JEL: A Multi-Domain Research Acronym

• JEL is a polysemous acronym that denotes distinct methodologies in enterprise NLP entity linking, economic classification, formal logic, and nonparametric inference.
• In enterprise NLP, JEL employs multi-network entity linking using fuzzy matching and context-aware representations to efficiently map news mentions to knowledge graphs.
• As a statistical method, JEL uses jackknife pseudo-values to linearize nonlinear constraints in U-statistics, facilitating robust likelihood-based inference with chi-square calibration.

JEL is a polysemous acronym with established meanings in several research communities. In contemporary technical literature it denotes, at minimum, an enterprise entity-linking architecture for mapping news mentions to knowledge-graph entities, the Journal of Economic Literature classification system used in economics, a formal Justification and Explanation Logic for multi-agent reasoning, and a family of jackknife empirical likelihood methods in nonparametric statistics (Kishelev et al., 9 Sep 2025, Heikkila, 2022, Letia et al., 2013, Chen et al., 2016). The term therefore cannot be interpreted without disciplinary context.

1. Acronymic scope and disciplinary disambiguation

Domain	Expansion	Core referent
Enterprise NLP	JEL	Neural entity-linking system for news and knowledge graphs
Economics	Journal of Economic Literature	Hierarchical classification code system
Multi-agent logic	Justification and Explanation Logic	Logic of justification, explanation, and commitments
Nonparametric inference	Jackknife Empirical Likelihood	EL methodology built on jackknife pseudo-values

These senses are unrelated in origin, formalism, and application, but each is technically mature within its own field (Kishelev et al., 9 Sep 2025, Heikkila, 2022, Letia et al., 2013, Chen et al., 2016). In NLP, JEL is a named system developed at JPMorgan to connect named entity mentions in news to a corporate knowledge graph. In economics, JEL denotes a subject taxonomy maintained by the American Economic Association. In formal logic, JEL is an axiomatic calculus combining justificatory and explanatory reasoning through commitment-based proof terms. In statistics, JEL is a modification of empirical likelihood that replaces nonlinear constraints by jackknife pseudo-values, especially for U-statistics.

The ambiguity is reinforced by the fact that each usage is internally canonical. Economics papers may state only “JEL codes”; statistical papers refer to “JEL ratio tests” or “JEL-based inference”; multi-agent systems papers use JEL as the name of a logic; and recent enterprise-NLP papers use JEL as a product-style model name. A common source of confusion is therefore to treat the economics expansion as globally dominant, even though current arXiv usage spans several incompatible meanings.

2. JEL as an enterprise entity-linking model

In enterprise NLP, JEL is JPMorgan’s end-to-end, multi-neural-network entity-linking model designed to connect named entity mentions in news to entities in a corporate knowledge graph (Kishelev et al., 9 Sep 2025). The problem is formulated over a knowledge graph $G=(V,E)$, a news document $D=(t_1,\dots,t_N)$, a mention set $M=\{m_1,\dots,m_K\}$, and context tokens $C_i$ surrounding mention $m_i$. A blocking layer first generates a candidate set $\Gamma(m_i)\subset V$, after which the system selects

$e^*=\arg\max_{e\in\Gamma(m_i)} s(m_i,e,C_i),$

optionally returning NIL when the best score falls below a link threshold.

The later system description emphasizes a four-component architecture: mention detection by NER, candidate generation by fuzzy matching, representation learning, and candidate ranking (Kishelev et al., 9 Sep 2025). Blocking averages cosine, Levenshtein, and Jaro similarities,

$S_{\text{block}}(m_i,e)=\frac{S_{\cos}+S_{\lev}+S_{\jaro}}{3},$

and retains candidates with $S_{\text{block}}\ge \tau_{\text{block}}$, with $\tau_{\text{block}}=0.5$ in the pipeline. Representation learning combines hierarchical surface signals and semantic signals. Surface representations proceed from character-level embeddings to word-level embeddings and then to an entity-level name embedding. Mention-side semantics are produced by BERT over local news context, while entity-side semantics are produced by a TripletNet trained on short entity descriptions using fastText vectors and triplet loss. Because entity descriptions are short, the paper explicitly states that BERT may be less effective there, motivating the TripletNet choice. The fused mention and entity vectors are projected and scored by a two-layer feed-forward network,

$D=(t_1,\dots,t_N)$0

The paper emphasizes local rather than global disambiguation: there is no explicit document-level coherence module such as a CRF or graph optimization layer (Kishelev et al., 9 Sep 2025). Efficiency derives from reducing the naive $D=(t_1,\dots,t_N)$1 search to $D=(t_1,\dots,t_N)$2 with $D=(t_1,\dots,t_N)$3, combined with a lightweight comparator network. Specific latency and throughput numbers are not reported, but the system is described as computationally efficient and productionized within the Galileo News Analytics platform, with an internally released SDK.

