoder remains frozen. Multi-level CLS token aggregation selects intermediate transformer layers, forms a token set with an additional learnable token, and passes it through a lightweight “FusionViT” so that the final image feature is the fused CLS token. Spherical feature interpolation augments two unit $i$5-normalized embeddings by spherical linear interpolation on the unit hypersphere. For each class $i$6, the prompt is defined as $i$7 “A photo of a ” $i$8 “, image description.” Training uses a symmetric CLIP contrastive loss together with cross-entropy classification loss, with total objective

$i$9

where $t$0 and $t$1 (Nguyen et al., 31 Aug 2025).

Training details are also specified. The system uses motion blur, Gaussian blur, color or contrast jitter, vertical flip, horizontal flip for symmetric classes, and random masking of $t$2 patches over approximately 10% of the area. Optimization runs for 50 total epochs with AdamW, learning rate $t$3, $t$4, default weight decay, and LoRA dropout $t$5. The batch size is not explicitly stated, and the hardware is not specified in the paper (Nguyen et al., 31 Aug 2025).

Reported results on ENTRep are strong across all three tasks. On the public test, the model achieves Accuracy $t$6, Precision $t$7, Recall $t$8, and F1 $t$9 for 7-way image classification; Recall@1 $p_i(t)$0 and MRR $p_i(t)$1 for image-to-image retrieval; and Recall@1 $p_i(t)$2 and MRR $p_i(t)$3 for text-to-image retrieval. On the private test, the reported values are Accuracy $p_i(t)$4 and F1 $p_i(t)$5 for classification, Recall@1 $p_i(t)$6 and MRR $p_i(t)$7 for image-to-image retrieval, and Recall@1 $p_i(t)$8 and MRR $p_i(t)$9 for text-to-image retrieval. Ablation results on the public split show a progression from baseline CLIP ViT-B/16 at Accuracy $n_i(t)$0, Recall@1 $n_i(t)$1, Recall@1 $n_i(t)$2, to $n_i(t)$3 at $n_i(t)$4, to $n_i(t)$5 at $n_i(t)$6, and finally to $n_i(t)$7 at $n_i(t)$8 (Nguyen et al., 31 Aug 2025).

The challenge overview provides a broader picture of method families and leaderboard behavior. For Task 1, WAS led the public test with 95.66 / 95.68 / 95.66 / 95.65 for Accuracy / Precision / Recall / F1 and maintained top private-test performance with 95.82% accuracy. For Task 2, Soft Mind_AIO led the public test with 94.51 / 96.88 for Recall@1 / MRR and the private test with 92.09% Recall@1. For Task 3, H3N1 led the public test with 95.11 / 97.52, while SoloL achieved 92.64% Recall@1 on the private test. Common methodological themes included anatomy-preserving augmentations, class-balanced sampling and focal-style loss, two-stage retrieval pipelines with gallery pruning by coarse classification, and minimal yet targeted ensembling and test-time augmentation (Nguyen et al., 6 Aug 2025).

6. ENTRep as intellectual management of enterprise

A third usage of ENTRep appears in “Intellectual Management of Enterprise,” where it denotes a thin-client platform that maps natural-language managerial requests to database queries, prediction modules, optimization routines, and adaptive-control inference. The major components are a Natural-Language Interface, Natural-Language Understanding, a Dialogue or Task Manager, a control-system database, mathematical engines for sales prediction and statistical analysis, a joint production-and-sales optimizer, an adaptive-control inference engine, and Natural-Language Generation. All manager questions are entered as free-form text; the NLU module uses object-oriented semantics to build discourse-level semantic frames; the Dialogue Manager routes each frame either to a simple database query, the sales-prediction engine, the joint-planning optimizer, or the adaptive-control engine; and results are returned in natural-language sentences, optionally augmented with charts (Ostapov, 2012).

The optimization module solves a joint production-and-sales planning problem. With decision variables $n_i(t)$9 for unit sale price, $\alpha_0=\beta_0=1$0 for sale volume, and $\alpha_0=\beta_0=1$1 for production volume, where $\alpha_0=\beta_0=1$2 and $\alpha_0=\beta_0=1$3 is opening stock, the objective is to maximize operating profit

$\alpha_0=\beta_0=1$4

The price–demand relation can be written either as a linear form,

$\alpha_0=\beta_0=1$5

or as a nonlinear regression $\alpha_0=\beta_0=1$6. Market constraints impose bounds on prices and volumes, and resource constraints are expressed through bottleneck-time inequalities or the matrix form $\alpha_0=\beta_0=1$7. If relations are linear, the system uses standard linear programming; if volumes are integer or price–demand curves are nonlinear, it uses mixed-integer or nonlinear programming. The prototype layers “statistical testing,” described as Monte-Carlo perturbation, on top of an LP kernel to explore scenario alternatives automatically (Ostapov, 2012).

Adaptive control is driven by internal state-of-plan streams and external critical events. A generic state representation is

$\alpha_0=\beta_0=1$8

where $\alpha_0=\beta_0=1$9 are event flags. When $$\alpha_i(t)=\alpha_0+p_i(t), \qquad \beta_i(t)=\beta_0+n_i(t),$$0, replanning is triggered: new resource constraints are asserted, the joint-planning optimizer is called again with updated capacities, and a short list of managerial decisions is produced. Although the exposition notes that one may view the process in a stochastic-decision-making perspective as a small Markov Decision Process, the prototype itself is rule-based with immediate re-optimization. The control-system database is implemented on MS SQL Server and includes ERP tables such as balans, money, finres, logist, docum, debitor, tovar, character, and calcul, as well as metadata tables for the NLU lexicon, grammar patterns, and semantic classes (Ostapov, 2012).

The reported case-study evidence is limited but explicit. The paper states that the prototype was deployed in a notebook-assembly enterprise and yielded simulated back-test gains over a 12-month horizon of profit improvement by 8%, sales-volume growth by 12%, accounts-receivable reduction by 20%, and monthly cash-flow variance down by 15%. In one Q3 2010 scenario, the joint optimizer recommended raising TN20A price by 5% in Region 1, lowering it by 3% in Region 2, reallocating 150 hours of bottleneck capacity from TN301 to TN20A, and changing the production mix to boost overall margin by 6 percentage points. The adaptive-control module reportedly caught a 10-day supplier delay of LCD panels and re-optimized the schedule within minutes, avoiding a projected 4% profit loss (Ostapov, 2012).

Taken together, these usages show that ENTRep denotes three different kinds of technical artifact: a Bayesian reputation model for enterprise expert ranking, a benchmark for bilingual ENT endoscopy classification and retrieval, and a natural-language enterprise management platform built on ERP data. The commonality is not domain but a recurrent emphasis on decision support under structured data, human-generated signals, and operational constraints.