Abelardo Carlos Martinez Lorenzo

Selected work

AI systems, from training to deployment

ImpactAI at the World Bank, domain entity linking, semantic parsing research code, and an incident-coordination app.

01The World Bank · 2024 – presentImpactAI: evidence systems for development policyA platform that turns impact-evaluation literature into structured, searchable evidence, then answers policy questions with retrieval, grouping and meta-analysis linked to the underlying studies.
02Sapienza NLP · Babelscape · The World Bank · 2020 – presentEncoder–decoder systems: from meaning graphs to tuned GeminiSequence-to-sequence models that write structured output. First, AMR parsers where each paper changed one part of the transformer: the output (CLAP), the attention (aligner), the encoder (LeakDistill), the input (AMRs Assemble!). Then generative models at scale: Gemini fine-tuned on Vertex AI, LLM-built datasets and LLM components in ImpactAI.
03The World Bank · 2024 – presentFine-tuned encoders: fast, cheap and domain-awareSmall encoders trained for one job each: a retriever and a reader that link text to a development-economics taxonomy in one pass, a relevance classifier, and an effect-size regressor, trained on data we created with LLMs.
04Research and ImpactAI · 2019 – presentTaxonomies and knowledge graphsStructuring a field’s knowledge in three layers: concepts (a taxonomy grown from millions of documents with LLMs and experts), relations (typed graphs from semantic parsing, in any language), and completion (missing links found by reasoning and by effect prediction).

Source-grounded by design

Claims should trace back to source records, passages, or structured evidence. The system should expose uncertainty instead of hiding it behind fluent text.

Structured before generative

Use extraction, entity linking, retrieval, and deterministic analysis where they make the system more inspectable; use LLMs where language flexibility is the real problem.

Evaluation as infrastructure

Benchmarks, traces, review bundles, and human calibration are part of product engineering, especially when answers influence research or policy decisions.

Research, made visible

Paying attention to cross-attention.

A scroll-through story of our ACL 2023 paper: how a Transformer parser, while turning a sentence into a meaning graph, already knows which words each concept comes from, in any language.

Explore the story All papers

Research profile

From multilingual NLP to evidence systems

My work sits between research and delivery: building systems that retrieve, interpret, and synthesize scientific evidence while remaining inspectable enough for real decision-making.

I care about domain-specialized language models, structured knowledge extraction, evidence synthesis, evaluation infrastructure, and AI systems that make public policy more rigorous and more accessible.

DCMI 2026 · PostersImpactAI · DCMIPresents ImpactAI, the World Bank’s generative AI evidence platform: a metadata layer over curated impact evaluations and a five-stage workflow (query understanding, retrieval, standardization, grounded generation, attribution) that returns source-grounded answers.

arXiv preprintQuery2EffectIntroduces Query2Effect, a benchmark of more than 72,000 synthetic natural-language queries, and a two-step framework that structures queries before predicting causal effect sizes.

VLDB 2025 Workshops · LS-NSLPoderosoPresents Poderoso, a modular framework combining knowledge-graph embeddings with rule-based reasoning for graph completion. The paper notes that the underlying work began during the author’s TU Berlin master’s thesis.

Findings of ACL 2024MSLIntroduces the Multilingual Semantic Layer and a semi-automatic multilingual dataset creation pipeline, then shows that manually refining 1,100 sentences across 11 languages helps reduce the semantic parsing performance gap for low-resource languages.

Source-grounded evidence systems

Building systems that retrieve, structure, and summarize rigorous findings while preserving source context.

Evaluation and quality gates

Designing benchmarks, traces, review workflows, and grounding checks for high-trust AI products.

Research-to-production NLP

Translating semantic parsing, entity linking, and scientific IE research into useful policy-facing systems.

About

Research, engineering, and policy in one track

Abelardo's work connects multilingual NLP and semantic parsing research with production systems for evidence synthesis. The same thread runs from Sapienza and Babelscape to the World Bank: make complex scientific knowledge easier to retrieve, structure, and use in practice.

Career path

Sep 2024 – Present · Washington, DC / Rome
Lead AI Scientist
The World Bank · ImpactAI
Lead an AI science team of six direct reports; own architecture, core backend implementation and code review for a conversational evidence-synthesis platform supporting meta-analysis and development policy.
Jan 2024 – Sep 2024 · Washington, DC / Rome
Senior AI Scientist
The World Bank
Designed and implemented a two-stage entity-linking system: an e5 retriever followed by a fine-tuned DeBERTa encoder that maps research text to an impact-evaluation taxonomy.
2023 – 2024 · Rome, Italy
NLP Researcher
Sapienza NLP · Sapienza University of Rome
Designed compact graph representations and encoder-decoder parsers in PyTorch Lightning, cutting training time by 82% and inference time by 78% versus SPRING while preserving parsing quality.
2020 – 2023 · Rome, Italy
NLP Researcher
Babelscape
Designed cross-lingual meaning representations and BART/T5 parsing and generation models; delivered Dockerized inference demos.

Full CV

Education

PhD in Deep Learning and Natural Language Processing

Sapienza University of Rome

Rome, Italy

Marie Skłodowska-Curie fellowship, Horizon 2020 KnowGraphs. Advisor: Roberto Navigli.

MSc in Big Data Management and Analytics

Erasmus Mundus Joint Programme · TU Berlin, UPC Barcelona, ULB Brussels

Berlin · Barcelona · Brussels

Final grade: 9.0/10.

BSc in Software Engineering

University of Málaga

Málaga, Spain

Final grade: 9.0/10.

MY JOURNEY · THE FILM

From Málaga to the world.

A three-minute pixel-art film of the path behind the career: a curious kid by the sea, a farm in Motril, Erasmus, a master across three cities, a PhD in Rome, and the ImpactAI team today.