The evaluation data are drawn from Dow Jones news and the DaVinci People Graph, which integrates internal systems with third-party sources including WealthX and FactSet (Kishelev et al., 9 Sep 2025). Two datasets are described: a high-confidence set built with fuzzy similarity at least $D=(t_1,\dots,t_N)$4 and a low-confidence set built with similarity at least $D=(t_1,\dots,t_N)$5. On the high-confidence benchmark with training/testing sizes $D=(t_1,\dots,t_N)$6, BLINK achieved Accuracy $D=(t_1,\dots,t_N)$7, Precision $D=(t_1,\dots,t_N)$8, Recall $D=(t_1,\dots,t_N)$9, and F1 $M=\{m_1,\dots,m_K\}$0, whereas JEL achieved Accuracy $M=\{m_1,\dots,m_K\}$1, Precision $M=\{m_1,\dots,m_K\}$2, Recall $M=\{m_1,\dots,m_K\}$3, and F1 $M=\{m_1,\dots,m_K\}$4. On clean versus noisy data, JEL recorded F1 $M=\{m_1,\dots,m_K\}$5 and AUC $M=\{m_1,\dots,m_K\}$6 on the clean split, and F1 $M=\{m_1,\dots,m_K\}$7 and AUC $M=\{m_1,\dots,m_K\}$8 on the noisy split. The paper attributes the gain over BLINK to the joint use of surface and semantic features.

An earlier system paper described a closely related JPMorgan JEL architecture in somewhat different terms, using spaCy NER, lexical overlap blocking, triplet-loss entity embeddings from minimal context, two short LSTMs with attention for mention semantics, and a wide-and-deep Siamese matcher that combines a linear character model with a semantic distance model (Ding et al., 2024). That version reported Precision $M=\{m_1,\dots,m_K\}$9, Recall $C_i$0, F1 $C_i$1, and Accuracy $C_i$2 on its test set for company-name linking in financial news. This suggests an evolving model family rather than a single frozen architecture. Across both papers, the consistent design principle is the fusion of appearance-based string evidence with context-sensitive semantics, tailored to enterprise entities that do not fit Wikipedia-centric entity-linking assumptions.

3. JEL as the Journal of Economic Literature classification system

In economics, JEL refers to the Journal of Economic Literature classification system, a domain-specific, hierarchical, alphanumeric scheme developed and maintained by the American Economic Association to organize economics research (Heikkila, 2022). It is the de facto standard used in EconLit to classify articles, dissertations, books, book reviews, and working papers. The AEA launched the Journal of Economic Literature in 1969, and the first JEL code version appeared in the same year. A major redesign in 1988–1990 culminated in the 1991 system, introducing the modern letter-level first tier and removing the long-debated “Theory” category.

The current system is hierarchical. At the top level it has 20 lettered main classes $C_i$3–$C_i$4; the second level appends one digit; the third level, most commonly used for indexing, is a letter plus two digits. The paper formalizes this as a hierarchy $C_i$5, with second-level sets $C_i$6 and third-level sets $C_i$7 (Heikkila, 2022). EconLit associates each document $C_i$8 with a multiset $C_i$9 of at most six three-digit codes. The system now comprises more than 850 three-digit subclasses, with examples including D12 Consumer Economics: Empirical Analysis, E71 Macro-Based Behavioral Economics, G12 Asset Markets and Pricing, H26 Tax Evasion and Avoidance, J24 Human Capital; Skills, Q54 Climate; Natural Disasters and Their Management; Global Warming, and Z13 Economic Sociology; Economic Anthropology; Language; Stratification.

EconLit itself is a large professional indexing infrastructure. The early Economic Literature Index was stored on magnetic tapes and mounted on DIALOG in 1981; the CD-ROM version was branded EconLit in 1991 (Heikkila, 2022). The database is updated weekly, includes over 1.6 million records spanning about 130 years and 74 countries, and as of February 2022 listed more than 2015 distinct ISSNs, with current coverage for about 1455 titles. Books historically received one or two categories, whereas articles can receive more cross-classifications.

The system is widely used beyond EconLit. Many publishers require or encourage JEL codes, often asking authors to provide up to six standard codes (Heikkila, 2022). IDEAS/RePEc and SSRN support browsing or filtering by JEL codes, and some journals provide “Browse by JEL code” interfaces. The codes also support bibliometric analyses, field delineation, and methodological controls in econometric studies. The literature summarized in the review notes, for example, that Microeconomics ($m_i$0) has historically been the most common first code, and that the breadth of a paper’s JEL portfolio has been used as a proxy for “innovational strength.”

