publications

2021

1. Generalising to German Plural Noun Classes, from the Perspective of a Recurrent Neural Network

   Dankers, Verna, Langedijk, Anna, McCurdy, Kate, Williams, Adina, and Hupkes, Dieuwke

   In Proceedings of the 25th Conference on Computational Natural Language Learning (CoNLL) 2021

   Abs

   Inflectional morphology has since long been a useful testing ground for broader questions about generalisation in language and the viability of neural network models as cognitive models of language. Here, in line with that tradition, we explore how recurrent neural networks acquire the complex German plural system and reflect upon how their strategy compares to human generalisation and rule-based models of this system. We perform analyses including behavioural experiments, diagnostic classification, representation analysis and causal interventions, suggesting that the models rely on features that are also key predictors in rule-based models of German plurals. However, the models also display shortcut learning, which is crucial to overcome in search of more cognitively plausible generalisation behaviour.

2. Adaptor Grammars for Unsupervised Paradigm Clustering

   McCurdy, Kate, Goldwater, Sharon, and Lopez, Adam

   In Proceedings of the 18th SIGMORPHON Workshop on Computational Research in Phonetics, Phonology, and Morphology 2021

   Abs

   This work describes the Edinburgh submission to the SIGMORPHON 2021 Shared Task 2 on unsupervised morphological paradigm clustering. Given raw text input, the task was to assign each token to a cluster with other tokens from the same paradigm. We use Adaptor Grammar segmentations combined with frequency-based heuristics to predict paradigm clusters. Our system achieved the highest average F1 score across 9 test languages, placing first out of 15 submissions.

2020

1. Conditioning, but on Which Distribution? Grammatical Gender in German Plural Inflection

   McCurdy, Kate, Lopez, Adam, and Goldwater, Sharon

   In Proceedings of the Workshop on Cognitive Modeling and Computational Linguistics Nov 2020

   Abs Data Poster

   Grammatical gender is a consistent and informative cue to the plural class of German nouns. We find that neural encoder-decoder models learn to rely on this cue to predict plural class, but adult speakers are relatively insensitive to it. This suggests that the neural models are not an effective cognitive model of German plural formation.

2. Inflecting When There’s No Majority: Limitations of Encoder-Decoder Neural Networks as Cognitive Models for German Plurals

   McCurdy, Kate, Goldwater, Sharon, and Lopez, Adam

   In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics Jul 2020

   Abs arXiv Blog Data

   Can artificial neural networks learn to represent inflectional morphology and generalize to new words as human speakers do? Kirov and Cotterell (2018) argue that the answer is yes: modern Encoder-Decoder (ED) architectures learn human-like behavior when inflecting English verbs, such as extending the regular past tense form /-(e)d/ to novel words. However, their work does not address the criticism raised by Marcus et al. (1995): that neural models may learn to extend not the regular, but the most frequent class — and thus fail on tasks like German number inflection, where infrequent suffixes like /-s/ can still be productively generalized. To investigate this question, we first collect a new dataset from German speakers (production and ratings of plural forms for novel nouns) that is designed to avoid sources of information unavailable to the ED model. The speaker data show high variability, and two suffixes evince ‘regular’ behavior, appearing more often with phonologically atypical inputs. Encoder-decoder models do generalize the most frequently produced plural class, but do not show human-like variability or ‘regular’ extension of these other plural markers. We conclude that modern neural models may still struggle with minority-class generalization.

3. Modeling grammatical gender and plural inflection in German

   McCurdy, Kate, Lopez, Adam, and Goldwater, Sharon

   In Proceedings of the 26 Architectures and Mechanisms for Language Processing Conference (AMLaP) Sep 2020

   Data

2019

1. Tutorbot Corpus: Evidence of Human-Agent Verbal Alignment in Second Language Learner Dialogues

   Sinclair, Arabella, McCurdy, Kate, Lucas, Christopher G, Lopez, Adam, and Gašević, Dragan

   In Proceedings of the 12th International Conference on Educational Data Mining 2019

