Skip to main content
Språkbanken Text is a department within Språkbanken.

BibTeX

@article{themistocleous-2019-dialect-341526,
	title        = {Dialect Classification From a Single Sonorant Sound Using Deep Neural Networks},
	abstract     = {During spoken communication, the fine acoustic properties of human speech can reveal vital sociolinguistic and linguistic information about speakers and thus, these properties can function as reliable identification markers of speakers' identity. One key piece of information speech reveals is speakers' dialect. The first aim of this study is to provide a machine learning method that can distinguish the dialect from acoustic productions of sonorant sounds. The second aim is to determine the classification accuracy of dialects from the temporal and spectral information of a single sonorant sound and the classification accuracy of dialects using additional co-articulatory information from the adjacent vowel. To this end, this paper provides two classification approaches. The first classification approach aims to distinguish two Greek dialects, namely Athenian Greek, the prototypical form of Standard Modern Greek and Cypriot Greek using measures of temporal and spectral information (i.e., spectral moments) from four sonorant consonants /m n l r/. The second classification study aims to distinguish the dialects using coarticulatory information (e.g., formants frequencies F1 - F5, F0, etc.) from the adjacent vowel in addition to spectral and temporal information from sonorants. In both classification approaches, we have employed Deep Neural Networks, which we compared with Support Vector Machines, Random Forests, and Decision Trees. The findings show that neural networks distinguish the two dialects using a combination of spectral moments, temporal information, and formant frequency information with 81% classification accuracy, which is a 14% accuracy gain over employing temporal properties and spectral moments alone. In conclusion, Deep Neural Networks can classify the dialect from single consonant productions, making them capable of identifying sociophonetic shibboleths.},
	journal      = {FRONTIERS IN COMMUNICATION},
	author       = {Themistocleous, Charalambos},
	year         = {2019},
	volume       = {4},
}

@inProceedings{themistocleous-muller-2015-intonation-232414,
	title        = {The intonation of Albanian polar questions and statements},
	abstract     = {This studyaims to provide an account of the effects of sentence type (statements vs. polar questions) on Standard Albanian prenuclear rises through a polynomial model representing the dynamic characteristics of tonal contours.Results show that the main difference in contour shape between Albanian statements and polar questions is located in the shape of the prenuclear rise, and this difference was significant; onset timing of the prenuclear rise, however, did not differ significantly betweenthe two types of sentence.},
	booktitle    = {6th International Conference of Experimental Linguistics. ExLing 2015, 26-27 June 2015, Athens, Greece / Edited by  Antonis Botinis},
	author       = {Themistocleous, Charalambos and Müller, Daniela},
	year         = {2015},
	publisher    = {University of Athens},
	address      = {Athens},
	ISBN         = {978-960-466-160-2},
}

@article{themistocleous-logotheti-2016-standard-239899,
	title        = {Standard Modern Greek and Cypriot Greek vowels: a sociophonetic study},
	abstract     = {This study is a comparative analysis of Standard Modern Greek (SMG) and Cypriot Greek (CG) vowels. Specifically, the study examines the effects of vowel (/e i a o u/), language variety (SMG vs CG), and stress (stressed vs unstressed vowels) on vowel formants F1 and F2, vowel duration, and fundamental frequency (f0). 45 female speakers were recorded: 20 SMG speakers and 25 CG speakers from Athens and Nicosia respectively. The results showed significant effects of vowel, stress, and language variety on formants, duration and f0. The study confirms the findings of earlier studies on SMG vowels, provides the first report on CG vowels’ acoustic structure, and constitutes the first comparative sociophonetic research on SMG and CG vowels.
},
	journal      = {Proceedings of the international conference on Modern Greek dialects and Linguistic Theory, Patras, 25-28  September 2014},
	author       = {Themistocleous, Charalambos and Logotheti, Angeliki},
	year         = {2016},
	volume       = {6},
	number       = {1},
	pages        = {178--184},
}

@article{fyndanis-themistocleous-2019-there-268753,
	title        = {Are there prototypical associations between time frames and aspectual values? Evidence from Greek aphasia and healthy ageing},
	abstract     = {Time reference, which has been found to be selectively impaired in agrammatic aphasia, is often interwoven with grammatical aspect. A recent study on Russian aphasia found that time reference and aspect interact: Past reference was less impaired when tested within a perfective aspect context (compared to when tested within an imperfective aspect context), and reference to the non-past was less impaired when tested within an imperfective aspect context (compared to when tested within a perfective aspect context). To explain this pattern, the authors argued that there are prototypical associations between time frames and aspectual values. The present study explores the relationship between time reference and aspect focusing on Greek aphasia and healthy ageing and using a sentence completion task that crosses time reference and aspect. The findings do not support prototypical matches between different time frames and aspectual values. Building on relevant studies, we propose that patterns of performance of healthy or language-impaired speakers on constrained tasks tapping different combinations of time frames with aspectual values should reflect the relative frequency of these combinations in a given language. The analysis of the results at the individual level revealed a double dissociation, which indicates that a given time frame–aspectual value combination may be relatively easy to process for some persons with aphasia but demanding for some others.},
	journal      = {Clinical Linguistics & Phonetics},
	author       = {Fyndanis, Valantis and Themistocleous, Charalambos},
	year         = {2019},
	volume       = {33},
	number       = {1-2},
	pages        = {191--217},
}

@misc{themistocleous-etal-2023-assessing-331090,
	title        = {Assessing Language Disorders using Artificial Intelligence: a Paradigm Shift
},
	abstract     = {Speech, language, and communication deficits are present in most neurodegenerative syndromes. They enable the early detection, diagnosis, treatment planning, and monitoring of neurocognitive disease progression as part of traditional neurological assessment. Nevertheless, standard speech and language evaluation is time-consuming and resource-intensive for clinicians. We argue that using machine learning methodologies, natural language processing, and modern artificial intelligence (AI) for Language Assessment is an improvement over conventional manual assessment. Using these methodologies, Computational Language Assessment (CLA) accomplishes three goals: (i) provides a neuro-cognitive evaluation of speech, language, and communication in elderly and high-risk individuals for dementia; (ii) facilitates the diagnosis, prognosis, and therapy efficacy in at-risk and language-impaired populations; and (iii) allows easier extensibility to assess patients from a wide range of languages. By employing AI models, CLA may inform neurocognitive theory on the relationship between language symptoms and their neural bases. Finally, it signals a paradigm shift by significantly advancing our ability to optimize the prevention and treatment of elderly individuals with communication disorders, allowing them to age gracefully with social engagement.
},
	author       = {Themistocleous, Charalambos and Tsapkini,  Kyrana and Kokkinakis, Dimitrios},
	year         = {2023},
	publisher    = {arXiv.org},
}

