FinNLP · FNP · LLMFinLegal · ACL 2025, pp. 207–213

AMWAL: Named Entity Recognition
for Arabic Financial News

The first domain-specific NER system for Arabic financial text — built from 26K articles, grounded in FIBO ontology, and achieving state-of-the-art performance across 20 entity types.

Muhammad S. Abdo  ·  Yash A. Hatekar  ·  Damir Cavar

Indiana University Bloomington

ARABIC NLP FINANCIAL NER ARABERT FIBO ONTOLOGY CORPUS LINGUISTICS SPACY MSA
Precision
96.08
vs. 91.00 (CamelBERT)
Recall
95.87
vs. 91.00 (CamelBERT)
F1 Score
95.97
vs. 80.00 (Wojood)
01Introduction & Motivation

Why Arabic Financial NER?

Financial news drives markets — predicting stock movements, measuring sentiment, informing investor decisions. Yet the overwhelming majority of NLP tools for financial text are English-only, leaving an enormous gap for Arabic, a language with 400M+ speakers and major financial centers from Morocco to the Gulf.

Even within Arabic NLP, existing NER systems are generic — built to recognize people, organizations, and countries. No domain-specific system existed for the financial domain until AMWAL.

The gap AMWAL fills: The first NER system specifically designed to extract financial entities from Arabic financial news, trained on a domain corpus, grounded in an international financial ontology (FIBO), and evaluated against state-of-the-art general Arabic NER systems.

What Makes Financial Arabic NER Hard?

🔠
Orthographic Variation
Diacritics, hamza, kashida create multiple spellings of the same token
🌐
Transliteration Chaos
Foreign company names have unpredictable, non-standardized Arabic renderings
🔀
Category Ambiguity
Company names overlap with months, nationalities, product names

The "Manufacturing" Problem — Orthographic Explosion

A single English word like "manufacturing" can produce many unpredictable Arabic transliterations — each a valid but distinct orthographic form:

تصنيعArabic translation
مانيوفاكتشرنجTransliteration v1
مانيوفاكتشرنغTransliteration v2
مانيوفكتشرنجTransliteration v3
صناعةAlternative
Implication: Rule-based systems that enumerate surface forms fail catastrophically. AMWAL adopts a corpus-driven lexical approach that captures entities as they naturally occur, regardless of which orthographic variant is used.

The "Nissan" Ambiguity

As a corporation
نيسان
Nissan (automotive brand)
In Levantine Arabic
نيسان
April (month name)

Identical spelling — different meaning depending on context. A key source of CORPORATION entity errors.

02Corpus Construction

Three Sources, 23 Years

26,231
Total articles collected
9.8M
Tokens in raw corpus
2000–2023
Time span covered
17,185
Annotated financial entities

Source Breakdown

Almal News: 11,012 articles; Al-Sharq: 8,106; Aljazeera business: 2,627; other 4,486.
Almal News — 11,012 articles (42%)
Al-Sharq — 8,106 articles (31%)
Aljazeera (Business) — 2,627 articles (10%)
Other / unclassified — 4,486 (17%)

Pre-processing Steps

Normalization reduces orthographic variation and prevents the model from treating the same token as different words:

① Remove All Diacritics

Strips vowel marks (harakat) so الزَّواجالزواج

② Normalize Hamza

All hamza variants → canonical form. Prevents إ / أ / آ / ا from counting as distinct tokens

③ Remove Kashida (Tatweel)

Strips decorative letter elongation: مبـلغمبلغ

These normalizations follow Hatekar & Abdo (2023) for consistency across the lab's Arabic NLP pipeline.
0320 Entity Types — Interactive Explorer
Ontology-grounded selection: Entities were not chosen arbitrarily — they derive from the Financial Industry Business Ontology (FIBO), supplemented with domain-relevant additions: BANK, GEOPOLITICAL, METRIC, STOCK EXCHANGE, MEDIA, and FINANCIAL MARKET.

Click any entity tile to explore it

Click an entity above to see its count, example tokens, and annotation challenges.

Entity Count Distribution — Figure 1

CORPORATION: 6840; QUANTITY OR UNIT: 2406; EVENT: 1417; PRODUCT OR SERVICE: 1222; PERSON: 1193; BANK: 1185...
04Semi-Automated Annotation & Training

The Two-Query Extraction Strategy

Using TXM (Textometry), two corpus queries were used to extract entities that are inherently labeled — the query structure itself provides the annotation context:

Query Pattern 1 — Hypernym–Hyponym

[Hypernym such as Hyponym]

بنوك مثل البنك الإسلامي الفلسطيني
"Banks such as Palestine Islamic Bank" → BANK labeled automatically
أدوات مالية مثل الأسهم
"Financial instruments such as stocks" → FINANCIAL INSTRUMENT labeled

Query Pattern 2 — Coordination

[Hyponym X and Hyponym Y]

بنك القاهرة وبنك الإسكندرية
"Cairo Bank and Alexandria Bank" → both labeled BANK
منتجات البترول والكيماويات
"Petroleum and chemical products" → PRODUCT OR SERVICE

Full Pipeline

Collect Corpus
26K articles from 3 newspapers (2000–2023)
Preprocess
Remove diacritics, normalize hamza & kashida
FIBO Entity Selection
20 entity types aligned to financial ontology
TXM Queries
Hypernym–Hyponym + coordination patterns
Frequency Analysis
Top-10 unigrams + bigrams as search seeds
Manual Review
17,185 entities verified for accuracy
Train AraBERT + spaCy
80/20 split, 20K steps, dropout 0.1

Training Configuration

Model:
Large AraBERT
Batch size:
50
Dropout:
0.1
Max steps:
20,000
Early stop:
patience=1600
Train/Test:
80% / 20%
Train files:
20,984
Test files:
5,247
Hardware:
1× GPU / 64GB
Why file-level split? Splitting at the article level (not randomly at the token level) ensures no overlapping context exists between training and test sets — preventing information leakage.
Frequency threshold: Only entities occurring ≥5 times in query results were retained, filtering noise while preserving genuine financial entities.
05Results — Interactive Comparison
Overall result: AMWAL achieves Precision 96.08 · Recall 95.87 · F1 95.97 — outperforming CamelBERT (91.00) and Wojood (80.00) and all cross-language financial NER systems cited in the paper.

