Simone Teufel

Josef Valvoda, Ryan Cotterell, Simone Teufel · cambridge

Every legal case sets a precedent by developing the law in one of the following two ways. It either expands its scope, in which case it sets positive precedent, or it narrows it, in which case it sets negative precedent. Legal outcome prediction, the prediction of positive outcome, is an increasingly popular task in AI. In contrast, we turn our focus to negative outcomes here, and introduce a new task of negative outcome prediction. We discover an asymmetry in existing models' ability to predict positive and negative outcomes. Where the state-of-the-art outcome prediction model we used predicts positive outcomes at 75.06 F1, it predicts negative outcomes at only 10.09 F1, worse than a random baseline. To address this performance gap, we develop two new models inspired by the dynamics of a court process. Our first model significantly improves positive outcome prediction score to 77.15 F1 and our second model more than doubles the negative outcome prediction performance to 24.01 F1. Despite this improvement, shifting focus to negative outcomes reveals that there is still much room for improvement for outcome prediction models.

Rowan Hall Maudslay, Simone Teufel

Linguists distinguish between novel and conventional metaphor, a distinction which the metaphor detection task in NLP does not take into account. Instead, metaphoricity is formulated as a property of a token in a sentence, regardless of metaphor type. In this paper, we investigate the limitations of treating conventional metaphors in this way, and advocate for an alternative which we name 'metaphorical polysemy detection' (MPD). In MPD, only conventional metaphoricity is treated, and it is formulated as a property of word senses in a lexicon. We develop the first MPD model, which learns to identify conventional metaphors in the English WordNet. To train it, we present a novel training procedure that combines metaphor detection with word sense disambiguation (WSD). For evaluation, we manually annotate metaphor in two subsets of WordNet. Our model significantly outperforms a strong baseline based on a state-of-the-art metaphor detection model, attaining an ROC-AUC score of .78 (compared to .65) on one of the sets. Additionally, when paired with a WSD model, our approach outperforms a state-of-the-art metaphor detection model at identifying conventional metaphors in text (.659 F1 compared to .626).

Rowan Hall Maudslay, Simone Teufel

A widely acknowledged shortcoming of WordNet is that it lacks a distinction between word meanings which are systematically related (polysemy), and those which are coincidental (homonymy). Several previous works have attempted to fill this gap, by inferring this information using computational methods. We revisit this task, and exploit recent advances in language modelling to synthesise homonymy annotation for Princeton WordNet. Previous approaches treat the problem using clustering methods; by contrast, our method works by linking WordNet to the Oxford English Dictionary, which contains the information we need. To perform this alignment, we pair definitions based on their proximity in an embedding space produced by a Transformer model. Despite the simplicity of this approach, our best model attains an F1 of .97 on an evaluation set that we annotate. The outcome of our work is a high-quality homonymy annotation layer for Princeton WordNet, which we release.

Chengzu Li, Chao Zhang, Simone Teufel et al. · cambridge

We are interested in the generation of navigation instructions, either in their own right or as training material for robotic navigation task. In this paper, we propose a new approach to navigation instruction generation by framing the problem as an image captioning task using semantic maps as visual input. Conventional approaches employ a sequence of panorama images to generate navigation instructions. Semantic maps abstract away from visual details and fuse the information in multiple panorama images into a single top-down representation, thereby reducing computational complexity to process the input. We present a benchmark dataset for instruction generation using semantic maps, propose an initial model and ask human subjects to manually assess the quality of generated instructions. Our initial investigations show promise in using semantic maps for instruction generation instead of a sequence of panorama images, but there is vast scope for improvement. We release the code for data preparation and model training at https://github.com/chengzu-li/VLGen.

Zhihan Cao, Hiroaki Yamada, Simone Teufel et al.

