Papers by Anders Søgaard

Antonia Karamolegkou, Jiaang Li, Li Zhou, Anders Søgaard

Challenge:	CWEB is a new benchmark for grammatical error correction (GEC) systems . website data contains far fewer grammamatical errors than learner essays .
Approach:	They propose to broaden the target domain of grammatical error correction (GEC) systems . website data contains far fewer grammamatical errors than learner essays .
Outcome:	The proposed model can't rely on a strong internal language model in low error density domains.

Copied to clipboard

How Do Multilingual Language Models Remember Facts? (2025.findings-acl)

Challenge:	a recent study examines the extent to which language models can memorize training data . a fair use exemption to copyright laws allows for limited use of copyrighted material .
Approach:	They examine the extent to which language models can redistribute copyrighted text . they use a range of popular books and coding problems to study copyright violations .
Outcome:	This study examines the extent to which language models can redistribute copyrighted text . it shows that language models may memorize entire chunks of training data .

Copied to clipboard

Constanza Fierro, Negar Foroutan, Desmond Elliott, Anders Søgaard

Unsupervised Cross-Lingual Representation Learning (P19-4)

Challenge:	Prior research has focused on English monolingual models, but how these mechanisms generalize to non-English languages remains unexplored.
Approach:	They analyze three multilingual LLMs to find out how they can generalize recall mechanisms . they find that subject enrichment is language-independent, object extraction is language dependent .
Outcome:	The proposed model performs better in multilingual contexts than in English models . the model is more efficient in multi-lingual context, but it is more complex in multilinguistic models compared to English models.

Copied to clipboard

Sebastian Ruder, Anders Søgaard, Ivan Vulić

We Need To Talk About Random Splits (2021.eacl-main)

Challenge:	a comprehensive survey of cutting-edge weakly-supervised and unsupervised cross-lingual word representations is presented .
Approach:	This tutorial provides a comprehensive survey of recent work on weakly-supervised and unsupervised cross-lingual word representations.
Outcome:	This tutorial provides a comprehensive survey of cutting-edge weakly-supervised and unsupervised word representations.

Copied to clipboard

Anders Søgaard, Sebastian Ebert, Jasmijn Bastings, Katja Filippova

Sluice Resolution without Hand-Crafted Features over Brittle Syntax Trees (N18-2)

Challenge:	argued that random splits, like standard splits lead to overly optimistic performance estimates.
Approach:	They argue that random splits, like standard splits lead to overly optimistic performance estimates.
Outcome:	The proposed method leads to more realistic performance estimates than standard splits.

Copied to clipboard

Ola Rønning, Daniel Hardt, Anders Søgaard

Attention Can Reflect Syntactic Structure (If You Let It) (2021.eacl-main)

Challenge:	sluice resolution in english is the problem of finding antecedents of wh-fronted ellipses . previous work relied on hand-crafted features over syntax trees that scale poorly to other languages and domains .
Approach:	They propose a model that uses partial parsing to find antecedents of wh-fronted ellipses in english . their model significantly outperforms previous work on available newswires .
Outcome:	The proposed model outperforms the only previous work on available newswires.

Copied to clipboard

Vinit Ravishankar, Artur Kulmizev, Mostafa Abdou, Anders Søgaard, Joakim Nivre

Defining Knowledge: Bridging Epistemology and Large Language Models (2024.emnlp-main)

Challenge:	a recent study has attempted to decode linguistic structure from the Transformer . but, much of the work focused on English, a language with rigid word order and a lack of inflectional morphology.
Approach:	They propose to fine-tune a feature encoder for BERT to learn linguistic structure from its multi-head attention mechanism.
Outcome:	The proposed model can decode full trees above baseline accuracy from single attention heads across languages.

Copied to clipboard

Constanza Fierro, Ruchira Dhar, Filippos Stamatiou, Nicolas Garneau, Anders Søgaard

Better, Faster, Stronger Sequence Tagging Constituent Parsers (N19-1)

Challenge:	Existing literature on large language models (LLMs) define knowledge as a fact if it correctly completes a cloze sentence . but the predictions of semantically equivalent clozing sentences are inconsistent .
Approach:	They review standard definitions of knowledge in epistemology and formalize interpretations applicable to LLMs.
Outcome:	The authors compare the preferences of philosophers and computer scientists in terms of knowledge definitions and evaluation protocols for testing knowledge in accordance with the most relevant definitions.

Copied to clipboard

David Vilares, Mostafa Abdou, Anders Søgaard

Adversarial Removal of Demographic Attributes Revisited (D19-1)

Challenge:	Existing efforts to speed up constituent parsing have focused on chart-based or shift-reduce parsers.
Approach:	They propose to use auxiliary losses and sentence-level fine-tuning to mitigate greedy decoding issues.
Outcome:	The proposed model surpasses the performance of sequence tagging constituent parsers on the English and Chinese Penn Treebank datasets and reduces their parsing time even further.

Copied to clipboard

Maria Barrett, Yova Kementchedjhieva, Yanai Elazar, Desmond Elliott, Anders Søgaard

Being Right for Whose Right Reasons? (2023.acl-long)

Challenge:	Several approaches have been proposed to learn classifiers that are invariant (unbiased with respect) to protected attributes.
Approach:	They propose to use a diagnostic classifier trained on a held-out subsample to find protected attributes for mention detection at above-chance levels.
Outcome:	The proposed classifier generalizes poorly to new in-domain and new domains, suggesting it relies on correlations specific to their particular data sample.

Copied to clipboard

Terne Sasha Thorn Jakobsen, Laura Cabello, Anders Søgaard

A Simple and Robust Approach to Detecting Subject-Verb Agreement Errors (N19-1)

Challenge:	Existing work has failed to acknowledge that what counts as a rationale is subjective.
Approach:	They propose to use demographic annotations to augment existing datasets to ask what demographics our models align with and whose reasoning patterns they align with.
Outcome:	The proposed model rationales align better with older and/or white annotators, and are biased towards older and white anorators.

Copied to clipboard

Simon Flachs, Ophélie Lacroix, Marek Rei, Helen Yannakoudakis, Anders Søgaard

Robustness in NLP

Challenge:	a recent study shows that neural sequential labelers overfit their training data to detect SVA errors.
Approach:	They propose a simple protocol that generates a neural sequential labeler from silver standard data and gold standard data.
Outcome:	The proposed method leads to more robust detection of SVA errors on silver standard data and gold standard data.