@article{themistocleous-etal-2018-identification-273026,
	title        = {Identification of Mild Cognitive Impairment From Speech in Swedish Using Deep Sequential Neural Networks},
	abstract     = {While people with mild cognitive impairment (MCI) portray noticeably incipient memory difficulty in remembering events and situations along with problems in decision making, planning, and finding their way in familiar environments, detailed neuropsychological assessments also indicate deficits in language performance. To this day, there is no cure for dementia but early-stage treatment can delay the progression of MCI; thus, the development of valid tools for identifying early cognitive changes is of great importance. In this study, we provide an automated machine learning method, using Deep Neural Network Architectures, that aims to identify MCI. Speech materials were obtained using a reading task during evaluation sessions, as part of the Gothenburg MCI research study. Measures of vowel duration, vowel formants (F1 to F5), and fundamental frequency were calculated from speech signals. To learn the acoustic characteristics associated with MCI vs. healthy controls, we have trained and evaluated ten Deep Neural Network Architectures and measured how accurately they can diagnose participants that are unknown to the model. We evaluated the models using two evaluation tasks: a 5-fold crossvalidation and by splitting the data into 90% training and 10% evaluation set. The findings suggest first, that the acoustic features provide significant information for the identification of MCI; second, the best Deep Neural Network Architectures can classify MCI and healthy controls with high classification accuracy (M = 83%); and third, the model has the potential to offer higher accuracy than 84% if trained with more data (cf., SD≈15%). The Deep Neural Network Architecture proposed here constitutes a method that contributes to the early diagnosis of cognitive decline, quantify the progression of the condition, and enable suitable therapeutics.},
	journal      = {Frontiers in Neurology},
	author       = {Themistocleous, Charalambos and Eckerström, Marie and Kokkinakis, Dimitrios},
	year         = {2018},
	volume       = {9},
	pages        = {1--10},
}

@article{themistocleous-2017-effects-259668,
	title        = {Effects of Two Linguistically Proximal Varieties on the Spectral and Coarticulatory Properties of Fricatives: Evidence from Athenian Greek and Cypriot Greek},
	abstract     = {Several studies have explored the acoustic structure of fricatives, yet there has been very little acoustic research on the effects of dialects on the production of fricatives. This article investigates the effects of two linguistically proximal Modern Greek dialects, Athenian Greek and Cypriot Greek on the temporal, spectral, and coarticulatory properties of fricatives and aims to determine the acoustic properties that convey information about these two dialects. Productions of voiced and voiceless labiodental, dental, alveolar, palatal, and velar fricatives were extracted from a speaking task from typically speaking female adult speakers (25 Cypriot Greek and 20 Athenian Greek speakers). Measures were made of spectral properties, using a spectral moments analysis. The formants of the following vowel were measured and second degree polynomials of the formant contours were calculated. The findings showed that Athenian Greek and Cypriot Greek fricatives differ in all spectral properties across all places of articulation. Also, the co-articulatory effects of fricatives on following vowel were different depending on the dialect. Duration, spectral moments, and the starting frequencies of F1, F2, F3, and F4 contributed the most to the classification of dialect. These findings provide a solid evidence base for the manifestation of dialectal information in the acoustic structure of fricatives.},
	journal      = {Frontiers in Psychology},
	author       = {Themistocleous, Charalambos},
	year         = {2017},
	volume       = {8},
	number       = {1945},
	pages        = {1--19},
}

@article{grohmann-etal-2017-acquiring-252175,
	title        = {Acquiring Clitic Placement in Bilectal Settings: Interactions between Social Factors},
	abstract     = {This paper examines the development of object clitic placement by children acquiring Cypriot Greek. Greek-speaking Cyprus is sociolinguistically characterized by diglossia between two varieties of Greek, the local Cypriot Greek and the official Standard Modern Greek. Arguably as a result of this situation, clitics may be placed post- (enclisis) or preverbally (proclisis) in the same syntactic environment; while the former is a property of Cypriot Greek, the latter is typically considered an effect of the standard language. The following issues are investigated here: (a) how such bilectal speakers distinguish between the two Greek varieties with respect to clitic placement; (b) how the acquisition of clitics develops over time; (c) how, and which, sociolinguistic factors determine clitic placement; and (d) how schooling may affect clitic placement. To address (a)–(d), a sentence completion task was used to elicit clitic productions, administered to 431 children around Cyprus ranging from 2;8 to 8;11. The C5.0 machine-learning algorithm was employed to model the interaction of (socio-)linguistic factors on the development of clitic placement. The model shows that speakers acquire the relevant features very early, yet compartmentalization of form and function according to style emerges only as they engage in the larger speech community. In addition, the effects of sociolinguistic factors on clitic placement appear gradually.},
	journal      = {Frontiers in Communication},
	author       = {Grohmann, Kleanthes and Papadopoulou, Elena and Themistocleous, Charalambos},
	year         = {2017},
	volume       = {2},
}

@article{themistocleous-2017-classifying-254040,
	title        = {Classifying linguistic and dialectal information from vowel acoustic parameters},
	abstract     = {This study provides a classification model of two Modern Greek dialects, namely Athenian Greek and Cypriot Greek, using information from formant dynamics of F1, F2, F3, F4 and vowel duration. To this purpose, a large corpus of vowels from 45 speakers of Athenian Greek and Cypriot Greek was collected. The first four formant frequencies were measured at multiple time points and modelled using second degree polynomials. The measurements were employed in classification experiments, using three classifiers: Linear Discriminant Analysis, Flexible Discriminant Analysis, and C5.0. The latter outperformed the other classification models, resulting in a higher classification accuracy of the dialect. C5.0 classification shows that duration and the zeroth coefficient of F2, F3 and F4 contribute more to the classification of the dialect than the other measurements; it also shows that formant dynamics are important for the classification of dialect.},
	journal      = {Speech Communication},
	author       = {Themistocleous, Charalambos},
	year         = {2017},
	volume       = {92},
	pages        = {13--22},
}

@article{themistocleous-2017-nature-251205,
	title        = {The Nature of Phonetic Gradience across a Dialect Continuum: Evidence from Modern Greek Vowels.},
	abstract     = {This study investigates the acoustic properties of vowels in 2 Modern Greek varieties: Standard Modern Greek (SMG) and Cypriot Greek (CG). Both varieties contain in their phonetic inventories the same 5 vowels. Forty-five female speakers between 19 and 29 years old participated in this study: 20 SMG speakers and 25 CG speakers, born and raised in Athens and Nicosia, respectively. Stimuli consisted of a set of nonsense CVCV and VCV words, each containing 1 of the 5 Greek vowels in stressed and unstressed position. Gaining insights from the controlled experimental design, the study sheds light on the gradient effects of vowel variation in Modern Greek. It shows that (1) stressed vowels are more peripheral than unstressed vowels, (2) SMG unstressed /i a u/ vowels are more raised than the corresponding CG vowels, (3) SMG unstressed vowels are shorter than CG unstressed vowels, and (4) SMG /i·u/ are more rounded than the corresponding CG vowels. Moreover, it shows that variation applies to specific subsystems, as it is the unstressed vowels that vary cross-varietally whereas the stressed vowels display only minor differences. The implications of these findings with respect to vowel raising and vowel reduction are discussed.},
	journal      = {Phonetica},
	author       = {Themistocleous, Charalambos},
	year         = {2017},
	volume       = {74},
	number       = {3},
	pages        = {157--172},
}

