Papers by Els Lefever
Lemmatisation of Medieval Greek: Against the Limits of Transformer’s Capabilities? (2024.lrec-main)
Copied to clipboard
| Challenge: | Existing lemmatisation algorithms display an accuracy drop of around 30pp when tested on unedited, Byzantine Greek epigrams. |
| Approach: | They propose to use transformer-based embeddings and a dictionary look-up to lemmatise unedited, Byzantine Greek epigrams. |
| Outcome: | The proposed method outperforms existing methods and provides detailed error analysis revealing why unedited, Byzantine Greek is so challenging for lemmatisation. |
Discovering the Language of Wine Reviews: A Text Mining Account (L18-1)
Copied to clipboard
| Challenge: | odors and flavors are often expressed in wine reviews, but they are often not. |
| Approach: | They use a corpus of wine reviews to find out what wine is like in a review . they use lexical bag-of-words features, domain-specific terminology features and word embedding features to train machine learning. |
| Outcome: | The proposed model predicts the wine's color, grape variety, and country of origin based on the review text alone. |
A Gold Standard for Multilingual Automatic Term Extraction from Comparable Corpora: Term Structure and Translation Equivalents (L18-1)
Copied to clipboard
| Challenge: | Terms are notoriously difficult to identify, both automatically and manually. |
| Approach: | They propose a method to annotate terms manually from a comparable corpus . they show that the gold standard provides a tool for evaluation and a rich source of information . |
| Outcome: | The proposed method provides a tool for evaluation and rich source of information about terms. |
When the Student Becomes the Master: Learning Better and Smaller Monolingual Models from mBERT (2022.coling-1)
Copied to clipboard
| Challenge: | Using a jointly trained model for 102 languages, monolingual models outperform the original model. |
| Approach: | They propose to distill monolingual models from a jointly trained model for 102 languages using a text corpus of 160 GB. |
| Outcome: | The proposed model outperforms the original model for 6 languages with varying amounts of resources and language families. |
Human and System Perspectives on the Expression of Irony: An Analysis of Likelihood Labels and Rationales (2024.lrec-main)
Copied to clipboard
| Challenge: | a new study examines the recognition of irony by humans and automatic systems . a fine-grained annotation scheme allows for improved modeling of ironity in automatic systems. |
| Approach: | They propose a fine-grained annotation scheme that allows for better recognition of irony by humans and automatic systems. |
| Outcome: | The proposed model improves on tweets annotated with high confidence and agreement . it also performs better on high-confidence and highagreement samples compared to automated systems . |
At the Crossroad of Cuneiform and NLP: Challenges for Fine-grained Part-of-speech Tagging (2024.lrec-main)
Copied to clipboard
| Challenge: | cuneiform texts are dominated by multiple languages and language families . the most dominant language written in cuniform is the Semitic Akkadian . existing cnl models are not suitable for digital editions of Akkadi . |
| Approach: | They focus on letters written in the Semitic Akkadian, a cuneiform language dominated by cuniform texts . they propose to use pre-trained embeddings, sentence segmentation and cnl to fine-tune language models . |
| Outcome: | The dominant language written in cuneiform is the Semitic Akkadian . the paper examines the input material and tries to initiate a discussion about best-practices . |
Lemmatisation & Morphological Analysis of Unedited Greek: Do Simple Tasks Need Complex Solutions? (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing systems for part-of-speech tagging of unedited Greek text have outperformed traditional methods for morphological analysis and lemmatisation. |
| Approach: | They propose to combine nominal features into a single label and combine the three most distinctive features of verbs into another unified label. |
| Outcome: | The proposed models outperform traditional models in lemmatisation and morphological analysis and show that multi-task learning improves performance by transferring parameters. |
MIsA: Multilingual “IsA” Extraction from Corpora (L18-1)
Copied to clipboard
| Challenge: | In this paper, we present a collection of hypernymy relations extracted from the Wikipedia corpus in five languages. |
| Approach: | They present a collection of hypernymy relations extracted from the Wikipedia corpus in five languages . they use existing or newly defined lexico-syntactic patterns to extract hyperniyms . |
| Outcome: | The proposed tool is based on a dictionary extracted from the full Wikipedia corpus. |
Misery Loves Complexity: Exploring Linguistic Complexity in the Context of Emotion Detection (2023.findings-emnlp)
Copied to clipboard
| Challenge: | a negative emotion is a cognitive bias that affects how we express thoughts and opinions online . a recent study shows that negative words generate more engagement and clicks than positive ones . |
| Approach: | They propose to use readability and linguistic complexity metrics to better understand emotions . they propose to fine-tune three state-of-the-art transformers to detect emotions based on a dataset . |
| Outcome: | The proposed model fails to predict emotions on complex texts, the authors show . they also show that more advanced models fail to predict complex texts . |
Evaluating Transformers for OCR Post-Correction in Early Modern Dutch Theatre (2025.coling-main)
Copied to clipboard
| Challenge: | a new study examines the effectiveness of two types of transformer models for OCR post-correction in early modern Dutch plays. |
| Approach: | They propose to use large generative models and sequence-to-sequence models for OCR post-correction in early modern Dutch plays. |
| Outcome: | The proposed model outperforms generative models on the OCR post-correction task . the model outpersforms the model with the lowest error rate on the historical English dataset . |
EnerGIZAr: Leveraging GIZA++ for Effective Tokenizer Initialization (2025.findings-acl)
Copied to clipboard
| Challenge: | Continual pre-training has long been considered the default strategy for adapting models to non-English languages, but struggles with initializing new embeddings, especially for non-Latin scripts. |
| Approach: | They propose a method that leverages statistical word alignment techniques to improve continual pre-training by leveraging word alignment matrix between source and target tokens. |
| Outcome: | The proposed method outperforms existing methods on key NLP tasks including POS tagging, Sentiment Analysis, NLI, and NER in Hindi, Basque, Arabic and Korean. |
Lost in Activations: A Neuron-level Analysis of Encoders for Cross-Lingual Emotion Detection (2026.eacl-short)
Copied to clipboard
| Challenge: | XLM-R models for multilingual emotion classification are still lacking in understanding of their internal decision-making mechanisms. |
| Approach: | They propose to use neuron-level activation analysis to study the inter-language differences between neurons. |
| Outcome: | The proposed model consistently encodes emotion-related concepts across languages, but others show strong monolingual specialization. |
Identifying Cognates in English-Dutch and French-Dutch by means of Orthographic Information and Cross-lingual Word Embeddings (2020.lrec-1)
Copied to clipboard
| Challenge: | Existing methods to identify cognate pairs in English-Dutch and French-Dutsch combine orthographic information with cross-lingual word embeddings. |
| Approach: | They combine traditional orthographic information with cross-lingual word embeddings to identify cognate pairs in English-Dutch and French-Dutsch. |
| Outcome: | The proposed classifier achieves good results on the basis of orthographic information but improves by including semantic information in the form of cross-lingual word embeddings. |