Is the Lottery Fair? Evaluating Winning Tickets Across Demographics (2021.findings-acl)

Copied to clipboard

Victor Petrén Bach Hansen, Anders Søgaard

The Impact of Positional Encodings on Multilingual Compression (2021.emnlp-main)

Challenge:	Recent studies suggest weight pruning compromises fairness of machine learning models . however, no empirical evaluation has been done in the context of natural language processing.
Approach:	They evaluate the fairness of lottery ticket extraction through layer-wise and global weight pruning across three languages and two tasks.
Outcome:	The proposed model is compared with two text classification datasets annotated with demographic information.

Copied to clipboard

Vinit Ravishankar, Anders Søgaard

Do Language Models Have Semantics? On the Five Standard Positions (2025.acl-long)

Challenge:	Several modifications have been proposed to improve monolingual language models, but none of them result in better multilingual models.
Approach:	They propose to add positional encodings to token embeddings to preserve word-order information in a non-autoregressive setting.
Outcome:	The proposed modifications tend to improve monolingual models, but none improve multilingual models.

Copied to clipboard

Ethical Concern Identification in NLP: A Corpus of ACL Anthology Ethics Statements (2025.naacl-long)

Challenge:	Large language models (LLMs) are trained to solve the so-called cloze task . solving clozing tasks is essentially a memorization task, says a recent study .
Approach:	They propose to use five positions to determine whether large language models exhibit semantic understanding . large language model is trained to solve the so-called cloze task .
Outcome:	The proposed theory is based on a pairwise comparison of five positions on semantic understanding in large language models and chatbots.

Copied to clipboard

Antonia Karamolegkou, Sandrine Schiller Hansen, Ariadni Christopoulou, Filippos Stamatiou, Anne Lauscher, Anders Søgaard

John praised Mary because _he_? Implicit Causality Bias and Its Interaction with Explicit Cues in LMs (2021.findings-acl)

Challenge:	a survey shows that laypeople express different ethical concerns than professionals . acl-code-ethics provides a taxonomy for ethical concerns .
Approach:	They propose to annotate a corpus of ethical concern statements from scientific papers . they extract ethical concern keywords from the statements and automate the process .
Outcome:	The proposed corpus of ethical concern statements compares with existing taxonomies and guidelines pointing to gaps and actionable insights.

Copied to clipboard

Yova Kementchedjhieva, Mark Anderson, Anders Søgaard

Some Languages Seem Easier to Parse Because Their Treebanks Leak (2020.emnlp-main)

Challenge:	Psycholinguists have identified one such cue in the implicit causality bias of interpersonal verbs.
Approach:	They propose to use pre-trained language models to encode IC bias at inference time . they hypothesize that when a cause is explicitly stated, an incongruent IC biased leads to a delay in human processing.
Outcome:	The results suggest that pre-trained language models tend to prioritize lexical patterns over higher-order signals.

Copied to clipboard

Group Fairness in Multilingual Speech Recognition Models (2024.findings-naacl)

Challenge:	Cross-language differences in (universal) dependency parsing performance are mostly attributed to treebank size, average sentence length, average dependency length, morphological complexity, and domain differences.
Approach:	They compute graph isomorphisms and find that treebank size is a factor that influences parsing performance.
Outcome:	The results show that the overlap between training and test graphs explain more of the observed variation than standard explanations such as the above.

Copied to clipboard

Anna Zee, Marc Zee, Anders Søgaard

Multimodality

Challenge:	a new study evaluates the performance disparities of ASR models across languages and demographics . a large amount of data is required to mitigate performance disparity, but this is computationally expensive .
Approach:	They evaluate the performance disparity of ASR models using a multilingual dataset . they find that model size correlates logarithmically with worst-case performance disparities .
Outcome:	The proposed models exhibit significant performance disparities across binary genders for adolescents.

Issue Framing in Online Discussion Fora (N19-1)

Copied to clipboard

Mareike Hartmann, Tallulah Jansen, Isabelle Augenstein, Anders Søgaard

Challenge:	In online discussion fora, speakers often make arguments by highlighting certain aspects of the topic.
Approach:	They propose to use a newswire and social media annotated corpus to detect issue frames in online discussions.
Outcome:	The proposed model can be applied to the domain of discussion fora using multi-task and adversarial training.

Debiasing Multilingual LLMs in Cross-lingual Latent Space (2025.emnlp-main)

Copied to clipboard

Qiwei Peng, Guimin Hu, Yekun Chai, Anders Søgaard

Robustness in NLP

Challenge:	Existing studies have evaluated their cross-lingual transferability by directly applying these methods to LLM representations, revealing their limited effectiveness across languages.
Approach:	They propose to perform debiasing in a joint latent space rather than directly on LLM representations by using an autoencoder trained on parallel TED talk scripts.
Outcome:	The proposed method improves both the overall debiasing performance and cross-lingual transferability of the proposed techniques across four languages.

A Danish FrameNet Lexicon and an Annotated Corpus Used for Training and Evaluating a Semantic Frame Classifier (L18-1)

Copied to clipboard

Bolette Pedersen, Sanni Nimb, Anders Søgaard, Mareike Hartmann, Sussi Olsen

Information Retrieval

Challenge:	a Danish FrameNet is a lexicon based on the Danish Thesaurus . it is significantly faster than building a new one from scratch .
Approach:	They propose a way to efficiently compile a Danish FrameNet based on the Danish Thesaurus . they present the corresponding corpus annotations of frames and roles and show how this can be used for a semantic frame classifier .
Outcome:	The proposed approach is faster than building a lexicon from scratch.

Are Pretrained Multilingual Models Equally Fair across Languages? (2022.coling-1)

Copied to clipboard

Laura Cabello Piqueras, Anders Søgaard

Grammatical Error Correction through Round-Trip Machine Translation (2023.findings-eacl)

Challenge:	Pretrained multilingual language models can help bridge the digital language divide, enabling high-quality NLP models for lower-resourced languages.
Approach:	They propose to use a multilingual dataset to examine whether multilingual models are equally fair across languages.
Outcome:	The proposed model enables apples-to-apples comparison across languages of group disparities in multilingual language models.

Copied to clipboard

Yova Kementchedjhieva, Anders Søgaard

Machine Translation

Challenge:	A decade ago the idea of using round-trip MT to guide grammatical error correction was not feasible due to the low quality of MT systems of the day.
Approach:	They propose to use round-trip machine translation to guide grammatical error correction to preserve meaning while mapping its surface form from one language into another.
Outcome:	The proposed system is re-examined across five languages and models of various sizes and yields consistent improvements.