@article{themistocleous-2016-bursts-243451,
	title        = {The bursts of stops can convey dialectal information},
	abstract     = {This study investigates the effects of the dialect of the speaker on the spectral properties of stop bursts. Forty-five female speakers—20 Standard Modern Greek and 25 Cypriot Greek speakers—participated in this study. The spectral properties of stop bursts were calculated from the burst spectra and analyzed using spectral moments. The findings show that besides linguistic information, i.e., the place of articulation and the stress, the speech signals of bursts can encode social information, i.e., the dialects. A classification model using decision trees showed that skewness and standard deviation have a major contribution for the classification of bursts across dialects.},
	journal      = {Journal of the Acoustical Society of America},
	author       = {Themistocleous, Charalambos},
	year         = {2016},
	volume       = {140},
	number       = {4},
	pages        = {EL334--EL339},
}

@inProceedings{themistocleous-etal-2016-effects-239893,
	title        = {Effects of stress on fricatives: Evidence from Standard Modern Greek},
	abstract     = {This study investigates the effects of stress on the spectral properties of fricative noise in Standard Modern Greek (SMG). Twenty female speakers of SMG participated in the study. Fricatives were produced in stressed and unstressed positions in two vowel place positions: back and front vowels. Acoustic measurements were taken and the temporal and spectral properties of fricatives using spectral moments were calculated. Stressed fricatives are produced with increased duration, center of gravity, standard deviation, and normalized intensity. The machine learning and classification algorithm C5.0 has been employed to estimate the contribution of the temporal and spectral parameters for the classification of fricatives. Overall, duration and center of gravity contribute the most to the classification of stressed vs. unstressed fricatives.},
	booktitle    = {17th Annual Conference of the International Speech Communication Association, Interspeech 2016  8-12 Sep 2016, San Francisco, USA },
	author       = {Themistocleous, Charalambos and Savva, Angelandria and Aristodemou, Andrie},
	year         = {2016},
	ISBN         = {978-1-5108-3313-5},
}

@inProceedings{themistocleous-etal-2020-automatic-305224,
	title        = {Automatic analysis of voice quality and prosody in patients with Mild Cognitive Impairment.},
	abstract     = {http://demo.spraakdata.gu.se/svedk/pbl/SNL2020.pdf},
	booktitle    = {The 12th Annual Society for the Neurobiology of Language Meeting (SNL) -- virtual conference},
	author       = {Themistocleous, Charalambos and Eckerström, Marie and Kokkinakis, Dimitrios},
	year         = {2020},
}

@inProceedings{themistocleous-etal-2020-automated-305223,
	title        = {Automated speech analysis improves MCI diagnosis},
	abstract     = {Mild Cognitive Impairment (MCI) is a condition characterized by cognitive
decline greater than expected for an individual's age and education level. In this
study, we are investigating whether acoustic properties of speech production
can improve the classification of individuals with MCI from healthy controls
augmenting the Mini Mental State Examination, a traditional screening tool,
with automatically extracted acoustic information. We found that just one
acoustic feature, can improve the AUC score (measuring a trade-off between
sensitivity and specificity) from 0.77 to 0.89 in a boosting classification task.
These preliminary results suggest that computerized language analysis can
improve the accuracy of traditional screening tools},
	booktitle    = {Proceedings of the 11th Experimental Linguistics Conference (ExLing)},
	author       = {Themistocleous, Charalambos and Eckerström, Marie and Kokkinakis, Dimitrios},
	year         = {2020},
}

@inProceedings{themistocleous-etal-2020-improving-305222,
	title        = {Improving the Diagnosis of Mild Cognitive Impairment in elderly individuals using a multifactorial automatic analysis of voice quality and prosody.},
	abstract     = {http://demo.spraakdata.gu.se/svedk/pbl/AEC-30-Paper.JPG},
	booktitle    = {30th Alzheimer Europe Conference #30AEC -- virtual conference },
	author       = {Themistocleous, Charalambos and Eckerström, Marie and Lundholm Fors, Kristina and Kokkinakis, Dimitrios},
	year         = {2020},
}

@inProceedings{themistocleous-kokkinakis-2019-speech-289021,
	title        = {Speech and Mild Cognitive Impairment detection},
	abstract     = {It is of great importance to detect objective markers that can enable the early and fast identification of individuals with Mild Cognitive Impairment (MCI) from healthy individuals to inform, patient care, family and treatment planning. Connected speech productions can offer such markers. This study analyses recordings from picture description tasks by Swedish individuals with MCI and healthy control individuals (HC) and shows that voice quality, periodicity, and speech rate distinguish individuals with MCI from HC. 
},
	booktitle    = {Proceedings of the 10th International Conference of Experimental Linguistics, 25-27 September 2019, Lisbon, Portugal},
	editor       = {Antonis Botinis},
	author       = {Themistocleous, Charalambos and Kokkinakis, Dimitrios},
	year         = {2019},
	publisher    = { ExLing Society},
	ISBN         = {978-618-84585-0-5},
}

@inProceedings{fyndanis-themistocleous-2018-morphosyntactic-271917,
	title        = {Morphosyntactic production in agrammatic aphasia: A cross-linguistic machine learning approach.},
	abstract     = {Introduction
Recent studies on agrammatic aphasia by Fyndanis et al. (2012, 2017) reported evidence against the cross-linguistic validity of unitary accounts of agrammatic morphosyntactic impairment, such as the Distributed Morphology Hypothesis (DMH) (Wang et al., 2014), the two versions of the Interpretable Features’ Impairment Hypothesis (IFIH-1: Fyndanis et al., 2012; IFIH-2: Fyndanis et al., 2018b), and the Tree Pruning Hypothesis (TPH) (Friedmann & Grodzinsky, 1997). However, some of the features/factors emphasized by the accounts above (i.e. involvement of inflectional alternations (DMH), involvement of integration processes (IFIH-1), involvement of both integration processes and inflectional alternations (IFIH-2), position of a morphosyntactic feature/category in the syntactic hierarchy (TPH)) may still play a role in agrammatic morphosyntactic production. These features may act in synergy with other factors in determining the way in which morphosyntactic production is impaired across persons with agrammatic aphasia (PWA) and across languages. Relevant factors may include language-independent and language-specific properties of morphosyntactic categories, as well as subject-specific and task/material-specific variables. The present study addresses which factors determine verb-related morphosyntactic production in PWA and what is their relative importance.

Methods		
We collapsed the datasets of the 24 Greek-, German-, and Italian-speaking PWA underlying Fyndanis et al.’s (2017) study, added the data of two more Greek-speaking PWA, and employed machine learning algorithms to analyze the data. The unified dataset consisted of data on subject-verb agreement, time reference (past reference, future reference), grammatical mood (indicative, subjunctive), and polarity (affirmatives, negatives). All items/conditions were represented as clusters of theoretically motivated features: ±involvement of integration processes, ±involvement of inflectional alternations, ±involvement of both integration processes and inflectional alternations, and low/middle/high position in the syntactic hierarchy. We included 14 subject-specific, category-specific and task/material-specific predictors: Verbal Working Memory (WM), (years of formal) Education, Age, Gender, Mean Length of Utterance in (semi)spontaneous speech (Index 1 of severity of agrammatism), Proportion of Grammatical Sentences in (semi)spontaneous speech (Index 2 of severity of agrammatism), Words per Minute in (semi)spontaneous speech (Index of fluency), Involvement of inflectional alternations, Involvement of integration processes, Involvement of both integration processes and inflectional alternations, Position of a given morphosyntactic category in the syntactic hierarchy (high, middle, low), Item Presentation mode (cross-modal, auditory), Response mode (oral, written), and Language (Greek, German, Italian). Different machine learning models were employed: Random Forest, C5.0 decision tree, RPart, and Support Vector Machine.