Recently, much work has concerned itself with the enigma of what exactly pretrained language models~(PLMs) learn about different aspects of language, and how they learn it. One stream of this type of research investigates the knowledge that PLMs have about semantic relations. However, many aspects of semantic relations were left unexplored. Generally, only one relation has been considered, namely hypernymy. Furthermore, previous work did not measure humans' performance on the same task as that performed by the PLMs. This means that at this point in time, there is only an incomplete view of the extent of these models' semantic relation knowledge. To address this gap, we introduce a comprehensive evaluation framework covering five relations beyond hypernymy, namely hyponymy, holonymy, meronymy, antonymy, and synonymy. We use five metrics (two newly introduced here) for recently untreated aspects of semantic relation knowledge, namely soundness, completeness, symmetry, prototypicality, and distinguishability. Using these, we can fairly compare humans and models on the same task. Our extensive experiments involve six PLMs, four masked and two causal language models. The results reveal a significant knowledge gap between humans and models for all semantic relations. In general, causal language models, despite their wide use, do not always perform significantly better than masked language models. Antonymy is the outlier relation where all models perform reasonably well. The evaluation materials can be found at https://github.com/hancules/ProbeResponses.

Igor Sterner, Simone Teufel

There is a lack of an evaluation methodology that estimates the extent to which large language models (LLMs) use code-switching (CS) in the same way as bilinguals. Existing methods do not have wide language coverage, fail to account for the diverse range of CS phenomena, or do not scale. We propose an intervention based on minimal pairs of CS. Each minimal pair contains one naturally occurring CS sentence and one minimally manipulated variant. We collect up to 1,000 such pairs each for 11 language pairs. Our human experiments show that, for every language pair, bilinguals consistently prefer the naturally occurring CS sentence. Meanwhile our experiments with current LLMs show that the larger the model, the more consistently it assigns higher probability to the naturally occurring CS sentence than to the variant. In accordance with theoretical claims, the largest probability differences arise in those pairs where the manipulated material consisted of closed-class words.

Igor Sterner, Simone Teufel

The theoretical code-switching (CS) literature provides numerous pointwise investigations that aim to explain patterns in CS, i.e. why bilinguals switch language in certain positions in a sentence more often than in others. A resulting consensus is that CS can be explained by the syntax of the contributing languages. There is however no large-scale, multi-language, cross-phenomena experiment that tests this claim. When designing such an experiment, we need to make sure that the system that is predicting where bilinguals tend to switch has access only to syntactic information. We provide such an experiment here. Results show that syntax alone is sufficient for an automatic system to distinguish between sentences in minimal pairs of CS, to the same degree as bilingual humans. Furthermore, the learnt syntactic patterns generalise well to unseen language pairs.

Andrii Nikolaiev, Yiannos Stathopoulos, Simone Teufel

In this paper we look at the ability of recent large language models (LLMs) at solving mathematical problems in combinatorics. We compare models LLaMA-2, LLaMA-3.1, GPT-4, and Mixtral against each other and against human pupils and undergraduates with prior experience in mathematical olympiads. To facilitate these comparisons we introduce the Combi-Puzzles dataset, which contains 125 problem variants based on 25 combinatorial reasoning problems. Each problem is presented in one of five distinct forms, created by systematically manipulating the problem statements through adversarial additions, numeric parameter changes, and linguistic obfuscation. Our variations preserve the mathematical core and are designed to measure the generalisability of LLM problem-solving abilities, while also increasing confidence that problems are submitted to LLMs in forms that have not been seen as training instances. We found that a model based on GPT-4 outperformed all other models in producing correct responses, and performed significantly better in the mathematical variation of the problems than humans. We also found that modifications to problem statements significantly impact the LLM's performance, while human performance remains unaffected.

Zhihan Cao, Hiroaki Yamada, Simone Teufel et al.

WordNet provides a carefully constructed repository of semantic relations, created by specialists. But there is another source of information on semantic relations, the intuition of language users. We present the first systematic study of the degree to which these two sources are aligned. Investigating the cases of misalignment could make proper use of WordNet and facilitate its improvement. Our analysis which uses templates to elicit responses from human participants, reveals a general misalignment of semantic relation knowledge between WordNet and human intuition. Further analyses find a systematic pattern of mismatch among synonymy and taxonomic relations~(hypernymy and hyponymy), together with the fact that WordNet path length does not serve as a reliable indicator of human intuition regarding hypernymy or hyponymy relations.

Rowan Hall Maudslay, Simone Teufel, Francis Bond et al.