JEL codes also function as topical metadata in applied research. A paper on FDI, banking crises, and growth explicitly uses F65, F36, G01, and G15, corresponding respectively to International Financial Markets: Globalization, Financial Aspects of International Macroeconomics, Financial Crises, and International Financial Markets (Gaies et al., 2019). Another paper connects Sustainable Development Goals to JEL categories and proposes a practical crosswalk, for example mapping SDG 13 to Q54, Q15, and Q58, and SDG 1 to I32, O15, and H53 (Heikkilä, 2020). This suggests that JEL classification is both a bibliographic control system and a substantive ontology for organizing economics research agendas.

4. JEL as Justification and Explanation Logic

In formal logic and multi-agent systems, JEL denotes Justification and Explanation Logic, an axiomatic framework introduced to unify justificatory reasoning and explanatory reasoning in institutionally governed interaction (Letia et al., 2013). The paper distinguishes normative discourse, grounded in obligations, rights, and institutional rules, from explicative discourse, grounded in preferences, intentions, capabilities, and situational facts. JEL integrates both within a justification-logic-inspired calculus and makes social commitments first-class proof terms.

The syntax contains proof terms $m_i$1 and formulas $m_i$2. Proof terms include variables, constants, application, sum, proof checkers, negative proof checkers, and a strength operator: $m_i$3 Formulas include propositional atoms, disjunction, negation, and two central modalities: $m_i$4 The intended readings are “$m_i$5 is a justification for $m_i$6” and “$m_i$7 is an explanation for $m_i$8.” The logic also uses commitments as proof terms. A conditional commitment $m_i$9 means that debtor $\Gamma(m_i)\subset V$0 promises creditor $\Gamma(m_i)\subset V$1 to bring about $\Gamma(m_i)\subset V$2 under condition $\Gamma(m_i)\subset V$3.

The axiomatics extend propositional logic with justification and explanation principles (Letia et al., 2013). Justification application is captured by

$\Gamma(m_i)\subset V$4

while explanation propagation is handled by parallel axioms in which explanatory status dominates: if any step is explanatory, only an explanation of the conclusion is obtained. Justification is monotonic under sum, but no analogous monotonicity axiom is given for explanations, which allows explanatory reasoning to remain non-monotonic. Additional axioms provide positive and negative introspection through $\Gamma(m_i)\subset V$5 and $\Gamma(m_i)\subset V$6.

A distinctive feature is the agent-indexed use of commitments. Expressions such as $\Gamma(m_i)\subset V$7 and $\Gamma(m_i)\subset V$8 capture justification and explanation from agent $\Gamma(m_i)\subset V$9’s perspective (Letia et al., 2013). The paper develops reusable discourse patterns, including Gratuitous Promise and Request/Directive as justification patterns, and Cognitive Consistency and Preferred Commitment as explanation patterns. Strength and preference over commitments guide conflict resolution; facts can be stronger than commitments to bring those facts about, and creditors and debtors may rank commitments differently.

The framework is presented axiomatically rather than semantically. The paper explicitly notes that it does not provide a formal satisfaction relation, Kripke frames, evidence functions, or a full model theory (Letia et al., 2013). It likewise omits completeness, decidability, and complexity results. Its contribution is instead a commitment-centric proof system for producing and evaluating justificatory and explanatory arguments in multi-agent settings, especially business interactions involving deadlines, promises, requests, and exceptions.

5. JEL as jackknife empirical likelihood

In nonparametric inference, JEL most commonly denotes jackknife empirical likelihood, a modification of empirical likelihood designed to handle nonlinear estimating equations, especially those arising from U-statistics (Chen et al., 2016). For a U-statistic $e^*=\arg\max_{e\in\Gamma(m_i)} s(m_i,e,C_i),$0 estimating a scalar parameter $e^*=\arg\max_{e\in\Gamma(m_i)} s(m_i,e,C_i),$1, the method forms jackknife pseudo-values

$e^*=\arg\max_{e\in\Gamma(m_i)} s(m_i,e,C_i),$2

where $e^*=\arg\max_{e\in\Gamma(m_i)} s(m_i,e,C_i),$3 is the leave-one-out statistic. Empirical likelihood is then applied to the linear moment condition $e^*=\arg\max_{e\in\Gamma(m_i)} s(m_i,e,C_i),$4, with