Improved Multi-label Classification under Temporal Concept Drift: Rethinking Group-Robust Algorithms in a Label-Wise Setting (2022.findings-acl)

Copied to clipboard

Ilias Chalkidis, Anders Søgaard

Information Retrieval

Challenge:	Large-scale multi-label document classification presents interesting challenges due to the large label space and two-tiered skewed label distributions.
Approach:	They evaluate several group-robust optimization algorithms proposed to mitigate temporal concept drift and class imbalance in document classification.
Outcome:	The proposed algorithms outperform sampling-based approaches to class imbalance and concept drift and lead to much better performance on minority classes.

Mapping Brains with Language Models: A Survey (2023.findings-acl)

Copied to clipboard

Antonia Karamolegkou, Mostafa Abdou, Anders Søgaard

LLM Beliefs Are in Their Heads (2026.acl-long)

Challenge:	accumulated evidence for brain and language model activations remains ambiguous, but correlations with model size and quality provide grounds for cautious optimism.
Approach:	They examine the evidence accumulated by 30 studies spanning 10 datasets and 8 metrics to determine whether there is any overlap between brain and language model activations.
Outcome:	The findings suggest that representations extracted from NLP models can (partially) explain the signal found in neural data.

Copied to clipboard

Alessandro Corona Mendozza, Anders Søgaard

Challenge:	Using linear controlled probes, we investigate belief-like representations in decoder-only autoregressive LLMs using residual stream activations and single attention heads.
Approach:	They develop four different experiments on decoder-only autoregressive LLMs and examine how they fare against these standards.
Outcome:	The proposed representations exhibit strong truth sensitivity and consistent accuracy across models and data sets.

FairLex: A Multilingual Benchmark for Evaluating Fairness in Legal Text Processing (2022.acl-long)

Copied to clipboard

Ilias Chalkidis, Tommaso Pasini, Sheng Zhang, Letizia Tomada, Sebastian Schwemer, Anders Søgaard

Spurious Correlations in Cross-Topic Argument Mining (2021.starsem-1)

Challenge:	Using pre-trained language models, we evaluate performance group disparities while none of these techniques guarantee fairness, nor consistently mitigate group disparity.
Approach:	They present a benchmark suite of four datasets for evaluating the fairness of pre-trained language models and the techniques used to fine-tune them for downstream tasks.
Outcome:	The proposed methods show that performance group disparities are vibrant in many cases, while none of these techniques guarantee fairness, nor consistently mitigate group disparity.

Copied to clipboard

Terne Sasha Thorn Jakobsen, Maria Barrett, Anders Søgaard

Reasoning

Challenge:	Recent work in cross-topic argument mining attempts to learn models that generalise across topics rather than relying on within-topic spurious correlations.
Approach:	They propose to use linear approximations of decision boundaries and manual feature grouping to learn models that generalise across topics rather than relying on within-topic spurious correlations.
Outcome:	The proposed model generalise across topics rather than relying on spurious correlations.

Square One Bias in NLP: Towards a Multi-Dimensional Exploration of the Research Manifold (2022.findings-acl)

Copied to clipboard

Sebastian Ruder, Ivan Vulić, Anders Søgaard

Higher-order Comparisons of Sentence Encoder Representations (D19-1)

Challenge:	a prototypical NLP experiment trains a standard architecture on labeled English data . a recent study shows that research often goes beyond the square one setup .
Approach:	They argue that the prototypical NLP experiment trains a standard architecture on labeled English data and optimizes for accuracy without accounting for other dimensions such as fairness, interpretability, or computational efficiency.
Outcome:	The proposed model steers and biases the research dynamics in the NLP community, the authors argue . they show that the prototype biased recent NLP research on English data is true .

Copied to clipboard

Mostafa Abdou, Artur Kulmizev, Felix Hill, Daniel M. Low, Anders Søgaard

What if Othello-Playing Language Models Could See? (2025.findings-emnlp)

Challenge:	a technique developed by neuroscientists compares activity patterns of different measurement modalities . a recent study examined the correspondence between popular pretrained language encoders and human processing difficulty .
Approach:	They employ a technique to compare activity patterns of different measurement modalities . they establish a correspondence between widely-employed pretrained language encoders and human processing difficulty .
Outcome:	The proposed technique can be used to compare representational geometries of neural models . it does not require large training samples and is not prone to overfitting, authors say .

Copied to clipboard

Xinyi Chen, Yifei Yuan, Jiaang Li, Serge Belongie, Maarten de Rijke, Anders Søgaard

The Impact of Differential Privacy on Group Disparity Mitigation (2024.findings-naacl)

Challenge:	a multi-modal model trained on move sequences and board images is a popular testbed for language models .
Approach:	They propose a multi-modal model trained jointly on move sequences and board images.
Outcome:	The proposed multi-modal model trains on move sequences and board images.

Copied to clipboard

Victor Hansen, Atula Neerkaje, Ramit Sawhney, Lucie Flek, Anders Søgaard

Why is unsupervised alignment of English embeddings from different algorithms so hard? (D18-1)

Challenge:	a recent study evaluated the impact of differential privacy on fairness across four tasks.
Approach:	They evaluate the impact of differential privacy on fairness across four diverse tasks . they train (,)-differentially private models with empirical risk minimization .
Outcome:	The proposed model shows that differential privacy increases performance differences between groups . the model also reduces performance differences in the robust setting .

Copied to clipboard

Mareike Hartmann, Yova Kementchedjhieva, Anders Søgaard

Zero-Shot Sequence Labeling: Transferring Knowledge from Sentences to Tokens (N18-1)

Challenge:	a new paper challenges word embedding algorithms to align independent English word embeds with 100% precision . authors show that when two different embeddables are used, they fail to do so .
Approach:	They propose to use unsupervised bilingual dictionary induction to study English-English alignments.
Outcome:	The proposed approach is more of a challenge than a technical contribution . it shows that the results challenge unsupervised bilingual dictionary induction algorithms .

Copied to clipboard

Marek Rei, Anders Søgaard

Trick or Neat: Adversarial Ambiguity and Language Model Evaluation (2025.findings-acl)

Challenge:	Recent work has used attention weights to visualize the focus of neural models in input data.
Approach:	They propose to use attention-based visualization techniques to infer token-level labels from a network trained only on sentence-level labeling.
Outcome:	The proposed approach outperforms gradient-based methods on four datasets and is expected to outperfect supervised methods.