The senses of a word exhibit rich internal structure. In a typical lexicon, this structure is overlooked: a word's senses are encoded as a list without inter-sense relations. We present ChainNet, a lexical resource which for the first time explicitly identifies these structures. ChainNet expresses how senses in the Open English Wordnet are derived from one another: every nominal sense of a word is either connected to another sense by metaphor or metonymy, or is disconnected in the case of homonymy. Because WordNet senses are linked to resources which capture information about their meaning, ChainNet represents the first dataset of grounded metaphor and metonymy.

Tiago Pimentel, Clara Meister, Elizabeth Salesky et al.

While there exist scores of natural languages, each with its unique features and idiosyncrasies, they all share a unifying theme: enabling human communication. We may thus reasonably predict that human cognition shapes how these languages evolve and are used. Assuming that the capacity to process information is roughly constant across human populations, we expect a surprisal--duration trade-off to arise both across and within languages. We analyse this trade-off using a corpus of 600 languages and, after controlling for several potential confounds, we find strong supporting evidence in both settings. Specifically, we find that, on average, phones are produced faster in languages where they are less surprising, and vice versa. Further, we confirm that more surprising phones are longer, on average, in 319 languages out of the 600. We thus conclude that there is strong evidence of a surprisal--duration trade-off in operation, both across and within the world's languages.

Tiago Pimentel, Clara Meister, Simone Teufel et al.

Homophony's widespread presence in natural languages is a controversial topic. Recent theories of language optimality have tried to justify its prevalence, despite its negative effects on cognitive processing time; e.g., Piantadosi et al. (2012) argued homophony enables the reuse of efficient wordforms and is thus beneficial for languages. This hypothesis has recently been challenged by Trott and Bergen (2020), who posit that good wordforms are more often homophonous simply because they are more phonotactically probable. In this paper, we join in on the debate. We first propose a new information-theoretic quantification of a language's homophony: the sample Rényi entropy. Then, we use this quantification to revisit Trott and Bergen's claims. While their point is theoretically sound, a specific methodological issue in their experiments raises doubts about their results. After addressing this issue, we find no clear pressure either towards or against homophony -- a much more nuanced result than either Piantadosi et al.'s or Trott and Bergen's findings.

Jan Wira Gotama Putra, Simone Teufel, Takenobu Tokunaga

Argumentative structure prediction aims to establish links between textual units and label the relationship between them, forming a structured representation for a given input text. The former task, linking, has been identified by earlier works as particularly challenging, as it requires finding the most appropriate structure out of a very large search space of possible link combinations. In this paper, we improve a state-of-the-art linking model by using multi-task and multi-corpora training strategies. Our auxiliary tasks help the model to learn the role of each sentence in the argumentative structure. Combining multi-corpora training with a selective sampling strategy increases the training data size while ensuring that the model still learns the desired target distribution well. Experiments on essays written by English-as-a-foreign-language learners show that both strategies significantly improve the model's performance; for instance, we observe a 15.8% increase in the F1-macro for individual link predictions.

Rowan Hall Maudslay, Hila Gonen, Ryan Cotterell et al.

This paper treats gender bias latent in word embeddings. Previous mitigation attempts rely on the operationalisation of gender bias as a projection over a linear subspace. An alternative approach is Counterfactual Data Augmentation (CDA), in which a corpus is duplicated and augmented to remove bias, e.g. by swapping all inherently-gendered words in the copy. We perform an empirical comparison of these approaches on the English Gigaword and Wikipedia, and find that whilst both successfully reduce direct bias and perform well in tasks which quantify embedding quality, CDA variants outperform projection-based methods at the task of drawing non-biased gender analogies by an average of 19% across both corpora. We propose two improvements to CDA: Counterfactual Data Substitution (CDS), a variant of CDA in which potentially biased text is randomly substituted to avoid duplication, and the Names Intervention, a novel name-pairing technique that vastly increases the number of words being treated. CDA/S with the Names Intervention is the only approach which is able to mitigate indirect gender bias: following debiasing, previously biased words are significantly less clustered according to gender (cluster purity is reduced by 49%), thus improving on the state-of-the-art for bias mitigation.