Mark Anderson

Anthi Papadopoulou, Pierre Lison, Mark Anderson et al.

Text sanitization is the task of redacting a document to mask all occurrences of (direct or indirect) personal identifiers, with the goal of concealing the identity of the individual(s) referred in it. In this paper, we consider a two-step approach to text sanitization and provide a detailed analysis of its empirical performance on two recently published datasets: the Text Anonymization Benchmark (Pilán et al., 2022) and a collection of Wikipedia biographies (Papadopoulou et al., 2022). The text sanitization process starts with a privacy-oriented entity recognizer that seeks to determine the text spans expressing identifiable personal information. This privacy-oriented entity recognizer is trained by combining a standard named entity recognition model with a gazetteer populated by person-related terms extracted from Wikidata. The second step of the text sanitization process consists in assessing the privacy risk associated with each detected text span, either isolated or in combination with other text spans. We present five distinct indicators of the re-identification risk, respectively based on language model probabilities, text span classification, sequence labelling, perturbations, and web search. We provide a contrastive analysis of each privacy indicator and highlight their benefits and limitations, notably in relation to the available labeled data.

Mark Anderson, Jose Camacho-Collados

The increase in performance in NLP due to the prevalence of distributional models and deep learning has brought with it a reciprocal decrease in interpretability. This has spurred a focus on what neural networks learn about natural language with less of a focus on how. Some work has focused on the data used to develop data-driven models, but typically this line of work aims to highlight issues with the data, e.g. highlighting and offsetting harmful biases. This work contributes to the relatively untrodden path of what is required in data for models to capture meaningful representations of natural language. This entails evaluating how well English and Spanish semantic spaces capture a particular type of relational knowledge, namely the traits associated with concepts (e.g. bananas-yellow), and exploring the role of co-occurrences in this context.

Alberto Muñoz-Ortiz, Mark Anderson, David Vilares et al.

PoS tags, once taken for granted as a useful resource for syntactic parsing, have become more situational with the popularization of deep learning. Recent work on the impact of PoS tags on graph- and transition-based parsers suggests that they are only useful when tagging accuracy is prohibitively high, or in low-resource scenarios. However, such an analysis is lacking for the emerging sequence labeling parsing paradigm, where it is especially relevant as some models explicitly use PoS tags for encoding and decoding. We undertake a study and uncover some trends. Among them, PoS tags are generally more useful for sequence labeling parsers than for other paradigms, but the impact of their accuracy is highly encoding-dependent, with the PoS-based head-selection encoding being best only when both tagging accuracy and resource availability are high.

Mark Anderson, Carlos Gómez-Rodríguez

We contribute to the discussion on parsing performance in NLP by introducing a measurement that evaluates the differences between the distributions of edge displacement (the directed distance of edges) seen in training and test data. We hypothesize that this measurement will be related to differences observed in parsing performance across treebanks. We motivate this by building upon previous work and then attempt to falsify this hypothesis by using a number of statistical methods. We establish that there is a statistical correlation between this measurement and parsing performance even when controlling for potential covariants. We then use this to establish a sampling technique that gives us an adversarial and complementary split. This gives an idea of the lower and upper bounds of parsing systems for a given treebank in lieu of freshly sampled data. In a broader sense, the methodology presented here can act as a reference for future correlation-based exploratory work in NLP.

Parameswaran Kamalaruban, Mark Anderson, Stuart Burrell et al.

Pre-trained foundation models can be adapted for specific tasks using Low-Rank Adaptation (LoRA). However, the fairness properties of these adapted classifiers remain underexplored. Existing fairness-aware fine-tuning methods rely on direct access to sensitive attributes or their predictors, but in practice, these sensitive attributes are often held under strict consumer privacy controls, and neither the attributes nor their predictors are available to model developers, hampering the development of fair models. To address this issue, we introduce a set of LoRA-based fine-tuning methods that can be trained in a distributed fashion, where model developers and fairness auditors collaborate without sharing sensitive attributes or predictors. In this paper, we evaluate three such methods - sensitive unlearning, adversarial training, and orthogonality loss - against a fairness-unaware baseline, using experiments on the CelebA and UTK-Face datasets with an ImageNet pre-trained ViT-Base model. We find that orthogonality loss consistently reduces bias while maintaining or improving utility, whereas adversarial training improves False Positive Rate Parity and Demographic Parity in some cases, and sensitive unlearning provides no clear benefit. In tasks where significant biases are present, distributed fairness-aware fine-tuning methods can effectively eliminate bias without compromising consumer privacy and, in most cases, improve model utility.

