Cambridge Language Sciences Symposium for Early-Careers Researchers 2019
Date and time
Location
The Cavonius Centre, Harvey Court
Gonville and Caius College West Road Cambridge CB3 9DS United KingdomDescription
Presenting work by early-careers researchers at the University of Cambridge
This free event is open to all language sciences researchers at the University of Cambridge and is supported by Cambridge Language Sciences IRC (Interdisciplinary Research Centre).
13.00 Registration & coffee
13.25 Welcome (Dr Andrew Caines)
13.30-14.00 Dr Tam Blaxter (Theoretical & Applied Linguistics)
14.00-14.30 Jonathan R Goodman (Leverhulme Centre for Human Evolutionary Studies)
14.30-15.00 Giulia Bovolenta (Theoretical & Applied Linguistics)
15.00-16.00 Poster exhibition & refreshment break
16.00-16.30 Dr Avin Mirawdeli (Education)
16.30-17.00 Christopher Davis (Computer Science & Technology)
17.00-17.30 Ediz Sohoglu (MRC Cognition & Brain Sciences Unit)
17.30-18.00 Dr Elaine Schmidt (Cambridge Assessment English)
18.00-19.30 Drinks & canapés
Abstracts
Variationist evidence in explanations for the loss of the nominative in Norwegian
Dr Tam Blaxter (Theoretical & Applied Linguistics)
Traditional accounts of the loss of case in Western European languages tell a story of phonological erosion leading to increased syncretism: sound change leaves so few morphological contexts in which contrasts are maintained that categories become unlearnable. Detailed investigations into the loss of the nominative-accusative distinction in Old French (Schøsler 2001; Ashdowne & Smith 2007) have shown such phonology-first accounts to be simplistic: there are semantic and morphological constraints on the change which cannot be explained by patterns of sound change and syncretism. For the same change in Middle Norwegian, inference from modern dialect data (Enger 2013) and small-scale textual investigations (Wetås 2008) also speak against pure phonological explanation. In contrast, other scholars find an external explanation in contact with Middle Low German and imperfect learning of Norwegian (O’Neil 1978; Werner 1984; Trudgill 2011; McWhorter 2011); a related possibility is that the change should be added to the list of results of shift from Sámi to Norwegian in northern Norway (Kusmenko 2008; Bull 2011). In this paper, I explore detailed variational evidence for the loss of the nominative in Middle Norwegian. Conditioning by morphological class might speak to a role for sound change and syncretism, yet conditioning by morphosyntactic factors undermines a pure phonological account. Geographical evidence can be used to test contact accounts (if contact is involved, innovations must have diffused from centres of contact) and phonological accounts (innovations should have started in dialects with the relevant sound changes).
Do barriers direct the evolution of linguistic signals?
Jonathan R Goodman (Leverhulme Centre for Human Evolutionary Studies)
The Red Queen model suggests that evolution is a persistent competitive race between lineages, each trying to stay ahead of the game (eg, arms races, predator-prey relations, hosts and parasites). This race may be ended, paused, or redirected, which may be explained by barrier theory, a recently advanced theory in cell biology. Barriers may be understood as naturally evolved or artificially implemented mechanisms for blocking the risk of exploitation within a group, and in turn predict how free-riders are likely to attempt to exploit systems in novel ways. The mode of exploitation is likely to be innovative, and therefore cannot be accounted for in a standard signalling game model. In this talk I will suggest that this theory is applicable to non-biological systems, including those centred on human language. If language is conceived as a competitive interaction between signaller and listener, then barrier theory may be a helpful metaphor for understanding how that competitive interaction may be stopped, slowed, or redirected. We might expect, for example, that if there are any linguistic mechanisms for signalling kinship affiliation, that some fakers will become adept at feigning those signals effectively, which will in turn lead to mechanisms for detecting fakes. I will give an overview of this position and discuss some of our research thus far.
Applying implicit learning paradigms to generate productive rule knowledge in L2 acquisition
Giulia Bovolenta (Theoretical & Applied Linguistics)
Implicit learning (IL) is the process through which we can incidentally learn a regularity while remaining unaware of what we have learnt. The application of IL paradigms to second language acquisition research has shown that IL can occur with a variety of linguistic features. However, these studies usually test for the development of implicit knowledge with comprehension tasks only; they do not check whether participants demonstrate sensitivity to the rule in production, even though this is a fundamental aspect of language acquisition. To address the question, we ran a series of experiments in which participants were exposed to a novel rule based on Czech spatial prepositions v and na, which both mean “in, at” but alternate according to the type of place described by the noun (open vs bounded space). Subjects were exposed aurally to English sentences in which spatial prepositions were replaced by pseudowords based on the novel rule, together with a visual stimulus depicting the content of the sentence. The exposure phase consisted of a short-term recall task in which pictures and auditory descriptions were presented together, followed by oral recall cued by each picture. Following training, participants were tested on both long-term cued recall of training items and elicited production with novel items (generalization task), which followed the same procedure as recall items. Awareness of the target rule was assessed by post-test questionnaire. Our results show that even participants who remained unaware of the rule showed above-chance accuracy in both long-term recall and generalization tasks, suggesting that rule knowledge acquired through implicit learning can be used productively to generate new instances. However, learning outcomes were affected by changes to the learning procedure, suggesting that mere exposure was not enough and that the recall task used was crucial in generating the relevant kind of knowledge.
