nicolauduran45/multilingual-scientific-triplets

📚 Multilingual Scientific Triplet Dataset (EN–ES–CA)

A high-quality 50,148-triplet multilingual dataset for fine-tuning Sentence Transformers and other contrastive embedding models.

Designed specifically for scientific document retrieval, semantic search, and cross-lingual embedding alignment across three languages:

• English (EN)
• Spanish (ES)
• Catalan (CA)

---

🔬 Dataset Summary

This dataset contains triplets of the form:

(query, positive, hard_negative, negative)

constructed from scientific titles, abstracts, and title+abstract combinations.

Composition:

• 25,074 monolingual triplets
• 25,074 cross-lingual mixed triplets
• Balanced across EN / ES / CA
• Includes hard negatives
• Ideal for MultipleNegativesRankingLoss training

🔧 Dataset Construction Methodology

This dataset is built using the Microsoft Academic Graph (MAG) Fields of Study (FoS) Level-0 categories, which are the most fine-grained scientific topics in the MAG hierarchy.

Examples of Level-0 FoS include:

• Accounting
• Aeronautics
• Artificial intelligence
• Biochemistry
• Computer vision
• Cognitive psychology
• Civil engineering
• Control theory
• Dermatology
• Econometrics
• Environmental chemistry
• Data mining
• Criminology
• Endocrinology
• … and hundreds more

All Level-0 FoS names were translated into Catalan and Spanish using Google Translate to enable multilingual queries.

🔹 Query Construction (MAG Level-1 FoS Labels)

The query is always a MAG FoS Level-1 field name, in one of the three languages (EN, ES, CA):

Example:

• "Artificial intelligence"
• "Inteligencia artificial"
• "Intel·ligència artificial"

This ensures that queries represent precise scientific topics.

---

🔹 Positive Construction (Same Level-1 FoS)

The positive passage is:

• A scientific abstract (or title/abstract)
• Labeled with the exact same Level-1 FoS as the query

This provides strong semantic similarity.

Example

Query (Level-1 FoS): Biochemistry
Positive: Abstract from a paper labeled Biochemistry

---

🔸 Hard Negative Construction (Different Level-1, Same Level-0 Parent)

MAG organizes Level-0 FoS into parent Level-0 disciplines.

A hard negative is:

• From a different Level-1 FoS
• But with the same Level-0 parent discipline

This makes it:

• Domain-related → challenging
• Incorrect → good for contrastive separation

Example

• Level-0 parent: Chemistry
  • Level-1 FoS: Biochemistry → query/positive
  • Level-1 FoS: Chromatography → hard negative
  • (Same parent → Chemistry)

This ensures meaningful hard negatives.

---

🔹 Negative Construction (Different Parent Discipline)

A negative passage is:

• From a Level-0 FoS whose Level-0 parent is different from the query’s parent discipline
• Farther in the taxonomy → easier negative

Example

• Query Level-0: Chemistry
• Negative Level-0: Psychology or Computer Science

This encourages strong contrast and prevents semantic drift.

---

🧪 Summary of Sampling Rules

Component	MAG FoS Relationship to Query	Purpose
Query	Level-1 FoS name	Topic anchor
Positive	Same Level-1 FoS	True semantic match
Hard Negative	Different Level-1, same Level-0 parent	Challenging distractor
Negative	Different Level-0 parent	Domain-distant negative

This hierarchical sampling approach provides:

• High-quality contrastive supervision
• True scientific domain understanding
• Strong multilingual & cross-lingual alignment
• Better document retrieval performance

---

🌐 Multilingual Expansion

• All Level-1 FoS names originally in English were translated into:
  • Catalan
  • Spanish
• Abstracts (passages) remain in the original language of the paper
• Thus, queries are multilingual while scientific content remains realistic

This ensures:

• High-quality multilingual query formulation
• Real-world cross-lingual retrieval scenarios
• Balanced EN/ES/CA distributions

---

🧩 Data Fields

Column	Description
query	Query text (title, abstract, or combined)
positive	Semantically relevant passage
hard_negative	Hard negative passage from similar domain
negative	Distant, semantically unrelated passage
lang_query	Language of query (en, es, ca)
lang_positive	Language of positive passage
lang_hard_negative	Language of hard negative
lang_negative	Language of negative
type	monolingual or crosslingual

---

🌍 Linguistic Composition

Query Language Distribution

• EN: 17,012
• ES: 16,592
• CA: 16,544

Passage Language Balance

All passage types (positive, hard negative, negative) are evenly balanced across EN/ES/CA.

---

🔁 Monolingual vs Cross-Lingual Structure

50% Monolingual Triplets

• Query and all passages share the same language.
• Even distribution: EN/ES/CA (~8.3k each).
• Reinforces strong intra-language semantic learning.

50% Cross-Lingual Mixed Triplets

• Query language is random.
• Passages are independently sampled from EN/ES/CA.
• At least one passage differs in language from the query.
• Hard negatives and negatives may be from different languages.
• Achieves true multilingual alignment, not just pairwise translation.

---

🎯 Intended Use-Cases

This dataset is ideal for:

• 🧠 Multilingual Sentence Transformer fine-tuning
• 🌐 Cross-lingual scientific document retrieval
• 🔎 Semantic search engines
• 📚 Scientific knowledge embedding
• 🤖 Zero-shot multilingual retrieval
• 🧬 Domain-specific embedding learning

---

📈 Example Triplet

{
  "query": "Química física",
  "positive": "Presentamos una investigación teórica...",
  "hard_negative": "Estudi computacional de sistemes catalítics...",
  "negative": "Adaptive control for dynamic systems...",
  "lang_query": "es",
  "lang_positive": "es",
  "lang_hard_negative": "ca",
  "lang_negative": "en",
  "type": "crosslingual"
}

🧮 Dataset Statistics

• Total samples: 50,148
• Monolingual: 25,074
• Crosslingual: 25,074

Passage Lengths (tokens)

Field	Mean
Query	1.82
Positive	126.3
Hard Negative	108.9
Negative	136.1

---

🧪 Dataset Splits

Pre-defined splits:

Split	Size
train	40,118 (80%)
validation	5,015 (10%)
test	5,015 (10%)