Pierre Lison, Mark Anderson

While de-identification models can help conceal the identity of the individual(s) mentioned in a document, they fail to address linkage risks, defined as the potential to map the de-identified text back to its source. One straightforward way to perform such linkages is to extract phrases from the de-identified document and then check their presence in the original dataset. This paper presents a method to counter search-based linkage attacks while preserving the semantic integrity of the text. The method proceeds in two steps. We first construct an inverted index of the N-grams occurring in the document collection, making it possible to efficiently determine which N-grams appear in less than $k$ documents (either alone or in combination with other N-grams). An LLM-based rewriter is then iteratively queried to reformulate those spans until linkage is no longer possible. Experimental results on a collection of court cases show that the method is able to effectively prevent search-based linkages while remaining faithful to the original content.

Luca Pion-Tonachini, Kristofer Bouchard, Hector Garcia Martin et al.

We outline emerging opportunities and challenges to enhance the utility of AI for scientific discovery. The distinct goals of AI for industry versus the goals of AI for science create tension between identifying patterns in data versus discovering patterns in the world from data. If we address the fundamental challenges associated with "bridging the gap" between domain-driven scientific models and data-driven AI learning machines, then we expect that these AI models can transform hypothesis generation, scientific discovery, and the scientific process itself.

Mark Anderson, Carlos Gómez-Rodríguez

We present the system submission from the FASTPARSE team for the EUD Shared Task at IWPT 2021. We engaged in the task last year by focusing on efficiency. This year we have focused on experimenting with new ideas on a limited time budget. Our system is based on splitting the EUD graph into several trees, based on linguistic criteria. We predict these trees using a sequence-labelling parser and combine them into an EUD graph. The results were relatively poor, although not a total disaster and could probably be improved with some polishing of the system's rough edges.

Mark Anderson, Mathieu Dehouck, Carlos Gómez Rodríguez

We evaluate the efficacy of predicted UPOS tags as input features for dependency parsers in lower resource settings to evaluate how treebank size affects the impact tagging accuracy has on parsing performance. We do this for real low resource universal dependency treebanks, artificially low resource data with varying treebank sizes, and for very small treebanks with varying amounts of augmented data. We find that predicted UPOS tags are somewhat helpful for low resource treebanks, especially when fewer fully-annotated trees are available. We also find that this positive impact diminishes as the amount of data increases.

Mark Anderson, Carlos Gómez Rodríguez

We evaluate three leading dependency parser systems from different paradigms on a small yet diverse subset of languages in terms of their accuracy-efficiency Pareto front. As we are interested in efficiency, we evaluate core parsers without pretrained language models (as these are typically huge networks and would constitute most of the compute time) or other augmentations that can be transversally applied to any of them. Biaffine parsing emerges as a well-balanced default choice, with sequence-labelling parsing being preferable if inference speed (but not training energy cost) is the priority.

Yova Kementchedjhieva, Mark Anderson, Anders Søgaard

Some interpersonal verbs can implicitly attribute causality to either their subject or their object and are therefore said to carry an implicit causality (IC) bias. Through this bias, causal links can be inferred from a narrative, aiding language comprehension. We investigate whether pre-trained language models (PLMs) encode IC bias and use it at inference time. We find that to be the case, albeit to different degrees, for three distinct PLM architectures. However, causes do not always need to be implicit -- when a cause is explicitly stated in a subordinate clause, an incongruent IC bias associated with the verb in the main clause leads to a delay in human processing. We hypothesize that the temporary challenge humans face in integrating the two contradicting signals, one from the lexical semantics of the verb, one from the sentence-level semantics, would be reflected in higher error rates for models on tasks dependent on causal links. The results of our study lend support to this hypothesis, suggesting that PLMs tend to prioritize lexical patterns over higher-order signals.

Mark Anderson, Anders Søgaard, Carlos Gómez Rodríguez

Søgaard (2020) obtained results suggesting the fraction of trees occurring in the test data isomorphic to trees in the training set accounts for a non-trivial variation in parser performance. Similar to other statistical analyses in NLP, the results were based on evaluating linear regressions. However, the study had methodological issues and was undertaken using a small sample size leading to unreliable results. We present a replication study in which we also bin sentences by length and find that only a small subset of sentences vary in performance with respect to graph isomorphism. Further, the correlation observed between parser performance and graph isomorphism in the wild disappears when controlling for covariants. However, in a controlled experiment, where covariants are kept fixed, we do observe a strong correlation. We suggest that conclusions drawn from statistical analyses like this need to be tempered and that controlled experiments can complement them by more readily teasing factors apart.