Assessing speech and language needs in children aged 4-6 through an app
Dr Avin Mirawdeli (Education)
Early identification and intervention for speech, language and communication Needs (SLCN) and disorders is important for optimal client outcomes. I will present preliminary findings from the development of a new app for Speech and Language Assessment in Schools and Homes, Splash, which aims to provide a first screening of SLCNs in children aged 4-6 years. The app aims to be easy-to administer, with an adult (such as a teacher or a parent) directing the child through the tasks. Three fun games have been developed to assess receptive and expressive language and connected speech, respectively. At this time, the connected speech assessment is being validated and is based on findings from Mirawdeli (2015) and Mirawdeli & Howell (2016) that showed a relatively simple speech sample could be used to successfully identify those at risk of SLCNs with good sensitivity and specificity. Currently in proof-of-concept mode, when complete Splash will use automatic spoken language processing to give an instant estimate of a child’s communication ability by providing a norm-referenced ‘risk’ score for each child, and information to support referral to specialist services, if this is indicated. The intention is to empower parents and teachers, and to give them confidence and information to support referral to specialist services.
Improving grammatical error detection for learner English text using first-language specific priors
Christopher Davis (Computer Science & Technology)
This talk will present a neural sequence labelling approach to grammatical error detection (GED) which makes use of first-language (L1) specific priors from errors found in learner English essays. This approach to GED is motivated by the general understanding that a learner's L1 influences their use of a second language. We quantify 17 grammatical error types observed in error-annotated essays, grouped by author L1, and hypothesize that this additional information will enable the sequence labeller to predict L1 transfer errors it would not otherwise identify. We investigate effective ways to represent grammatical error vectors for 16 L1s and propose an approach to incorporating such priors into the neural network. We experimentally demonstrate that our approach not only improves the performance of a state-of-the-art GED model but also substantially improves the detection of specific error types. Finally, we group the L1s into 9 language genera (Germanic, Romance, Slavic, etc) and investigate how feature vectors at the genus level further assist in GED. Such groupings may be important in the scenario where an L1 is poorly represented in the training data but data for related languages are more abundant.
Testing computational theories of spoken language processing using brain imaging
Ediz Sohoglu (MRC Cognition & Brain Sciences Unit)
Spoken language is rich in structure, and individual elements like phonemes can be predicted even before they are heard. What are the neural computations by which predictions are combined with speech input? According to one proposal, speech representations are enhanced by prediction to leave a ‘sharpened’ version of the sensory input (McClelland and Elman, 1986)? An alternative account, however, proposes that predictions are subtracted away from speech representations to leave only those parts that are unexpected i.e. ‘prediction error’ (Rao and Ballard, 1999)? Recent work suggests that for degraded speech, brain responses are best explained by prediction error representations (Blank and Davis, 2016). However, that study used an artificial listening situation in which speech was highly distorted and strong predictions obtained from prior written text.
In this talk I will describe magnetoencephalography (MEG) evidence of prediction error representations in a more naturalistic listening situation in which speech is clearly presented and predictions obtained directly from the speech signal itself (i.e. from lexical content). These findings suggest that prediction errors are a general and fundamental component of neural processing for speech.
Processing global and local information in text vs. images: the contribution of visual and verbal working memory
Dr Elaine Schmidt (Cambridge Assessment English)
Studies on language comprehension frequently use multimodal situations, i.e., text information is written whereas comprehension is assessed through images. Despite this, no study as yet has looked at whether central (global) or peripheral (local) information in a text affects processing across modalities, and whether individual differences on working memory have an impact on this distinction. In this study, participants were presented with short texts followed by images which either matched or differed in global/local information while we tracked their eye movements. Comprehension questions assessed whether participants noticed mismatches, and background tasks measured working memory in participants. Eye movement results demonstrated that higher verbal working memory was associated with shorter fixation durations when grasping local but not global text information. In contrast, visual working memory did not explain comprehension differences for global or local information on images. Our study sheds light on the relationship between the nature of information (global vs local) and the modality of presentation (linguistic vs pictorial). These findings have important implications for the design and assessment of language comprehension tests.
Posters
Poster exhibition organised by Dr Andrew Caines (ALTA Institute/Computer Science & Technology).
Detecting personal attributes through analyzing online forums
Zhilin Wang, Xiaodong Wu, Weizhe Lin and Elena Rastorgueva, Faculty of Education
Talking to think, thinking to talk: a mixed methods study of the development of creativity in L2 talk
Hui Ki Chan, Faculty of Education
A computational mechanistic account of hemisphere differences in language processing
Ya-Ning Chang, Matthew Lambon Ralph, MRC Cognition and Brain Sciences Unit
Autoencoding Pixies: amortised variational inference for functional distributional semantics using graph convolutions
Guy Emerson, Dept. of Computer Science & Technology