Results & Discussion
Random Forest model outperformed all the other models achieving the highest accuracy (0.786). As shown in Figure 1, the best predictors of accuracy on tasks tapping morphosyntactic production were the involvement of both integration processes and inflectional alternations (categories involving both integration processes and inflectional alternations were more impaired than categories involving one or neither of them), verbal WM capacity (the greater the WM capacity, the better the morphosyntactic production), and severity of agrammatism (the more severe the agrammatism, the worse the morphosyntactic production). Results are consistent with IFIH-2 (Fyndanis et al., 2018b) and studies highlighting the role of verbal WM in morphosyntactic production (e.g., Fyndanis et al., 2018a; Kok et al., 2007).},
	booktitle    = {Frontiers in Human Neuroscience. Academy of Aphasia 56th Annual Meeting, Montreal, Canada, 21 Oct - 23 Oct, 2018. },
	author       = {Fyndanis, Valantis and Themistocleous, Charalambos},
	year         = {2018},
}

@inProceedings{kokkinakis-etal-2019-multifaceted-278217,
	title        = {A Multifaceted Corpus for the Study of Cognitive Decline in a Swedish Population},
	abstract     = {A potential, early-stage diagnostic marker for neurodegenerative diseases, such as Alzheimer’s disease, is the onset of language disturbances which is often characterized by subtle word-finding difficulties, impaired spontaneous speech, slight speech hesitancy, object naming difficulties and phonemic errors. Connected speech provides valuable information in a non-invasive and easy-to-assess way for determining aspects of the severity of language impairment. Data elicitation is an established method of obtaining highly constrained samples of connected speech that allows us to study the intricate interactions between various linguistic levels and cognition. In the paper, we describe the collection and content of a corpus consisting of spontaneous Swedish speech from individuals with Mild Cognitive Impairment (MCI), with Subjective Cognitive Impairment SCI) and healthy, age-matched controls (HC). The subjects were pooled across homogeneous subgroups for age and education, a sub-cohort from the Gothenburg-MCI study. The corpus consists of high quality audio recordings (including transcriptions) of several tasks, namely:
(i)	a picture description task – the Cookie-theft picture, an ecologically valid approximation to spontaneous discourse that has been widely used to elicitate speech from speakers with different types of language and communication disorders; 
(ii)	a read aloud task (including registration of eye movements) – where participants read a text from the IREST collection twice, both on a computer screen (while eye movements are registered), and the same text on paper;
(iii)	a complex planning task – a subset of executive functioning that tests the ability to identify, organize and carry out (complex) steps and elements that are required to achieve a goal;
(iv)	a map task – a spontaneous speech production/semi-structured conversation in which the participants are encouraged to talk about a predefined, cooperative task-oriented topic;
(v)	a semantic verbal fluency task – category animals: where participants have to produce as many words as possible from a category in a given time (60 seconds). The fluency tests require an elaborate retrieval of words from conceptual (semantic) and lexical (phonetic) memory involving specific areas of the brain in a restricted timeframe. 
All samples are produced by Swedish speakers after obtaining written consent approved by the local ethics committee. Tasks (i) and (ii) have been collected twice in a diachronically apart period of 18 months between 2016 and 2018.
The corpus represents an approximation to speech in a natural setting: The material for elicitation is controlled in the sense that the speakers are given specific tasks to talk about, and they do so in front of a microphone. The corpus may serve as a basis for many linguistic and/or speech technological investigations and has being already used for various investigations of language features.},
	booktitle    = {CLARe4 : Corpora for Language and Aging Research, 27 February – 1 March 2019, Helsinki, Finland},
	author       = {Kokkinakis, Dimitrios and Lundholm Fors, Kristina and Fraser, Kathleen and Eckerström, Marie and Horn, Greta and Themistocleous, Charalambos},
	year         = {2019},
}

@inProceedings{fraser-etal-2018-improving-264397,
	title        = {Improving the Sensitivity and Specificity of MCI Screening with Linguistic Information.},
	abstract     = {The Mini-Mental State Exam (MMSE) is a screening tool for cognitive impairment. It has been extensively validated and is widely used, but has been criticized as not being effective in detecting mild cognitive impairment (MCI). In this study, we examine the utility of augmenting MMSE scores with automatically extracted linguistic information from a narrative speech task to better differentiate between individuals with MCI and healthy controls in a Swedish population. We find that with the addition of just four linguistic features, the F score (measuring a trade-off between sensitivity and specificity) is improved from 0.67 to 0.81 in logistic regression classification. These preliminary results suggest that the accuracy of traditional screening tools may be improved through the addition of computerized language analysis.},
	booktitle    = {Proceedings of the LREC workshop: Resources and ProcessIng of linguistic, para-linguistic and extra-linguistic Data from people with various forms of cognitive/psychiatric impairments (RaPID-2). 8th  of May 2018,  Miyazaki, Japan  / Dimitrios  Kokkinakis  (ed.)},
	author       = {Fraser, Kathleen and Lundholm Fors, Kristina and Eckerström, Marie and Themistocleous, Charalambos and Kokkinakis, Dimitrios},
	year         = {2018},
	ISBN         = {979-10-95546-26-9},
}

@inProceedings{themistocleous-etal-2018-acoustic-271915,
	title        = {Acoustic markers of PPA variants using machine learning.},
	abstract     = {Introduction. Speakers’ acoustic profile carries significant linguistic and non-linguistic information. Employed in clinical practice, it can provide behavioral markers for a quick assessment of primary progressive aphasia (PPA). PPA is a complex language syndrome where different speech and language properties such as prosody, lexical retrieval, and motor speech functioning may be affected.  It is classified into three main variants: the nonfluent (nfvPPA), semantic (svPPA), and logopenic (lvPPA). Primary progressive apraxia of speech (PPAOS) is also distinguished (Duffy et al. 2017) but may fall into the category of nfvPPA (Gorno-Tempini et al. 2011). The present study aims to determine the contribution of the acoustic properties of vowels, prosody, and voice quality in the classification of PPA variants by using machine learning models.

Methods. Oral samples from picture description tasks of 50 individuals with PPA (lvPPA:17, svPPA:14, nfvPPA:11, PPAOS:8) were automatically transcribed and segmented into vowels and consonants using the new acoustic analysis platform THEMIS. From the segmented vowels, we measured: i. Vowel formants (F1…F5) (den Ouden, et al. 2017); ii. vowel duration (Duffy, et al., 2017); iii. Mean fundamental frequency (F0), min F0 and max F0 (Hillis, 2014); iv. Pause duration (Mack et al. 2015), and v. H1–H2, H1–A1, H1–A2, H1–A3 measures of voice quality. We compared three machine learning models: support vector machines (SVM) (Cortes and Vapnik, 1995), random forests (RF) (Breiman, 2001), and decision trees (DT) (Hastie et al. 2009) in an one-against all strategy, where each variant was tested against all others. We run all models with a 3-fold group-cross-validation to ensure that the speakers in the training and evaluation sets are different. The models were implemented in Python (Pedregosa et al. 2011).

