nicolauduran45/multilingual_research_pairs

🧩 Multilingual Research Pairs (CA–ES–EN)

📘 Dataset Summary

This dataset provides multilingual scientific and research-oriented text pairs designed for retrieval, semantic similarity, and cross-lingual embedding fine-tuning tasks.

Each sample links a query (a keyword or natural-language question/claim) with a text (typically a publication title + abstract or a project description).
Pairs are created in Catalan (CA), Spanish (ES), and English (EN), with balanced monolingual and cross-lingual directions.

The dataset includes detailed linguistic and semantic metadata — such as keyword length, lexical vs semantic match, and frequency quartiles — enabling fine-grained evaluation across languages and domains.

---

⚙️ Dataset Construction

1️⃣ Source and Sampling

The dataset originates from scientific publications and research project descriptions containing multilingual metadata (titles, abstracts, and keywords).
To ensure diversity and multilingual balance:

• We sampled 10,000 documents each for Catalan, Spanish, and English.
• Keywords were sampled from the English metadata as the initial reference set.
• Corresponding titles and abstracts were paired with these keywords to form the English baseline corpus.

---

2️⃣ Machine Translation for Multilingual Alignment

To extend coverage across languages, Google Translate was used to generate high-quality translations of:

• Titles
• Abstracts
• Keywords (English → Spanish / Catalan)

This ensured that every English document has aligned multilingual versions (text_lang ∈ {en, es, ca}) with consistent semantic structure.

All text was then normalized and truncated to a maximum of 512 tokens.

---

3️⃣ Cleaning and Normalization

Steps:

• Remove HTML entities
• Preserve accents and case (for multilingual consistency)
• Deduplicate using MD5 of abstract (abs_hash)
• Add text_lang column (en, es, or ca)

---

4️⃣ Multilingual Keyword Explosion

Keywords from all three languages were split and aligned across positions, then exploded so each row corresponds to a single multilingual keyword triplet:

Column	Description
keyword_en	English keyword
keyword_es	Spanish keyword
keyword_ca	Catalan keyword

This process created a multilingual keyword-level dataset supporting both monolingual and cross-lingual retrieval.

---

5️⃣ Metadata Enrichment

Each keyword instance was enriched with linguistic and statistical metadata:

Feature	Description
token_count	number of tokens in English keyword
length_type	short (1 token), medium (2–3 tokens), long (≥ 4)
match_type	whether the keyword appears lexically in title/abstract
count	raw frequency across all English keywords
freq	frequency quartile (1 = rare → 5 = very frequent)

Example statistics:

• 🧠 Lexical vs Semantic: 54.7 % semantic / 45.3 % lexical
• 🔠 Length: 59.4 % medium / 34.2 % short / 6.4 % long
• ⚙️ Frequency quartiles: evenly distributed across 1–5

---

6️⃣ Pair Generation

Pairs were generated with a balanced mono vs cross-lingual setup:

Pair type	Description	Proportion
keyword→text	keyword (any language) → (title + abstract / title / abstract)	~90 %
title→abstract	title → abstract	~10 %

Each pair retains its abs_hash (document ID) and dataset split (train, dev, test).
All six bilingual directions are represented (EN↔ES, EN↔CA, ES↔CA).

---

7️⃣ Query Variants (Natural-Language Prompts)

To simulate real information-retrieval behavior, each query was expanded using language-specific templates (questions, definitions, and research claims).

Examples:

• ❓ “What is {q}?”
• 💬 “Applications of {q}.”
• 📈 “Studies show that {q} improves performance.”

Acronyms (e.g., MC, CSI) are preserved, while regular capitalized words are normalized
(Embarazo molar parcial → embarazo molar parcial).

---

📊 Dataset Statistics

✅ Dataset Sizes

Split	# Pairs
Train	61 083
Dev	7 665
Test	7 618

🔍 Pair Type Distribution

pair_type	%
keyword→text	89.9
title→abstract	10.1

🌍 Language Distribution (query_lang)

Language	%
ca	33.4
en	33.3
es	33.3

---

Keyword & Semantic Statistics

Keyword Length Distribution (%)

length_type	percentage
medium	59.4
short	34.2
long	6.4

Lexical vs Semantic Distribution (%)

match_type	percentage
semantic	54.7
lexical	45.3

Frequency Quartiles (1 = rare → 5 = very frequent)

freq	percentage
1	29.0
2	13.9
3	17.4
4	19.7
5	20.0

---

📁 Data Fields

Column	Description
query	keyword or generated natural-language query
text	title + abstract, title, or abstract text
query_lang	ISO-639-1 language code of query
text_lang	ISO-639-1 language code of text
pair_type	“keyword→text” or “title→abstract”
abs_hash	MD5 hash of abstract (unique document ID)
split	train / dev / test
token_count	number of tokens in English keyword
length_type	short / medium / long
match_type	lexical / semantic
count	raw keyword frequency
freq	frequency quartile (1–5)
all_positive_keywords	concatenation of all unique multilingual keywords
query_variant	generated natural-language query variant

---

🧠 Example Record

Field	Example
query	Base de dades espacial
text	MEDIS — Una base de datos espacial completa sobre…
query_lang	ca
text_lang	ca
pair_type	keyword→text
length_type	medium
match_type	lexical
freq	3
all_positive_keywords	biodiversity hotspot; biogeography; conservation biology; database
query_variant	Beneficis de base de dades espacial.

---

🎯 Intended Uses

• Fine-tuning multilingual retrieval and sentence-embedding models (e.g. SentenceTransformers, E5, mBGE)
• Evaluating cross-lingual semantic search and R&I retrieval
• Studying keyword semantics, lexical vs semantic matching, and multilingual IR