Copied to clipboard

Antonia Karamolegkou, Oliver Eberle, Phillip Rust, Carina Kauf, Anders Søgaard

Error Analysis and the Role of Morphology (2021.eacl-main)

Challenge:	Direct prompting fails to detect ambiguity while linear probes can decode ambiguities with high accuracy, sometimes exceeding 90%.
Approach:	They introduce an adversarial ambiguity dataset that includes syntactic, lexical, and phonological ambiguities along with adversarials.
Outcome:	The proposed dataset includes syntactic, lexical, and phonological ambiguities along with adversarial variations.

Copied to clipboard

Marcel Bollmann, Anders Søgaard

A strong baseline for question relevancy ranking (D18-1)

Challenge:	Using morphological features does improve error prediction across tasks, but is less pronounced in morphology-complex languages.
Approach:	They propose to use morphological features to improve error prediction across four different tasks and up to 57 languages to test their hypothesis.
Outcome:	The proposed model is more discriminative in morphologically simple languages than in simple ones.

Copied to clipboard

Ana Gonzalez, Isabelle Augenstein, Anders Søgaard

Challenge:	SemEval-16 and Semeval-17 community question answering shared tasks require complex pipelines and manual feature engineering to beat the IR baseline.
Approach:	They train a multi-task feed forward network on a bag of 14 distance measures for the input question pair and train it using language-independent features.
Outcome:	The proposed model outperforms the best shared task systems on the task of retrieving relevant previously asked questions.

Neural Speed Reading Audited (2020.findings-emnlp)

Copied to clipboard

Information Extraction & Text Mining

Challenge:	Several approaches to neural speed reading have been presented at major NLP and machine learning conferences in 2017–20.
Approach:	They propose to model "human speed reading" for more efficient NLP, including document classification and named entity recognition.
Outcome:	The proposed approach has 7% error reduction and 136x speed-up over the state-of-the-art in neural speed reading.

Model-based Annotation of Coreference (2020.lrec-1)

Copied to clipboard

Rahul Aralikatte, Anders Søgaard

Information Extraction & Text Mining

Challenge:	Annotators are asked to annotate coreferent spans of text, which is unnatural . we present an alternative in which annotators can preprocess documents and assign pronouns to entities.
Approach:	They propose an alternative in which annotators are asked to assign pronouns to entities and preprocess documents to create a knowledge base.
Outcome:	The proposed model-based approach leads to faster annotation and higher inter-annotator agreement and opens up an alternative approach to coreference resolution.

Rewarding Coreference Resolvers for Being Consistent with World Knowledge (D19-1)

Copied to clipboard

Rahul Aralikatte, Heather Lent, Ana Valeria Gonzalez, Daniel Hershcovich, Chen Qiu, Anders Sandholm, Michael Ringaard, Anders Søgaard

Information Extraction & Text Mining

Challenge:	Unresolved coreference is a bottleneck for relation extraction systems . a state-of-the-art system may be able to infer the relation using distributional information about the phrase the Sunshine State, but is likely to have limited evidence for the decision that it is coreferential with Florida rather than with Skynyrd.
Approach:	They propose to forward coreference input to relation extraction system and reward them for producing triples that are found in knowledge bases.
Outcome:	The proposed approach improves over the state-of-the-art by forwarding their input to a relation extraction system and rewarding resolvers for producing triples that are found in knowledge bases.

On the Limitations of Unsupervised Bilingual Dictionary Induction (P18-1)

Copied to clipboard

Anders Søgaard, Sebastian Ruder, Ivan Vulić

Word Order Does Matter and Shuffled Language Models Know It (2022.acl-long)

Challenge:	Unsupervised machine translation does not require cross-lingual supervision, whether a dictionary, translations, or comparable corpora.
Approach:	They propose an adversarial, unsupervised cross-lingual word embedding technique for bilingual dictionary induction that exploits a weak supervision signal from identical words.
Outcome:	The proposed model relies heavily on an adversarial, unsupervised cross-lingual word embedding technique for bilingual dictionary induction.

Copied to clipboard

Mostafa Abdou, Vinit Ravishankar, Artur Kulmizev, Anders Søgaard

Do Transformer Models Show Similar Attention Patterns to Task-Specific Human Gaze? (2022.acl-long)

Challenge:	Recent studies have shown that language models pretrained and/or fine-tuned on randomly permuted sentences exhibit competitive performance on GLUE, putting into question the importance of word order information.
Approach:	They propose a transformer-based BERT architecture that uses a fixed, sinusoidal position embedding added to each token embeddable to compensate for this absence of linear order.
Outcome:	The proposed model retains word order information because of the dependencies between sentence length and unigram probabilities.

Copied to clipboard

Oliver Eberle, Stephanie Brandl, Jonas Pilot, Anders Søgaard

What does the Failure to Reason with “Respectively” in Zero/Few-Shot Settings Tell Us about Language Models? (2023.acl-long)

Challenge:	We compare attention functions in pre-trained language models to human eye fixation patterns during task-specific reading tasks.
Approach:	They compare attention functions in large-scale pre-trained language models to classical cognitive models of human attention by using a dataset with eye-tracking recordings of native speakers of English.
Outcome:	The proposed model is as predictive of human eye fixation patterns as classical cognitive models of human attention.

Copied to clipboard

Ruixiang Cui, Seolhwa Lee, Daniel Hershcovich, Anders Søgaard

MuLan: A Study of Fact Mutability in Language Models (2024.naacl-short)

Challenge:	In the context of natural language inference, we examine how language models reason with respective readings from two perspectives: syntactic-semantic and commonsense-world knowledge.
Approach:	They propose a controlled synthetic dataset WikiResNLI and a naturally occurring dataset NatResLI to encompass various explicit and implicit realizations of "respectively".
Outcome:	The proposed datasets include explicit and implicit readings of "respectively" the proposed dataset shows that fine-tuned models struggle with understanding readings without explicit supervision.

Copied to clipboard

Constanza Fierro, Nicolas Garneau, Emanuele Bugliarello, Yova Kementchedjhieva, Anders Søgaard

DaNE: A Named Entity Resource for Danish (2020.lrec-1)

Challenge:	Pretrained and large language models encode factual knowledge, but factual information changes over time and mutates with the passage of time.
Approach:	They propose to use a model to evaluate the ability of English language models to anticipate time-contingency by comparing their models to a benchmark model.
Outcome:	The proposed model can predict the president of a country or the winner of sa championship in time, but it is difficult to update them due to their mutability.