Mark Anderson, Carlos Gómez-Rodríguez

We present an error analysis of neural UPOS taggers to evaluate why using gold standard tags has such a large positive contribution to parsing performance while using predicted UPOS tags either harms performance or offers a negligible improvement. We evaluate what neural dependency parsers implicitly learn about word types and how this relates to the errors taggers make to explain the minimal impact using predicted tags has on parsers. We also present a short analysis on what contexts result in reductions in tagging performance. We then mask UPOS tags based on errors made by taggers to tease away the contribution of UPOS tags which taggers succeed and fail to classify correctly and the impact of tagging errors.

Mark Anderson, Carlos Gómez-Rodríguez

We present an analysis on the effect UPOS accuracy has on parsing performance. Results suggest that leveraging UPOS tags as features for neural parsers requires a prohibitively high tagging accuracy and that the use of gold tags offers a non-linear increase in performance, suggesting some sort of exceptionality. We also investigate what aspects of predicted UPOS tags impact parsing accuracy the most, highlighting some potentially meaningful linguistic facets of the problem.

Mark Anderson, Carlos Gómez-Rodríguez

The carbon footprint of natural language processing research has been increasing in recent years due to its reliance on large and inefficient neural network implementations. Distillation is a network compression technique which attempts to impart knowledge from a large model to a smaller one. We use teacher-student distillation to improve the efficiency of the Biaffine dependency parser which obtains state-of-the-art performance with respect to accuracy and parsing speed (Dozat and Manning, 2017). When distilling to 20\% of the original model's trainable parameters, we only observe an average decrease of $\sim$1 point for both UAS and LAS across a number of diverse Universal Dependency treebanks while being 2.30x (1.19x) faster than the baseline model on CPU (GPU) at inference time. We also observe a small increase in performance when compressing to 80\% for some treebanks. Finally, through distillation we attain a parser which is not only faster but also more accurate than the fastest modern parser on the Penn Treebank.

Mathieu Dehouck, Mark Anderson, Carlos Gómez-Rodríguez

We present the system submission from the FASTPARSE team for the EUD Shared Task at IWPT 2020. We engaged with the task by focusing on efficiency. For this we considered training costs and inference efficiency. Our models are a combination of distilled neural dependency parsers and a rule-based system that projects UD trees into EUD graphs. We obtained an average ELAS of 74.04 for our official submission, ranking 4th overall.

Mark Anderson, Carlos Gómez-Rodríguez

A wide variety of transition-based algorithms are currently used for dependency parsers. Empirical studies have shown that performance varies across different treebanks in such a way that one algorithm outperforms another on one treebank and the reverse is true for a different treebank. There is often no discernible reason for what causes one algorithm to be more suitable for a certain treebank and less so for another. In this paper we shed some light on this by introducing the concept of an algorithm's inherent dependency displacement distribution. This characterises the bias of the algorithm in terms of dependency displacement, which quantify both distance and direction of syntactic relations. We show that the similarity of an algorithm's inherent distribution to a treebank's displacement distribution is clearly correlated to the algorithm's parsing performance on that treebank, specifically with highly significant and substantial correlations for the predominant sentence lengths in Universal Dependency treebanks. We also obtain results which show a more discrete analysis of dependency displacement does not result in any meaningful correlations.

Mark Anderson, David Vilares, Carlos Gómez-Rodríguez

We introduce a language-agnostic evolutionary technique for automatically extracting chunks from dependency treebanks. We evaluate these chunks on a number of morphosyntactic tasks, namely POS tagging, morphological feature tagging, and dependency parsing. We test the utility of these chunks in a host of different ways. We first learn chunking as one task in a shared multi-task framework together with POS and morphological feature tagging. The predictions from this network are then used as input to augment sequence-labelling dependency parsing. Finally, we investigate the impact chunks have on dependency parsing in a multi-task framework. Our results from these analyses show that these chunks improve performance at different levels of syntactic abstraction on English UD treebanks and a small, diverse subset of non-English UD treebanks.

John Collins, Brian Farrimond, David Flower et al.

Computer programs often behave differently under different compilers or in different computing environments. Relative debugging is a collection of techniques by which these differences are analysed. Differences may arise because of different interpretations of errors in the code, because of bugs in the compilers or because of numerical drift, and all of these were observed in the present study. Numerical drift arises when small and acceptable differences in values computed by different systems are integrated, so that the results drift apart. This is well understood and need not degrade the validity of the program results. Coding errors and compiler bugs may degrade the results and should be removed. This paper describes a technique for the comparison of two program runs which removes numerical drift and therefore exposes coding and compiler errors. The procedure is highly automated and requires very little intervention by the user. The technique is applied to the Weather Research and Forecasting model, the most widely used weather and climate modelling code.