Results. We report the mean cross-validated accuracy of the best performing model that resulted from model comparison: i. RF model provided the highest classification accuracy for nfvPPA [Mean 82%, SD: 9%], ii. SVM had the highest accuracy for svPPA [Mean 66%, SD: 8%], iii. RF had the highest accuracy for lvPPA [Mean 57%, SD: 15%] and iv. RF provided the highest classification accuracy for PPAOS [Mean 80%, SD: 8%] (Figure 1). In all models, pause duration and F0 measures were ranked higher than most other features (Figure 2).

Discussion. This study employed an innovative method for the classification of PPA variants, using an automated speech transcription, segmentation, feature extraction and modeling. Using just acoustic features the best model classified nfvPP, svPPA, and PPAOS with high accuracy. However, acoustic features alone could not classify lvPPA with such high accuracy. More linguistic markers might be needed for a more accurate classification of lvPPA. Furthermore, we showed that prosody, which is measured by fundamental frequency and pause duration, contributes more than any other factor to the classification of PPA variants as alluded in previous research by our group and others (Hillis 2014, Patel et al. 2018, Mack 2015). Finally, the findings demonstrate the potential benefit of using machine learning models in clinical practice for the subtyping of PPA variants.},
	booktitle    = {Frontiers in Human Neuroscience. Conference Abstract: Academy of Aphasia 56th Annual Meeting, October 21-23, 2018, Montreal, Canada},
	author       = {Themistocleous, Charalambos and Ficek, Bronte and Webster , Kimberly and Wendt, Haley  and Hillis, Argye and  Den Ouden , Dirk-Bart  and Tsapkini, Kyrana},
	year         = {2018},
}

@inProceedings{neofytou-etal-2018-understanding-271916,
	title        = {Understanding and classifying the different variants of Primary Progressive Aphasia based on spelling performance},
	abstract     = {Introduction: Previous findings suggest differences in the written spelling performance between the three variants of Primary Progressive Aphasia (PPA) - semantic (svPPA), logopenic (lvPPA) and non-fluent (nfvPPA) (Shim et al., 2012; Sepelyak et al., 2011). However, no attempts have been made to systematically distinguish the three variants in terms of their spelling performance. The challenges of classification are considerable and given the ease of administering a spelling test, we aimed to determine to what extent a spelling task can provide accurate classification of the  PPA variants.
Method: Thirty-three participants with PPA were included - 14 lvPPAs, 11 nfvPPAs and 8 svPPAs – originally classified using the neuropsychological and spoken language criteria defined by Gorno-Tempini et al. (2011). Data were collected prior to spelling treatment, using a spelling to dictation task with both real-words and pseudowords (92-138 items/per participant), scored for each grapheme (i.e., letter) and analyzed for each participant individually using generalized linear mixed effects models (GLMEM) for real-words and pseudowords separately. The variables of interest for both real-words and pseudowords were word length, phoneme-grapheme conversion probability and grapheme position. The real-word models also included frequency, imageability, and the orthographic and phonological neighborhood density of the target words. 
The coefficients from the output of the GLMEMs, together with 3 additional variables – verb/noun  and pseudoword/word accuracy differences from the spelling task, and language impairment severity according to FTD-CDR (Knopman, 2008) - were used as predictors in a Random Forests (RFs) model implemented in Python, to identify the variables that contribute the most in distinguishing the three variants. Then, the three most significant predictors identified with RFs were used in multinomial models implemented in R to classify the PPA variants. The model was trained on a training set of all participants minus one (i.e. the left-out participant) and evaluated on the left-out participant, known as Leave-One-Out cross-validation. This process was repeated 33 times to evaluate all participants. 
Results: The three most significant predictors of the RFs analysis were: (1) grapheme position in real-words, (2) pseudoword/word accuracy difference, and (3) length of real-words (Figure 1). The overall accuracy of the multinomial models with these three predictors only was 67%: lvPPA=71%, nfvPPA=64% and svPPA=63%. When severely impaired cases (language severity =3 in Knopman et al., 2008; FTD-CDR criteria) were excluded (giving a new dataset of 22 participants), the overall accuracy increased to 91%: lvPPA=90%, nfvPPA=86% and svPPA=100%.
Discussion: Our study provides evidence of the value of considering spelling performance in understanding and classifying the different variants of PPA. The results suggest that lexical status, word length and grapheme position are useful parameters for classification, which index key components of the cognitive architecture of spelling (Rapp, 2002). Also, the finding that prediction accuracy increased when more severe cases were excluded supports  previous findings (Mesulam et al., 2012), as severity increases variants become less differentiated and classification is more difficult. In sum, a relatively short, easy-to-administer spelling test, provides useful information for PPA variant classification and can  potentially be used as a clinical tool.},
	booktitle    = {Frontiers in Human Neuroscience},
	author       = {Neofytou, Kyriaci and Themistocleous, Charalambos and Wiley,  Robert  and Tsapkini, Kyrana and Rapp, Brenta},
	year         = {2018},
}

@inProceedings{themistocleous-etal-2018-effects-270215,
	title        = {Effects of Mild Cognitive Impairment on vowel duration
},
	abstract     = {Mild cognitive impairment (MCI) is a neurological condition, which is characterized
by a noticeable decline of cognitive abilities, including communicative and linguistic
skills. In this study, we have measured the duration of vowels produced in a reading
task by 55 speakers— 30 healthy controls and 25 MCI—. The main results showed
that MCI speakers differed significantly from HC in vowel duration as MCI speakers
produced overall longer vowels. Also, we found that gender effects on vowel duration
were different in MCI and HC. One significant aspect of this finding is that they highlight
the contribution of vowel acoustic features as markers of MCI.},
	booktitle    = {Proceedings of the 9th Tutorial & Research Workshop on Experimental Linguistics, 28 - 30 August 2018, Paris, France},
	editor       = {Antonis Botinis},
	author       = {Themistocleous, Charalambos and Kokkinakis, Dimitrios and Eckerström, Marie and Fraser, Kathleen and Lundholm Fors, Kristina},
	year         = {2018},
	ISBN         = {978-960-466-162-6 },
}

@inProceedings{themistocleous-kokkinakis-2018-themis-265112,
	title        = {THEMIS-SV: Automatic classification of language disorders from speech signals},
	abstract     = {Background and Aims:
Brain injuries resulting from stroke can affect the production of speech resulting in different types of language impairments, such as aphasia.
Studying these productions manually is an extremely cumbersome and time consuming process. 
The aim of this paper is to present THEMIS-SV: a system that enables the automatic transcription of speech signals and the segmentation of vowels and consonants in Swedish.

Method:
The input of the system are recordings of speech. The system processes the recordings and returns an output with three tiers: the utterance tier, the word tier, and the vowels/consonants tier. 

Results:
The output of the system is a fast and reliable transcription and segmentation of speech, which is very close to transcriptions and segmentations performed manually. 
The automatic segmentation of speech enables targeted acoustic measurements, such as measurements of consonant spectra, formant frequencies of vowels, fundamental frequency, pauses, speech rate, etc. and other acoustic measurements that have been known to differentiate between the different types of language disorders.