Copied to clipboard

Rasmus Hvingelby, Amalie Brogaard Pauli, Maria Barrett, Christina Rosted, Lasse Malm Lidegaard, Anders Søgaard

On the Interaction of Belief Bias and Explanations (2021.findings-acl)

Challenge:	a named entity annotation for the Danish Universal Dependencies treebank is the largest publicly available named entity gold annotation.
Approach:	They propose a named entity annotation for the Danish Universal Dependencies treebank using the CoNLL-2003 annotation scheme DaNE.
Outcome:	The proposed annotations improve Danish named entity recognition over a recent cross-lingual approach and over norwegian training set.

Copied to clipboard

Ana Valeria González, Anna Rogers, Anders Søgaard

Multi-Task Semantic Dependency Parsing with Policy Gradient for Learning Easy-First Strategies (P19-1)

Challenge:	Existing methods to evaluate explainability fail to account for belief biases affecting human performance . previous studies have shown that neural models can make confident predictions relying on artifacts .
Approach:	They propose to account for belief bias in explainability by using models of varying quality and adversarial examples.
Outcome:	The proposed methods show that results change when using models of varying quality and adversarial examples.

Copied to clipboard

Shuhei Kurita, Anders Søgaard

On Mitigating Performance Disparities in Multilingual Speech Recognition (2024.emnlp-main)

Challenge:	Existing dependency parsing algorithms do not support directed acyclic graphs . a a systole-based dependency parses sentences using binary semantic relations that are not trees .
Approach:	They propose an iterative predicate selection algorithm for semantic dependency parsing . they train the algorithm using multi-task learning and task-specific policy gradient training .
Outcome:	The proposed model achieves a new state of the art on the SemEval 2015 task 18 dataset .

Copied to clipboard

Monorama Swain, Anna Zee, Anders Søgaard

Parameter sharing between dependency parsers for related languages (D18-1)

Challenge:	Automatic Speech Recognition systems are not always equally effective for all users, and gender disparity in their performance is a significant concern.
Approach:	They compare performance of different fine-tuning algorithms for multilingual speech recognition across languages and genders.
Outcome:	The proposed algorithms improve performance and parity across languages and languages.

Copied to clipboard

Miryam de Lhoneux, Johannes Bjerva, Isabelle Augenstein, Anders Søgaard

Information Retrieval

Challenge:	Using parameter sharing between parsers of related languages can improve performance, but there is no consensus on what parameters to share.
Approach:	They propose a model where transition classifier parameters are shared and word and character parameters are controlled by a parameter that can be tuned on validation data.
Outcome:	The proposed model improves on a monolingually trained baseline.

Zero-Shot Dependency Parsing with Worst-Case Aware Automated Curriculum Learning (2022.acl-short)

Copied to clipboard

Miryam de Lhoneux, Sheng Zhang, Anders Søgaard

Understanding Subword Compositionality of Large Language Models (2025.emnlp-main)

Challenge:	Large multilingual pretrained language models such as mBERT and XLM-RoBERTa have been found to be effective for cross-lingual transfer of syntactic parsing models but only between related languages.
Approach:	They propose to use multi-task learning to dynamically optimize for parsing performance on outlier languages by using a multi-level learning approach.
Outcome:	The proposed method significantly outperforms uniform and size-proportional sampling in the zero-shot setting.

Copied to clipboard

Qiwei Peng, Yekun Chai, Anders Søgaard

A Discriminative Latent-Variable Model for Bilingual Lexicon Induction (D18-1)

Challenge:	Large language models (LLMs) take sequences of subwords as input, requiring them to compose subword representations into meaningful word-level representations.
Approach:	They propose to probe how large language models compose subword information . they find structural similarity, semantic decomposability, and form retention are key aspects .
Outcome:	The proposed models can be classified into three distinct groups, the authors show . they show that they can achieve great performance when probing layer by layer their sensitivity to semantic decompositionality .

Copied to clipboard

Sebastian Ruder, Ryan Cotterell, Yova Kementchedjhieva, Anders Søgaard

Type B Reflexivization as an Unambiguous Testbed for Multilingual Multi-Task Gender Bias (2020.emnlp-main)

Challenge:	Existing methods for bilingual lexicon induction take advantage of word embeddings, but our model is not as efficient as previous work.
Approach:	They propose a discriminative latent-variable model for bilingual lexicon induction that combines the bipartite matching dictionary prior and an embedding-based approach.
Outcome:	The proposed model outperforms existing models on six language pairs and shows that it mitigates hubness problem.

Copied to clipboard

Ana Valeria González, Maria Barrett, Rasmus Hvingelby, Kellie Webster, Anders Søgaard

X-WikiRE: A Large, Multilingual Resource for Relation Extraction as Machine Comprehension (D19-61)

Challenge:	English challenge datasets highlight gender-ambiguous occurrences of ‘doctor’ as male doctors, but they are not useful for other languages.
Approach:	They propose to build multi-task challenge datasets for detecting gender bias that lead to unambiguously wrong model predictions for languages with type B reflexivization.
Outcome:	The proposed dataset can detect gender bias in languages with type B reflexivization and spans four languages and four NLP tasks.

Copied to clipboard

Mostafa Abdou, Cezar Sas, Rahul Aralikatte, Isabelle Augenstein, Anders Søgaard

Ellipsis Resolution as Question Answering: An Evaluation (2021.eacl-main)

Challenge:	Existing knowledge bases are heavily biased towards English, but Wikipedias cover very different topics in different languages.
Approach:	They propose a multilingual dataset that frams relation extraction as a machine reading problem.
Outcome:	The proposed model can be used to transfer models cross-lingually and improves knowledge base completion across languages.

Copied to clipboard

Rahul Aralikatte, Matthew Lamm, Daniel Hardt, Anders Søgaard

Natural Language Interfaces

Challenge:	Existing models for ellipsis resolution in English are expensive and cumbersome . ellipas are hard, open problems in NLP, and can cause errors in translation, question answering, and dialogue understanding.
Approach:	They propose an alternative approach to ellipsis resolution based on question answering architectures.
Outcome:	The proposed model outperforms the current state of the art for ellipsis resolution in English . it shows that annotations can be useful for a subset of the known ellipas .

