Papers by Dieuwke Hupkes
Compute Optimal Scaling of Skills: Knowledge vs Reasoning (2025.findings-acl)
Copied to clipboard
| Challenge: | Scaling laws are a critical component of the LLM development pipeline, but little is known about whether the COs of individual skills such as mathematical reasoning, question answering (QA) or coding, align with these APEs. |
| Approach: | They examine knowledge-based QA and code generation to find out whether skill-dependent scaling is an artefact of the pretraining datamix. |
| Outcome: | The proposed scaling laws are skill-dependent, and knowledge and code exhibit fundamental differences in scaling behaviour when corrected for datamix differences. |
Lost in Inference: Rediscovering the Role of Natural Language Inference for Large Language Models (2025.naacl-long)
Copied to clipboard
| Challenge: | In the recent past, a popular way of evaluating natural language understanding was to consider a model’s ability to perform natural language inference (NLI) tasks. |
| Approach: | They focus on five different NLI benchmarks across six models of different scales and examine how their accuracies develop during training. |
| Outcome: | The softmax distributions of models align with human label distributions in cases where statements are ambiguous or vague. |
Masked Language Modeling and the Distributional Hypothesis: Order Word Matters Pre-training for Little (2021.emnlp-main)
Copied to clipboard
| Challenge: | masked language models (MLMs) pre-train to model higher-order word co-occurrence statistics . authors suggest that such models have learned to represent syntactic structures prevalent in classical NLP pipelines . purely distributional information largely explains the success of pre-training, authors say . |
| Approach: | They propose to pre-train masked language models on sentences with random shuffled word order and show they still achieve high accuracy after fine-tuning on many downstream tasks. |
| Outcome: | The proposed model performs well according to parametric syntactic probes . the authors argue that the model is not all that different from earlier distributional models . |
Location Attention for Extrapolation to Longer Sequences (2020.acl-main)
Copied to clipboard
| Challenge: | Neural networks are surprisingly good at interpolating, but they are often unable to extrapolate patterns beyond the seen data. |
| Approach: | They propose to use a special type of extrapolation for natural language processing to generalize to sequences that are longer than the training ones. |
| Outcome: | The proposed model is more likely to extrapolate than models with common attention mechanisms. |
Can Transformers Process Recursive Nested Constructions, Like Humans? (2022.coling-1)
Copied to clipboard
| Challenge: | A recent study evaluated recursive processing in recurrent neural language models (RNN-LMs) and showed that such models perform below chance level on embedded dependencies within nested constructions. |
| Approach: | They evaluated recursive processing in recurrent neural language models and found that Transformers perform below chance level on embedded dependencies within nested constructions. |
| Outcome: | The proposed models perform below chance level on embedded dependencies within nested constructions, compared to humans. |
Internal and external pressures on language emergence: least effort, object constancy and frequency (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Existing studies show that the emergent languages rarely display salient features inherent to natural languages, such as compositionality of meaning and generalisation to novel objects. |
| Approach: | They propose to formalise the principle of least effort through an auxiliary objective and explore several game variants inspired by the principle 'object constancy' they find that the proposed sources of pressure result in emerging languages with less redundancy, more focus on high-level conceptual information, and better abilities of generalisation. |
| Outcome: | The proposed sources of pressure result in emerging languages with less redundancy, more focus on high-level conceptual information, and better abilities of generalisation. |
Memorisation Cartography: Mapping out the Memorisation-Generalisation Continuum in Neural Machine Translation (2023.emnlp-main)
Copied to clipboard
| Challenge: | Using the counterfactual memorisation metric, we find that when training neural networks, models will memorise some inputs but not others. |
| Approach: | They use the counterfactual memorisation metric to build a resource that places 5M NMT datapoints on a memorisations-generalisation map and describe how the datapoint’s surface-level characteristics and a models’ per-datum training signals are predictive of memorising in NMT. |
| Outcome: | The proposed model places 5M NMT datapoints on a memorisation-generalisation map and shows how their surface-level characteristics and models’ per-datum training signals are predictive of memorising in NMT. |
Language Models Use Monotonicity to Assess NPI Licensing (2021.findings-acl)
Copied to clipboard
| Challenge: | Neural language models (LMs) have become powerful approximators of human language . fewer studies have been done on what kind of formal semantic features are encoded by LMs . |