Conclusion:
The method proposed here can be employed for the analysis of speech of individuals with post-stroke aphasia and other language disorders and constitutes a promising step towards a fully automated differential diagnostic tool for language disorders. },
	booktitle    = {Abstracts of the 4th European Stroke Organisation Conference ​(ESOC 2018). Gothenburg, Sweden, 16-18 May, 2018. },
	author       = {Themistocleous, Charalambos and Kokkinakis, Dimitrios},
	year         = {2018},
}

@inProceedings{kokkinakis-etal-2018-textforskning-265113,
	title        = {Kan textforskning bidra till tidigare och säkrare demensdiagnostik?},
	abstract     = {Tidigare forskning har visat att subtila språkstörningar kan finnas vid de tidigaste förstadierna till demens, flera år innan en klinisk diagnos kan ställas. Inom ramen för projektet ”Språkliga och extra-lingvistiska parametrar för tidig upptäckt av kognitiv svikt” (finansierat av Riksbankens Jubileumsutlysning, 2016-19) undersöker vi med hjälp av språkteknologi och språkanalysstudier hur dessa språkstörningar yttrar sig. Kan språkteknologi användas för att upptäcka dessa tidiga språkrelaterade symtom och därmed bidra med nyanserad, komplementär och användbar kunskap? Kan användning av språkteknologi särskilja personer med de allra tidigaste kognitiva avvikelserna från personer med mer godartad, åldersrelaterad kognitiv svikt? Vilka språkliga förmågor drabbas? Hur yttrar sig dessa förändringar och vilka slags empiriska material finns att tillgå? Dessa är några av de frågor vi söker svar på. Vi gör inspelningar som vi analyserar för att kunna ta fram ny kunskap om subtila språkliga kännetecken som kan föregå demensutveckling. Denna kunskap kan användas för att eventuellt kunna förutsäga vilka individer som befinner sig i riskzonen för att utveckla demens, och kan vara användbar som komplementerande beslutsunderlag till domänexperter. Vi utvinner, analyserar och undersöker om det finns samband mellan olika språkrelaterade parametrar från spontan talinteraktion, transkriptioner men även ögonrörelser och neuropsykologiska tester från personer med subjektiv eller lindrig kognitiv nedsättning och friska kontrollpersoner. Många gånger är det svårt att avgöra huruvida lindriga kognitiva symtom är en del av det normala åldrandet eller början på en neurodegenerativ process. Vi förväntar oss inte heller att varje enskild person med kognitiv nedsättning kommer att uttrycka sig eller läsa på samma sätt utan snarare att dessa personer tidigt i sjukdomsförloppet kommer att börja uppvisa olika slags avvikande läsmönster, eller göra fonologiska, lexikala, syntaktiska eller semantiska fel. I studien utvecklar vi verktyg för att automatiskt hitta dessa avvikelser, och målet är att detta sedan ska kunna användas som komplement till tidig diagnostik samt som prognostiskt eller screeningverktyg. Deltagarna i vår studie har rekryterats från en pågående longitudinell studie, ”Demens i Tidigt Skede”, (eng. ”The Gothenburg MCI study”) på Minnesmottagningen i Göteborg, och vårt projekt har godkänts av den lokala etiknämnden. Alla deltagare i studien (kontrollgruppen [HC], personer med subjektiv kognitiv nedsättning [SCI] och personer med mild kognitiv nedsättning [MCI]) har genomgått baslinjeundersökning och gett informerat skriftligt samtycke (demografisk information finns i tabell 1). Vårt projekt är f.n. pågående och vi kommer presentera resultat baserade på inspelningstillfälle nr ett (aug. 2016-mars 2017). En ny inspelningsomgång, med samma deltagare, började i februari 2018 och förväntas vara avslutat i december 2018. Under presentationen kommer vi ge exempel på olika tal-, text- och ögonrörelseanalyser vi har genomfört och diskutera metodval och resultat baserade på studiens första fas. Vi kommer vidare ge en kort inblick i den nya, pågående inspelningsomgången och de nya testmoment vi använder. Vi vill med vårt arbete visa hur språkteknologisk analys kan bidra till att utöka vår kunskap inom området så att den kan vara användbar för tidig diagnostik och optimal omvårdnad. Enligt Socialstyrelsen (2017) finns det i Sverige över 160 000 personer med någon demenssjukdom. Våra resultat kan ha en betydelse för vårdpersonal som snabbare vill diagnostisera och identifiera individer med olika former av kognitiv funktionsnedsättning innan allvarliga symtom blir påtagliga. Utvecklingsmöjligheterna är många: nya eller förbättrade kognitiva screeningtester som skulle kunna användas inom primär- och specialistvården, samt utveckling och tillämpning av insatser som kan påverka beteendemönster och träna upp individens kommunikativa förmåga, kan på sikt leda till positiva konsekvenser som minskade vårdköer samt effektivare behandling avseende kostnader och behandlingsutfall.},
	booktitle    = {Forum för textforskning 13 , Lund 7 – 8 juni 2018},
	author       = {Kokkinakis, Dimitrios and Lundholm Fors, Kristina and Eckerström, Marie and Themistocleous, Charalambos},
	year         = {2018},
}

@inProceedings{themistocleous-etal-2018-classification-268340,
	title        = {A classification study of the variants of Primary Progressive Aphasia using Machine Learning.},
	abstract     = {Introduction: Primary Progressive Aphasia (PPA) is a neurodegenerative syndrome in which linguistic abilities become gradually impaired. There are three primary variants of PPA: the non-fluent agrammatic PPA, the fluent type semantic PPA, and the logopenic PPA, which is also considered an atypical form of Alzheimer’s disease (Mesulam et al., 1982; Gorno-Tempini et al., 2011). Along with the three main variants, a fourth variant has been proposed, a non-fluent apraxia of speech (AOS), though this is currently the subject of an open debate (e.g., Duffy et al., 2017; Henry et al., 2013). According to sophisticated criteria established a few years ago, PPA subtyping for a given patient presented in clinic requires clinical, neuropsychological, and imaging information (Gorno-Tempini et al., 2011). Nevertheless, quantifying the decline of linguistic abilities and subtyping its variants manually is both hard and laborious, so there is a great demand for algorithms that subtype a given patient automatically. Picture description samples of connected speech and random forests techniques have been used for this purpose (de Aguiar et al., 2017; Wilson et al., 2010, Fraser et al. 2013, 2014). In the present study, we compared existing models and we propose a new one.
Aims: In this study, we provide an automated classification model of the four PPA variants trained on known morphological and acoustic predictors and on predictors related to the clinical and linguistic profile of individuals with PPA (e.g., Mack et al., 2015; Gorno-Tempini et al., 2011; Wilson et al., 2010). 
Method: Speech materials for this study come from the Transcranial Direct Current Stimulation for Primary Progressive Aphasia study at Johns Hopkins University. Twenty-six individuals with PPA (Mean(SD) age = 68.6 (7.8) years, Mean(SD) education = 16.1 (2.9) years) participated in this study. PPA participants were diagnosed based on the established consensus criteria (Gorno-Tempini et al., 2011) based on imaging, clinical, and neuropsychological examination by trained neurologists. Individuals with PPA included non-fluent AOS (N=5), non fluent (N=7), logopenic (N=8), and semantic (N=6) variants. 
Recordings of the Cookie Theft picture description from the Boston Diagnostic Aphasia Examination (BDAE) were computationally analyzed. All speech productions were automatically transcribed and segmented using an end-to-end speech-to-transcription platform. From the speech signals, we measured morphological and acoustic predictors, including vowel formants F1 ... F3, measured at 15%, 50%, and 75% of vowel’s duration, vowel duration, fundamental frequency, and pause duration. The analysis and the statistics were conducted using Python and R programming languages (R Core Team, 2017; Rossum, 1995). 
Three different machine learning algorithms: C5.0 decision trees, Classification and Regression Trees (CART) and random forests were trained on the predictors (Breiman, 2001; Quinlan, 1993; Hastie et al., 2009). All models were trained on the 80% of the speakers (training set), with 3-fold cross-validation. All predictor variables were centered and scaled. C5.0 was trained with winnowing and without winnowing. (Winnowing facilitates the automatic pre-selection of the predictors that are used in the decision tree.) After the training we evaluated the trained models on the unknown dataset, namely the 20% of the speakers (evaluation set).
Results: C5.0 provided 86% (95% CI[81, 88], kappa = 0.76) and Random Forests 85% (95% CI[81, 88], kappa = 0.76) classification accuracy on the test data; CART provided the lowest overall classification accuracy. Overall, C5.0 outperformed both the random forests and CART, with high classification accuracy on unknown data. Non-fluent AOS was correctly predicted by both C5.0 and random forests.
Discussion: C5.0 classification model provides support for the known predictors employed in the literature. Also, it provides initial support for the distinct properties of the non-fluent AOS variant and corroborate research on classification of AOS using acoustic properties especially those related to vowel production (Den Ouden et al. 2017). However, given the low number of participants employed in this study, further research is required, with a largest number of participants. Nevertheless, the proposed methods employed here constitute a promising step towards a computational differential diagnostic tool of PPA that is easy to use, quick and accurate. },
	booktitle    = {Clinical Aphasiology Conference, CAC 2018, Austin, Texas USA},
	author       = {Themistocleous, Charalambos and Ficek, Bronte and Webster, Kimberly and Wendt, Haley and Hillis, Argye E. and den Ouden , Dirk Bart and Tsapkini, Kyrana},
	year         = {2018},
}