$e^*=\arg\max_{e\in\Gamma(m_i)} s(m_i,e,C_i),$5

and ratio statistic

$e^*=\arg\max_{e\in\Gamma(m_i)} s(m_i,e,C_i),$6

Under regularity conditions, the paper proves the Wilks-type limit

$e^*=\arg\max_{e\in\Gamma(m_i)} s(m_i,e,C_i),$7

A central refinement is adjusted jackknife empirical likelihood (AJEL), which addresses the convex-hull restriction by augmenting the pseudo-values with

$e^*=\arg\max_{e\in\Gamma(m_i)} s(m_i,e,C_i),$8

where $e^*=\arg\max_{e\in\Gamma(m_i)} s(m_i,e,C_i),$9; the paper recommends $S_{\text{block}}(m_i,e)=\frac{S_{\cos}+S_{\lev}+S_{\jaro}}{3},$0 (Chen et al., 2016). AJEL preserves the same first-order $S_{\text{block}}(m_i,e)=\frac{S_{\cos}+S_{\lev}+S_{\jaro}}{3},$1 limit while guaranteeing feasibility for all $S_{\text{block}}(m_i,e)=\frac{S_{\cos}+S_{\lev}+S_{\jaro}}{3},$2. A further robustification is weighted JEL (WJEL), which assigns smaller weights to outliers using data depth and yields a statistic converging to a multiple of a chi-square variable; a self-normalized version restores a standard chi-square limit (Sang et al., 2019).

The methodology has been specialized to a wide range of inferential targets in the supplied literature. These include confidence intervals and tests for probability weighted moments (Bhati et al., 2018), S-Gini indices (N et al., 2017), a new generalized income inequality measure based on extrema of $S_{\text{block}}(m_i,e)=\frac{S_{\cos}+S_{\lev}+S_{\jaro}}{3},$3 draws (Kattumannil et al., 2023), categorical Gini correlation (Hewage et al., 2023), log symmetry (S et al., 2024), equality of upper semivariance (Suresh et al., 2024), independence between a continuous and a categorical variable through modified categorical Gini covariance (Suresh et al., 2024), independence of time to failure and cause of failure in competing risks (N. et al., 2021), multivariate-sample U-statistics such as differences in VUS measurements (Garg et al., 2024), a divergence measure based on survival extropy (Garg et al., 21 Jul 2025), and poverty indices such as the Sen and Sen–Shorrocks–Thon indices (N et al., 18 Mar 2026). Across these applications, the recurring attraction of JEL is variance-free or variance-light likelihood-based inference for nonlinear functionals with asymptotic $S_{\text{block}}(m_i,e)=\frac{S_{\cos}+S_{\lev}+S_{\jaro}}{3},$4 calibration.

The application papers are explicit about practical trade-offs. Normal-based tests often require difficult plug-in variance estimators; standard empirical likelihood may face nonlinear constraints or high computational cost; and bootstrap calibration can be expensive (S et al., 2024, Garg et al., 2024). JEL addresses these issues by linearizing the constraint through pseudo-values. This does not eliminate all difficulties—leave-one-out recomputation can still be expensive, and weighted variants need a defensible weighting scheme—but it explains why “JEL ratio test” has become a recognizable phrase across recent arXiv statistics papers.

6. Cross-domain significance and persistent ambiguity

The four major senses of JEL occupy distinct methodological regimes. The enterprise-NLP JEL is an operational model family for entity disambiguation in news streams and proprietary knowledge graphs (Kishelev et al., 9 Sep 2025). The economics JEL is a hierarchical metadata system for classification, discovery, and bibliometric analysis (Heikkila, 2022). The logic JEL is an axiomatic proof system centered on commitments, explanations, and justifications (Letia et al., 2013). The statistical JEL is a nonparametric inference technique for U-statistics and related estimating equations (Chen et al., 2016).

Because these uses are structurally unrelated, abbreviation expansion is indispensable. A common misconception is that JEL automatically means the Journal of Economic Literature codes. The current literature shows otherwise: “JEL” may denote a production entity-linking pipeline, a formal logic, or a class of empirical-likelihood methods, depending on venue and disciplinary context (Kishelev et al., 9 Sep 2025, Letia et al., 2013, Chen et al., 2016). Another misconception is that the statistical usage is a single test; in fact, it is a methodology with many variants, including AJEL and WJEL, and with applications ranging from dependence measures to poverty indices.

A plausible implication is that acronym-level search and indexing are intrinsically lossy when expansions are omitted. In bibliographic systems, enterprise model names, formal logics, classification schemes, and inference frameworks all compete for the same three-letter token. For technical writing, the safest convention is therefore explicit first-use expansion together with domain cues such as “JEL codes,” “JEL ratio test,” or “JEL entity linking.” Across fields, the acronym survives because each community has stabilized a local meaning; across fields, it remains ambiguous because no single expansion is globally dominant.