| Approach: | They propose a series of experiments that investigate the semantic knowledge of language models . they use diagnostic classifiers, linguistic acceptability tasks and a ranking method to investigate the models' inner workings. |
| Outcome: | The proposed method can be applied to LMs trained on filtered corpora and gain stronger insights into their generalizations. |
Language Modelling as a Multi-Task Problem (2021.eacl-main)
Copied to clipboard
| Challenge: | Using multitask learning, humans are optimising their behaviour towards a multitude of objectives to reach their goals in dayto-day life. |
| Approach: | They propose to study language modelling as a multi-task problem by examining the generalisation behaviour of language models as they learn the linguistic concept of Negative Polarity Items. |
| Outcome: | The proposed model is able to learn the linguistic concept of Negative Polarity Items (NPIs) and is a multi-task learning model. |
The Curious Case of Absolute Position Embeddings (2022.findings-emnlp)
Copied to clipboard
| Challenge: | In natural language, it is not absolute position that matters, but relative position . et al., 2017) language models incorporate positional encodings that encode absolute (linear) word order. |
| Approach: | They find that Transformer language models encode word order using positional information . they also find that models that use absolute position embeddings over-rely on positional data . |
| Outcome: | The results raise questions about the efficacy of APEs to model the relativity of position information. |
Interpretability of Language Models via Task Spaces (2024.acl-long)
Copied to clipboard
| Challenge: | linguistic interpretability is a method used to assess language models' ability to interpret outputs. |
| Approach: | They propose a method to assess LMs' language conceptualisations by 'similarity probing' and a technique to fine tune them via gradient differentials to disentangle the learning signals of linguistic phenomena. |
| Outcome: | The proposed method generalises larger models to overarching general concepts for linguistic tasks, and the generalisation patterns are stable throughout training and not marked by incisive stages. |
Co-evolution of language and agents in referential games (2021.eacl-main)
Copied to clipboard
| Challenge: | Referential games allow neural agents to learn language, but they do not take into account the learning biases of the learners. |
| Approach: | They propose to model cultural and architectural evolution in a population of agents to take into account learning biases of the language learners and let them co-evolve. |
| Outcome: | The proposed model outperforms cultural transmission in a population of agents and takes into account learning biases of the learners. |
The Grammar of Emergent Languages (2020.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies on emergent languages focus on semantics, but lack tools to analyse their properties. |
| Approach: | They propose to use unsupervised grammar induction techniques to analyse emergent languages and to examine their syntactic properties. |
| Outcome: | The proposed techniques are appropriate to analyse emergent languages and show that they exhibit syntactic properties similar to those observed in human language. |
The emergence of number and syntax units in LSTM language models (N19-1)
Copied to clipboard
| Challenge: | a recent study shows that LSTMs can capture syntax-sensitive generalizations such as long-distance number agreement. |
| Approach: | They investigate the inner mechanics of number tracking in LSTMs at the single neuron level . they find that long-distance number information is largely managed by two "number units" importantly, the behaviour of these units is partially controlled by other units to track syntactic structure . |
| Outcome: | The proposed model is based on a language model with a long-distance number agreement task. |
Text Characterization Toolkit (TCT) (2022.aacl-demo)
Copied to clipboard
Daniel Simig, Tianlu Wang, Verna Dankers, Peter Henderson, Khuyagbaatar Batsuren, Dieuwke Hupkes, Mona Diab
| Challenge: | Text Characterization Toolkit (TCT) is a tool that researchers can use to study characteristics of large datasets. |
| Approach: | They propose a text characterization toolkit that researchers can use to study characteristics of large datasets. |
| Outcome: | The proposed tool can be used to study characteristics of large datasets and to understand the influence of attributes on models’ behaviour. |
The Paradox of the Compositionality of Natural Language: A Neural Machine Translation Case Study (2022.acl-long)
Copied to clipboard
| Challenge: | Obtaining human-like performance in NLP is often argued to require compositional generalisation. |
| Approach: | They re-instantiate three compositionality tests from the literature and reformulate them for neural machine translation. |
| Outcome: | The proposed models are more compositional than models trained on more data, the authors show . they also show that some non-compositional behaviours are mistakes, whereas others reflect natural variation in data. |