@inProceedings{denouden-etal-2018-comparison-268339,
	title        = {Comparison of Automated Methods for Vowel Segmentation and Extraction of Acoustic Variables},
	abstract     = {Introduction: Primary Progressive Aphasia (PPA) is a neurodegenerative syndrome in which linguistic abilities become gradually impaired. There are three primary variants of PPA: the non-fluent agrammatic PPA, the fluent type semantic PPA, and the logopenic PPA, which is also considered an atypical form of Alzheimer’s disease (Mesulam et al., 1982; Gorno-Tempini et al., 2011). Along with the three main variants, a fourth variant has been proposed, a non-fluent apraxia of speech (AOS), though this is currently the subject of an open debate (e.g., Duffy et al., 2017; Henry et al., 2013). According to sophisticated criteria established a few years ago, PPA subtyping for a given patient presented in clinic requires clinical, neuropsychological, and imaging information (Gorno-Tempini et al., 2011). Nevertheless, quantifying the decline of linguistic abilities and subtyping the variants of PPA manually is both hard and laborious, so there is great demand for algorithms that subtype a given patient automatically. Picture description samples of connected speech and random forests techniques have been used for this purpose (de Aguiar et al., 2017; Wilson et al., 2010, Fraser et al. 2013, 2014). In the present study, we compared existing models and we propose a new one.
Aims: In this study, we provide an automated classification model of PPA variants trained on known morphological and acoustic predictors and on predictors related to the clinical and linguistic profile of individuals with PPA (e.g., Mack et al., 2015; Gorno-Tempini et al., 2011; Wilson et al., 2010). 
Method: Speech materials for this study come from the Transcranial Direct Current Stimulation for Primary Progressive Aphasia study at Johns Hopkins University. Twenty-six individuals with PPA (Mean(SD) age = 68.6 (7.8) years, Mean(SD) education = 16.1 (2.9) years) participated in this study. PPA participants were diagnosed based on the established consensus criteria (Gorno-Tempini et al., 2011), i.e., imaging, clinical, and neuropsychological examination by trained neurologists. Individuals with PPA included non-fluent with AOS (N=5), non fluent without AOS (N=7), logopenic (N=8), and semantic (N=6) variants. 
Recordings of the Cookie Theft picture description from the Boston Diagnostic Aphasia Examination (BDAE) were computationally analyzed. All speech productions were automatically transcribed and segmented using an end-to-end speech-to-transcription platform. From the speech signals, we measured morphological and acoustic predictors, including vowel formants F1 ... F3, measured at 15%, 50%, and 75% of vowel’s duration, vowel duration, fundamental frequency, and pause duration. The analysis and the statistics were conducted using Python and R programming languages (R Core Team, 2017; Rossum, 1995). 
Three different machine learning algorithms: C5.0 decision trees, Classification and Regression Trees (CART) and random forests were trained on the predictors (Breiman, 2001; Quinlan, 1993; Hastie et al., 2009). All models were trained on the 80% of the speakers (training set), with 3-fold cross-validation. All predictor variables were centered and scaled. C5.0 was trained with winnowing and without winnowing. (Winnowing facilitates the automatic pre-selection of the predictors that are used in the decision tree.) After the training we evaluated the trained models on the unknown dataset, namely the 20% of the speakers (evaluation set).
Results: C5.0 provided 86% (95% CI[81, 88], kappa = 0.76) and Random Forests 85% (95% CI[81, 88], kappa = 0.76) classification accuracy on the test data; CART provided the lowest overall classification accuracy. Overall, C5.0 outperformed both the random forests and CART, with high classification accuracy on unknown data. Non-fluent PPA with AOS was correctly predicted by both C5.0 and random forests.
Discussion: The C5.0 classification model provides support for the known predictors employed in the literature. Also, it provides some objective ways to distinguish the presence of AOS in PPA and corroborate research on classification of AOS using acoustic properties especially those related to vowel production (Den Ouden et al. 2017). However, given the low number of participants employed in this study, further research is required, with a larger number of participants. Nevertheless, the proposed methods employed here constitute a promising step towards a computational differential diagnostic tool of PPA that is easy to use, quick and accurate. },
	booktitle    = {Clinical Aphasiology Conference, CAC 2018, Austin, Texas USA.},
	author       = {den Ouden, Dirk B. and Hutchinson, Angelica and Tsapkini, Kyrana and Themistocleous, Charalambos},
	year         = {2018},
}

@inProceedings{angelopoulou-etal-2018-pause-268338,
	title        = {Pause patterns and speech errors in stroke patients with aphasia: cross-linguistic evidence from narrative speech.},
	booktitle    = {Clinical Aphasiology Conference, CAC 2018, Austin, Texas USA.},
	author       = {Angelopoulou, Georgia and Kiran, Swathi and  Kasselimis, Dimitrios and Varkanitsa, Maria and Meier, Erin and Yue, Pan and Tsolakopoulos, Dimitrios and Themistocleous, Charalambos and Vassilopoulou, Sofia and Korompoki , Eleni and Tountopoulou, Argyro and Papageorgiou, Georgios and Goutsos, Dionysis, and Evdokimidis, Ioannis and Potagas, Constantin},
	year         = {2018},
}