Unlocking Markets: A Multilingual Benchmark to Cross-Market Question Answering (2024.emnlp-main)

Copied to clipboard

Yifei Yuan, Yang Deng, Anders Søgaard, Mohammad Aliannejadi

Natural Language Interfaces

Challenge:	Product-related question answering (PQA) involves utilizing product-related resources to provide precise answers to users.
Approach:	They propose a task of multilingual cross-market product-based question answering that combines product-related questions with product-specific questions from a multilingual marketplace.
Outcome:	The proposed task provides answers to product-related questions in a multilingual marketplace even in fewer languages.

What a Creole Wants, What a Creole Needs (2022.lrec-1)

Copied to clipboard

Heather Lent, Kelechi Ogueji, Miryam de Lhoneux, Orevaoghene Ahia, Anders Søgaard

Low-Resource NLP

Challenge:	Recent efforts to improve the quality of high-resource languages focus on translating existing datasets into other languages, but this approach ignores that different language communities have different needs.
Approach:	They examine how things needed from language technology can change dramatically from one language to another.
Outcome:	The proposed method ignores that different language communities have different needs.

The Sensitivity of Language Models and Humans to Winograd Schema Perturbations (2020.acl-main)

Copied to clipboard

Mostafa Abdou, Vinit Ravishankar, Maria Barrett, Yonatan Belinkov, Desmond Elliott, Anders Søgaard

Few-Shot and Zero-Shot Learning for Historical Text Normalization (D19-61)

Challenge:	Large-scale pre-trained language models are driving recent improvements in perfromance on the Winograd Schema Challenge . a diagnostic dataset shows that these models are sensitive to linguistic perturbations that minimally affect human understanding .
Approach:	They propose to use a dataset to test pre-trained language models for the Winograd Schema Challenge . they show that these models are sensitive to linguistic perturbations that minimally affect human understanding .
Outcome:	The proposed models are sensitive to linguistic perturbations that minimally affect human understanding.

Copied to clipboard

Marcel Bollmann, Natalia Korchagina, Anders Søgaard

A Two-Sided Discussion of Preregistration of NLP Research (2023.eacl-main)

Challenge:	Historical text normalization often relies on small training datasets.
Approach:	They evaluate 63 multi-task learning configurations for sequence-to-sequence-based historical text normalization across ten datasets from eight languages.
Outcome:	The proposed learning architecture outperforms the simple, but strong identity baseline.

Copied to clipboard

Anders Søgaard, Daniel Hershcovich, Miryam de Lhoneux

Ethical NLP

Challenge:	et al. (2021) suggest NLP research should adopt preregistration to prevent fishing expeditions and promote publication of negative results.
Approach:	et al. suggest NLP research should adopt preregistration to prevent fishing expeditions and promote publication of negative results.
Outcome:	The proposed approach solves many methodological problems with NLP research.

Factual Consistency of Multilingual Pretrained Language Models (2022.findings-acl)

Copied to clipboard

Constanza Fierro, Anders Søgaard

Concept Space Alignment in Multilingual LLMs (2024.emnlp-main)

Challenge:	Recent work shows that monolingual English language models fill-in-the-blank differently for paraphrases describing the same fact.
Approach:	They propose a resource to analyze consistency of English language models . they find that mBERT is as inconsistent as English BERT in paraphrases .
Outcome:	The proposed model is as inconsistent as English BERT in English paraphrases, but it is more so for all the other 45 languages.

Copied to clipboard

Qiwei Peng, Anders Søgaard

FoodieQA: A Multimodal Dataset for Fine-Grained Understanding of Chinese Food Culture (2024.emnlp-main)

Challenge:	Multilingual large language models generalize somewhat across languages, but it is unclear whether this is a result of improved, implicit alignment, or of something else, e.g., linguistic overlap or semi-parallel subsets of training data.
Approach:	They hypothesize that implicit alignment is the reason for generalization in multilingual large language models.
Outcome:	The proposed model generalizes well across languages, but lacks linearity.

Copied to clipboard

Wenyan Li, Crystina Zhang, Jiaang Li, Qiwei Peng, Raphael Tang, Li Zhou, Weijia Zhang, Guimin Hu, Yifei Yuan, Anders Søgaard, Daniel Hershcovich, Desmond Elliott

Multimodality

Challenge:	FoodieQA is a manually curated, fine-grained image-text dataset capturing the intricate features of food cultures across various regions in China.
Approach:	They evaluate vision–language Models and large language models on unseen food images and corresponding questions.
Outcome:	The proposed dataset evaluates vision–language Models and large language models on unseen food images and corresponding questions.

Unsupervised Induction of Linguistic Categories with Records of Reading, Speaking, and Writing (N18-1)

Copied to clipboard

Maria Barrett, Ana Valeria González-Garduño, Lea Frermann, Anders Søgaard

Should We Ban English NLP for a Year? (2022.emnlp-main)

Challenge:	a few researchers have shown that data traces from human processing can be used to improve NLP models.
Approach:	They propose to use data readily available for most languages to improve unsupervised induction . they find that english unsupervised POS induction achieves an error reduction of 1.5% .
Outcome:	The proposed model improves on Ontonotes domains with a word embeddings.

Copied to clipboard

Lexi: A tool for adaptive, personalized text simplification (C18-1)

Challenge:	aaron carroll: two thirds of NLP research is devoted to developing technology for speakers of English . carroll says this bias feeds into consumer technologies to widen existing inequality gaps . he says we need to consider more concrete measures to mitigate climate change .
Approach:	a new paper argues that NLP is contributing to global inequalities through a digital language divide . a carbon tax, cap-and-trade and car-free Sundays are examples of measures to mitigate climate change .
Outcome:	a new paper argues that NLP is contributing to global inequalities through a digital language divide . a carbon tax, cap-and-trade and car-free Sundays are examples of measures to mitigate climate change .

Copied to clipboard

Joachim Bingel, Gustavo Paetzold, Anders Søgaard

Paraphrasing

Challenge:	Existing research on text simplification has aimed to develop generic solutions . instead, we need to develop customized simplification systems for individual users .
Approach:	They propose a framework for adaptive lexical simplification and introduce Lexi, a free open-source tool for personalized text simplification.
Outcome:	The proposed framework is based on a free open-source tool for adaptive, personalized text simplification.