Figure 2: System-Level Comparison

AMWAL outperforms both baselines across all three metrics.

Cross-Language Financial NER Context

These cross-language comparisons are contextual only, not direct benchmarks (different languages and corpora).

AMWAL (Arabic): 95.97; Chinese (Wang 2014): 92; German (Hillebrand 2022): ~88; French (Jabbari 2020): 73; Turkish (Dinç 2022): ~80; Arabic SMS (Kumar 2023): 65.4.

Per-Entity Performance — Hover to Compare

The table below shows F1 scores across all 20 entities for all three systems. Green = AMWAL wins; blue = competitive; orange = lower; gray = zero.

Entity AMWAL CamelBERT Wojood
P R F1 P R F1 P R F1

Figure 3: AMWAL vs CamelBERT — F1 by Entity

Most entities fall above the diagonal line, indicating AMWAL outperforms CamelBERT on nearly every entity type.

Points above the diagonal = AMWAL wins. Financial-domain entities (BANK, METRIC, STOCK EXCHANGE) show the largest gains.

06Error Analysis

CORPORATION — Lowest F1 (81)

Root cause: Company names semantically overlap with other entity types — products, services, nationalities, and even time references. The model cannot always resolve this ambiguity without broader discourse context.

Company ↔ Product overlap

يوروميد للصناعات الطبية
Euromed for Medical Industries — contains "Medical Industries" which can read as PRODUCT OR SERVICE

Company ↔ Nationality overlap

ويند إيطاليا
Wind Italy — contains "Italy" (NATIONALITY)

Company ↔ Month ambiguity

نيسان ← نيسان
Nissan (car brand) vs. April (month) — identical spelling in Arabic

PERSON — F1 (80)

Root cause: Arabic personal names sometimes include embedded nationality adjectives (nisba forms), blurring the PERSON / NATIONALITY boundary.

Name ↔ Nationality overlap

السويدي
"The Swedish" — could be a person's nisba surname OR a nationality label

High-Confidence Entity Types

CURRENCY
P:99 R:99 F1:99
99
TIME
P:99 R:99 F1:99
99
EVENT
P:98 R:98 F1:98
98
STOCK EXCHANGE
P:98 R:98 F1:98
98
FIN. INSTRUMENT
P:97 R:97 F1:97
97
COUNTRY
P:97 R:97 F1:97
97

Where Baselines Completely Fail — AMWAL's Biggest Margins

AMWAL uniquely handles financial-domain entities that CamelBERT and Wojood cannot recognize at all.
07Key Findings, Limitations & Future Work
🏆
Finding 1
First Arabic Financial NER
AMWAL is the first NER system specifically designed and trained for the Arabic financial domain. Its 20-category schema derived from FIBO is more comprehensive than any prior Arabic NER system.
📊
Finding 2
Domain Beats Generality
AMWAL's 95.97 F1 vs. CamelBERT's 91.00 and Wojood's 80.00 confirms that domain-specific training yields significant gains over general-purpose systems — even when the baseline uses the same backbone (AraBERT).
🔍
Finding 3
New Entities = Big Wins
AMWAL uniquely handles FINANCIAL MARKET, STOCK EXCHANGE, and GOVERNMENT ENTITY — all scoring 0 on both baselines. Domain specificity is the sole reason for coverage.
🌐
Finding 4
Cross-Language SOTA
At 95.97 F1, AMWAL outperforms financial NER systems in Chinese (92), French (73), Turkish (~80), and German (~88) — despite those languages having far more NLP resources than Arabic.
⚠️
Limitation 1
MSA Only
AMWAL is trained on Modern Standard Arabic from formal newspapers. It will not generalize to dialectal Arabic, social media text, or informal financial blogs without fine-tuning.
⚠️
Limitation 2
Category Overlap Errors
Corporation, Person, and Nationality categories are the most error-prone due to inherent ambiguity in Arabic naming conventions and transliterated entity names.
🧪
Limitation 3
Seen-Data Bias
High performance partly reflects strong tagging of entities seen during training. Generalization to genuinely novel entity mentions (zero-shot) remains a challenge shared with all NER systems.
🔬
Future Work 1
Hierarchical Entity Schema
Restructure flat 20-type schema into FIBO-aligned hierarchies (e.g., BANK as subtype of FINANCIAL INSTITUTION), enabling more nuanced representation.
🕸️
Future Work 2
Arabic Financial Knowledge Graph
The stated ultimate goal: extend from entity recognition to relation extraction, building a full Arabic financial KG to serve investors, regulators, and intelligence analysts.
📦
Future Work 3
Data Augmentation
Expand training set with more ambiguous/overlapping category examples and apply augmentation to improve robustness on Corporation and Person edge cases.
Open source: AMWAL's best model, training files, and test files are available on GitHub: https://github.com/Muhsabrys/AMWAL/