   Abs

   Prior research has shown that, under certain conditions, Human-Agent (H-A) alignment exists to a stronger degree than that found in Human-Human (H-H) communication. In an H-H Second Language (L2) setting, evidence of alignment has been linked to learning and teaching strategy. We present a novel analysis of H-A and H-H L2 learner dialogues using automated metrics of alignment. Our contributions are twofold: firstly we replicated the reported H-A alignment within an educational context, finding L2 students align to an automated tutor. Secondly, we performed an exploratory comparison of the alignment present in comparable H-A and H-H L2 learner corpora using Bayesian Gaussian Mixture Models (GMMs), finding preliminary evidence that students in H-A L2 dialogues showed greater variability in engagement.

2017

1. Grammatical gender associations outweigh topical gender bias in crosslinguistic word embeddings

   McCurdy, Katherine, and Serbetçi, Oğuz

   2017

   Abs arXiv Code

   Recent research has demonstrated that vector space models of semantics can reflect undesirable biases in human culture. Our investigation of crosslinguistic word embeddings reveals that topical gender bias interacts with, and is surpassed in magnitude by, the effect of grammatical gender associations, and both may be attenuated by corpus lemmatization. This finding has implications for downstream applications such as machine translation.

2. Linked Data for Language-Learning Applications

   Loughnane, Robyn, McCurdy, Kate, Kolb, Peter, and Selent, Stefan

   In Proceedings of the 12th Workshop on Innovative Use of NLP for Building Educational Applications Sep 2017

   Abs

   The use of linked data within language-learning applications is an open research question. A research prototype is presented that applies linked-data principles to store linguistic annotation generated from language-learning content using a variety of NLP tools. The result is a database that links learning content, linguistic annotation and open-source resources, on top of which a diverse range of tools for language-learning applications can be built.

2013

1. Implicit prosody and contextual bias in silent reading

   McCurdy, Kate, Kentner, Gerrit, and Vasishth, Shravan

   Journal of Eye Movement Research Jul 2013

   Abs

   Eye-movement research on implicit prosody has found effects of lexical stress on syntactic ambiguity resolution, suggesting that metrical well-formedness constraints interact with syntactic category assignment. Building on these findings, the present eyetracking study investigates whether contextual bias can modulate the effects of metrical structure on syntactic ambiguity resolution in silent reading. Contextual bias and potential stress-clash in the ambiguous region were crossed in a 2 2 design. Participants read biased context sentences followed by temporarily ambiguous test sentences. In the three-word ambiguous region, main effects of lexical stress were dominant, while early effects of context were absent. Potential stress clash yielded a significant increase in first-pass regressions and re-reading probability across the three words. In the disambiguating region, the disambiguating word itself showed increased processing difficulty (lower skipping and increased re-reading probability) when the disambiguation engendered a stress clash configuration, while the word immediately following showed main effects of context in those same measures. Taken together, effects of lexical stress upon eye movements were swift and pervasive across first-pass and second-pass measures, while effects of context were relatively delayed. These results indicate a strong role for implicit meter in guiding parsing, one that appears insensitive to higher-level constraints. Our findings are problematic for two classes of models, the two-stage garden-path model and the constraint-based competition-integration model, but can be explained by a variation on the two-stage model, the unrestricted race model.

2010

1. Poetic rhyme reflects cross-linguistic differences in information structure

   Wagner, Michael, and McCurdy, Katherine

   Cognition Nov 2010

   Abs

   Identical rhymes (right/write, attire/retire) are considered satisfactory and even artistic in French poetry but are considered unsatisfactory in English. This has been a consistent generalization over the course of centuries, a surprising fact given that other aspects of poetic form in French were happily applied in English. This paper puts forward the hypothesis that this difference is not merely one of poetic tradition, but is grounded in the distinct ways in which information-structure affects prosody in the two languages. A study of rhyme usage in poetry and a perception experiment confirm that native speakers’ intuitions about rhyming in the two languages indeed differ, and a further perception experiment supports the hypothesis that this fact is due to a constraint on prosody that is active in English but not in French. The findings suggest that certain forms of artistic expression in poetry are influenced, and even constrained, by more general properties of a language.