Dynamic Forecasting of Conversation Derailment (2021.emnlp-main)

Copied to clipboard

Yova Kementchedjhieva, Anders Søgaard

Dialogue Systems & Conversational Agents

Challenge:	a pretrained language encoder can predict derailment in online conversations . this is a useful task for detecting and preventing abusive language .
Approach:	They extend a task to predict derailment in online conversations by using a pretrained language encoder.
Outcome:	The proposed task outperforms previous approaches in terms of performance and quality.

Replicating and Extending “Because Their Treebanks Leak”: Graph Isomorphism, Covariants, and Parser Performance (2021.acl-short)

Copied to clipboard

Mark Anderson, Anders Søgaard, Carlos Gómez-Rodríguez

Challenges and Strategies in Cross-Cultural NLP (2022.acl-long)

Challenge:	a small sample size and unreliable results suggest a correlation between parser performance and graph isomorphism is not observed in the wild.
Approach:	They propose to replicate a study which found graph isomorphism is a non-trivial variable . they also bin sentences by length and find correlation between parser performance and isopathism disappears .
Outcome:	The results show that the original analysis was unreliable and had methodological issues . the study also bin sentences by length and shows that the correlation between parser performance and graph isomorphism disappears when controlling for covariants.

Copied to clipboard

Daniel Hershcovich, Stella Frank, Heather Lent, Miryam de Lhoneux, Mostafa Abdou, Stephanie Brandl, Emanuele Bugliarello, Laura Cabello Piqueras, Ilias Chalkidis, Ruixiang Cui, Constanza Fierro, Katerina Margatina, Phillip Rust, Anders Søgaard

The Effect of Round-Trip Translation on Fairness in Sentiment Analysis (2021.emnlp-main)

Challenge:	Various efforts have been made to accommodate linguistic diversity and serve speakers of many different languages.
Approach:	They propose a framework to examine cultural differences in NLP to better serve users . they argue that cultural knowledge, preferences and values can affect NLP practices .
Outcome:	The proposed framework examines how cultural knowledge, preferences and values can affect NLP practices.

Copied to clipboard

Jonathan Gabel Christiansen, Mathias Gammelgaard, Anders Søgaard

Text Classification

Challenge:	Sentiment analysis systems exhibit sensitivity to protected attributes, while round-trip translation has been shown to normalize text.
Approach:	They propose to use round-trip translation to normalize text to reduce the fairness gap between groups in sentiment analysis.
Outcome:	The proposed method reduces the fairness gap between groups by up to 47%.

FrameNet-Cultures: A Benchmark for Evaluating LLMs via Cross-Cultural Frame Semantics (2026.findings-acl)

Copied to clipboard

Neda Jamshidi, Anders Søgaard, Monica Bianchini

Sociolectal Analysis of Pretrained Language Models (2021.emnlp-main)

Challenge:	Existing evaluation paradigms for large language models lack rigorous methods to evaluate cultural alignment . FRAMENET-CULTURES is an open-ended benchmark for evaluating cultural alignment in LLMs .
Approach:	They propose a benchmark for evaluating cultural alignment in large language models based on Fillmore-style frame semantics.
Outcome:	The proposed benchmark is based on Fillmore-style frame semantics.

Copied to clipboard

Sheng Zhang, Xin Zhang, Weiming Zhang, Anders Søgaard

Multi-Task Learning of Pairwise Sequence Classification Tasks over Disparate Label Spaces (N18-1)

Challenge:	Using data from English cloze tests, we demonstrate wide performance gaps across demographic groups and show that pretrained language models disfavor young non-white male speakers.
Approach:	They use data from English cloze tests to examine performance differences of pretrained language models across demographic groups.
Outcome:	The models disfavor young non-white male speakers, but larger models reduce performance gaps between majority and minority groups.

Copied to clipboard

Isabelle Augenstein, Sebastian Ruder, Anders Søgaard

NLP for Social Good: A Survey and Outlook of Challenges, Opportunities and Responsible Deployment (2026.eacl-long)

Challenge:	Multi-task learning and semi-supervised learning are successful paradigms for learning in scenarios with limited labelled data.
Approach:	They propose to induce a joint embedding space between disparate label spaces and learning transfer functions between label embeddments to leverage unlabelled data and auxiliary, annotated datasets.
Outcome:	The proposed approach outperforms strong single and multi-task baselines and achieves state of the art on aspect-based and topic-based sentiment analysis.

Copied to clipboard

Antonia Karamolegkou, Angana Borah, Eunjung Cho, Sagnik Ray Choudhury, Martina Galletti, Pranav Gupta, Oana Ignat, Priyanka Kargupta, Neema Kotonya, Hemank Lamba, Sun-Joo Lee, Arushi Mangla, Ishani Mondal, Fatima Zahra Moudakir, Deniz Nazar, Poli Nemkova, Dina Pisarevskaya, Naquee Rizwan, Nazanin Sabri, Keenan Samway, Dominik Stammbach, Anna Steinberg Schulten, David Tomás, Steven R Wilson, Bowen Yi, Jessica H Zhu, Arkaitz Zubiaga, Anders Søgaard, Alexander Fraser, Zhijing Jin, Rada Mihalcea, Joel R. Tetreault, Daryna Dementieva

QLEVR: A Diagnostic Dataset for Quantificational Language and Elementary Visual Reasoning (2022.findings-naacl)

Challenge:	This paper surveys work in "NLP for Social Good" across nine domains relevant to global development and risk agendas.
Approach:	This paper analyzes work in "NLP for Social Good" across nine domains relevant to global development and risk agendas.
Outcome:	The paper analyzes work in "NLP for Social Good" across nine domains relevant to global development and risk agendas.

Copied to clipboard

Zechen Li, Anders Søgaard

Visual Data in NLP

Challenge:	Synthetic datasets have been used to test visual question-answering datasets for reasoning abilities.
Approach:	They propose a visual question-answering dataset that is minimally biased and diagnostic . they propose to use the dataset to test visual reasoning abilities .
Outcome:	The proposed dataset is compared with existing models and shows it is far superior to existing models.

WikiBank: Using Wikidata to Improve Multilingual Frame-Semantic Parsing (2020.lrec-1)

Copied to clipboard

Cezar Sas, Meriem Beloucif, Anders Søgaard

Evaluating Multimodal Language Models as Visual Assistants for Visually Impaired Users (2025.acl-long)

Challenge:	Frame-semantic annotations exist for a tiny fraction of the world’s languages, however, Wikidata provides a common, distant supervision signal for semantic parsers.
Approach:	They propose a multilingual resource with partial semantic dependency structures that can be used to extend pre-existing resources rather than creating new man-made resources from scratch.
Outcome:	The proposed resource can be used to augment pre-existing resources or reduce the annotation effort for low-resource languages.

Copied to clipboard

Antonia Karamolegkou, Malvina Nikandrou, Georgios Pantazopoulos, Danae Sanchez Villegas, Phillip Rust, Ruchira Dhar, Daniel Hershcovich, Anders Søgaard

Multimodality

Challenge:	Despite high adoption rate of Large Language Models, there are limitations related to contextual understanding, cultural sensitivity, and complex scene understanding.
Approach:	They conduct a user survey to identify adoption patterns and key challenges users face with such technologies.
Outcome:	The proposed models have high adoption rates but still face limitations in visual aids.

Evaluating Webcam-based Gaze Data as an Alternative for Human Rationale Annotations (2024.lrec-main)

Copied to clipboard

Stephanie Brandl, Oliver Eberle, Tiago Ribeiro, Anders Søgaard, Nora Hollenstein

Explainability & Interpretability in NLP

Challenge:	We compare webcam-based eye-tracking recordings with human-annotated rationales to evaluate importance scores.
Approach:	They compare webcam-based eye-tracking recordings with attention-based importance scores for 4 different multilingual Transformer-based language models.
Outcome:	The proposed method is comparable to human rationales in linguistic analysis.

How Conservative are Language Models? Adapting to the Introduction of Gender-Neutral Pronouns (2022.naacl-main)

Copied to clipboard

Stephanie Brandl, Ruixiang Cui, Anders Søgaard

Lost in Embeddings: Information Loss in Vision–Language Models (2025.findings-emnlp)

Challenge:	a recent study shows that gender-neutral pronouns are not associated with processing difficulties . linguistic scholars have observed how technology has altered the course of language evolution .
Approach:	They show that gender-neutral pronouns in Danish, English and Swedish are not associated with processing difficulties.
Outcome:	a new study shows that gender-neutral pronouns are not associated with human processing difficulties . the findings suggest that such conservativity in language models may limit widespread adoption .

Copied to clipboard

Wenyan Li, Raphael Tang, Chengzu Li, Caiqi Zhang, Ivan Vulić, Anders Søgaard

Lost in Evaluation: Misleading Benchmarks for Bilingual Dictionary Induction (D19-1)

Challenge:	Experiments reveal connectors substantially distort the local geometry of visual representations, with k-nearest neighbors diverging by 40–60% post-projection, correlating with degradation in retrieval performance.
Approach:	They propose two approaches to examine and quantify information loss by analyzing latent representation space.
Outcome:	The proposed model improves retrieval performance by analyzing changes in k-nearest neighbor relationships between image representations before and after projection.

Copied to clipboard

Yova Kementchedjhieva, Mareike Hartmann, Anders Søgaard

Minimax and Neyman–Pearson Meta-Learning for Outlier Languages (2021.findings-acl)

Challenge:	a quarter of the data consists of proper nouns, which can be hardly indicative of BDI performance, and there are pervasive gaps in the gold-standard targets.
Approach:	They examine the composition and quality of test sets for five different languages . they suggest future research avoids drawing conclusions from quantitative results .
Outcome:	The results show that a quarter of the data consists of proper nouns, which can be hardly indicative of BDI performance, and there are gaps in the gold-standard targets.

Copied to clipboard

Edoardo Maria Ponti, Rahul Aralikatte, Disha Shrivastava, Siva Reddy, Anders Søgaard

Robustness in NLP

Challenge:	Model-agnostic meta-learning (MAML) is a strategy to learn resource-poor languages in a sample-efficient fashion.
Approach:	They propose a model-agnostic meta-learning strategy that minimizes the expected risk across languages with a uniform prior . they propose 'minimax' and 'neyman-pearson' models that constrain the risk in each language to a maximum threshold.
Outcome:	The proposed model reduces the maximum risk across languages while constraining the risk in each language to a maximum threshold.

Locke’s Holiday: Belief Bias in Machine Reading (2021.emnlp-main)

Copied to clipboard

Historical Text Normalization with Delayed Rewards (P19-1)

Challenge:	Existing machine reading systems fail to answer What did Elizabeth want? correctly in the context of ‘My kingdom for a cough drop, cried Queen Elizabeth.’ Biased by co-occurrence statistics in the training data of pretrained language models, systems predict my kingdom, rather than a lung drop.
Approach:	They propose to use a dataset to quantify belief bias in machine reading based on pre-trained language models to examine the pervasiveness of belief bias.
Outcome:	The proposed dataset shows that machine reading models fail when contexts do not align with common beliefs.

Copied to clipboard

Simon Flachs, Marcel Bollmann, Anders Søgaard

LeXFiles and LegalLAMA: Facilitating English Multinational Legal Language Model Development (2023.acl-long)

Challenge:	Recent work on a novel approach to historical text normalization has shown that policy gradient fine-tuning improves accuracy across languages.
Approach:	They propose to train sequence-to-sequence models with simple token-level log-likelihood with reinforcement learning to optimize for exact matches.
Outcome:	The proposed model outperforms phrase-based models in the evaluation metric for historical text normalization across languages.

Copied to clipboard

Ilias Chalkidis, Nicolas Garneau, Catalina Goanta, Daniel Katz, Anders Søgaard

Generalized Quantifiers as a Source of Error in Multilingual NLU Benchmarks (2022.naacl-main)

Challenge:	In this study, we examine the performance of legal-oriented pre-trained language models.
Approach:	They conduct a detailed analysis on the performance of legal-oriented pre-trained language models by examining their original objective, acquired knowledge, and legal language understanding capacities.
Outcome:	The results show that the models' size and pre-training corpora are important for the development of domain-specific models.

Copied to clipboard

Ruixiang Cui, Daniel Hershcovich, Anders Søgaard

Are All Good Word Vector Spaces Isomorphic? (2020.emnlp-main)

Challenge:	Quantifiers are pervasive in NLU benchmarks and their occurrence at test time is associated with performance drops.
Approach:	They propose a generalized quantifier NLI task to quantify their contribution to the errors of NLU models.
Outcome:	The proposed model is based on a generalized quantifier theory and is compared with pre-trained models.

Copied to clipboard

Ivan Vulić, Sebastian Ruder, Anders Søgaard