@inProceedings{fyndanis-etal-2017-time-260585,
	title        = {Time reference and aspect in agrammatic aphasia: Evidence from Greek},
	abstract     = {Time reference, which has been found to be selectively impaired in agrammatic aphasia (e.g., Bastiaanse et al., 2011), is often interwoven with grammatical aspect. Dragoy and Bastiaanse (2013) investigated the relationship between time reference/tense and aspect focusing on Russian aphasia and found that the two interact: past reference was less impaired when tested within perfective aspect (compared to when tested within imperfective aspect), and reference to the nonpast was less impaired when tested within imperfective aspect (compared to when tested within perfective aspect). To account for this pattern, Dragoy and Bastiaanse (2013: 114) claimed that “perfectives primarily refer to completed, past events while imperfectives prototypically describe ongoing, non-past events”.
This study explores the relationship between time reference and aspect focusing on Greek aphasia. In Greek, verb forms referring to the past and future encode the perfective-imperfective contrast. Dragoy and Bastiaanse (2013) would make predictions PR1–PR4 for Greek. 
(PR1) past reference within perfective aspect > past reference within imperfective aspect;
(PR2) future reference within perfective aspect < future reference within imperfective aspect;
(PR3) perfective aspect within past reference > imperfective aspect within past reference;
(PR4) perfective aspect within future reference < imperfective aspect within future reference.

Methods
Eight Greek-speaking persons with agrammatic aphasia (PWA) and eight controls were administered a sentence completion task consisting of 128 experimental source sentence (SS)-target sentence (TS) pairs. There were eight subconditions, each of which consisted of 16 items: past reference within perfective aspect; past reference within imperfective aspect; future reference within perfective aspect; future reference within imperfective aspect; perfective aspect within past reference; imperfective aspect within past reference; perfective aspect within future reference; imperfective aspect within future reference. Participants were auditorily presented with a SS and the beginning of the TS, and were asked to orally complete the TS producing the missing Verb Phrase. We fitted generalized linear mixed-effect models and employed Fisher’s exact tests to make within-participant comparisons.

Results
Overall, the aphasic group fared significantly worse than the control group (p < 0.001). At the group level, none of the four relevant comparisons (see PR1–PR4) yielded significant differences for PWA (Table 1). Four PWA (P1, P3, P7, P8) exhibited dissociations, with three of them making up a double dissociation: P1 performed better on imperfective aspect-future reference than on perfective aspect-future reference (p < 0.001), and P7 and P8 exhibited the opposite pattern (p = 0.016 and p < 0.001 for P7 and P8, respectively). 

Discussion
Results are not consistent with Dragoy and Bastiaanse’s (2013) findings, which challenges the idea of prototypical and non-prototypical associations between time reference and aspect. The double dissociation that emerged in the aspect condition indicates that a given time reference-aspect combination may be relatively easy to process for some PWA but demanding for some others. Thus, studies investigating tense/time reference in aphasia should ensure that this grammatical/semantic category is not confounded by aspect.
},
	booktitle    = { Front. Hum. Neurosci. Conference Abstract: Academy of Aphasia, 55th Annual Meeting, Baltimore, United States, 5 Nov - 7 Nov, 2017. },
	author       = {Fyndanis, Valantis and Themistocleous, Charalambos and Christidou, Paraskevi},
	year         = {2017},
}

@inProceedings{bernardy-themistocleous-2017-modelling-258661,
	title        = {Modelling prosodic structure using Artificial Neural Networks},
	abstract     = {The ability to accurately perceive whether a speaker is asking a question or is making a statement is crucial for any successful interaction. However, learning and classifying tonal patterns has been a challenging task for automatic speech recognition and for models of tonal representation, as tonal contours are characterized by significant variation. This paper provides a classification model of Cypriot Greek questions and statements. We evaluate two state-of-the-art network architectures: a Long Short-Term Memory (LSTM) network and a convolutional network (ConvNet). The ConvNet outperforms the LSTM in the classification task and exhibited an excellent performance with 95% classification accuracy.},
	booktitle    = {ExLing 2017. Proceedings of 8 th Tutorial and Research Workshop on Experimental Linguistics, 19-22 June 2017, Heraklion, Crete, Greece},
	editor       = {Antonis Botinis},
	author       = {Bernardy, Jean-Philippe and Themistocleous, Charalambos},
	year         = {2017},
	publisher    = {University of Athens},
	address      = {Athens},
	ISBN         = {978-960-466-162-6},
}

@article{themistocleous-2016-seeking-239901,
	title        = {Seeking an Anchorage. Stability and Variability in Tonal Alignment of Rising Prenuclear Pitch Accents in Cypriot Greek},
	abstract     = {Although tonal alignment constitutes a quintessential property of pitch accents, its exact characteristics remain unclear. This study, by exploring the timing of the Cypriot Greek L*+H prenuclear pitch accent, examines the predictions of three hypotheses about tonal alignment: the invariance hypothesis, the segmental anchoring hypothesis, and the segmental anchorage hypothesis. The study reports on two experiments: the first of which manipulates the syllable patterns of the stressed syllable, and the second of which modifies the distance of the L*+H from the following pitch accent. The findings on the alignment of the low tone (L) are illustrative of the segmental anchoring hypothesis predictions: the L persistently aligns inside the onset consonant, a few milliseconds before the stressed vowel. However, the findings on the alignment of the high tone (H) are both intriguing and unexpected: the alignment of the H depends on the number of unstressed syllables that follow the prenuclear pitch accent. The ‘wandering’ of the H over multiple syllables is extremely rare among languages, and casts doubt on the invariance hypothesis and the segmental anchoring hypothesis, as well as indicating the need for a modified version of the segmental anchorage hypothesis. To address the alignment of the H, we suggest that it aligns within a segmental anchorage–the area that follows the prenuclear pitch accent–in such a way as to protect the paradigmatic contrast between the L*+H prenuclear pitch accent and the L+H* nuclear pitch accent.},
	journal      = {Language and Speech},
	author       = {Themistocleous, Charalambos},
	year         = {2016},
	volume       = {59},
	number       = {4},
	pages        = {433--461},
}

@incollection{armosti-etal-2014-addressing-232415,
	title        = {Addressing writing system issues in dialectal lexicography: the case of Cypriot Greek},
	booktitle    = {Dialogue on Dialect Standardization / editor(s): Carrie Dyck, Tania Granadillo, Keren Rice, Jorge Emilio Rosés Labrada },
	author       = {Armosti, Spyros and Christodoulou, Kyriaci and Katsoyannou, Marianna and Themistocleous, Charalambos},
	year         = {2014},
	publisher    = {Cambridge Scholars Publishing},
	address      = {Newcastle},
	ISBN         = {978-1-4438-6661-3},
	pages        = {23--38},
}

@inProceedings{aristodemou-etal-2015-acoustics-239890,
	title        = {The Acoustics of Cypriot Greek Fricatives},
	booktitle    = {Proceedings of the 6th ISEL Conference on Experimental Linguistics ExLing 2015 26 - 27 June 2015 Athens, Greece Edited by Antonis Botinis    },
	author       = {Aristodemou, Andrie and Savva, Angelandria and Themistocleous, Charalambos},
	year         = {2015},
	publisher    = {University of Athens},
	address      = {Athens},
	pages        = {9--12},
}

@incollection{tsakmakis-themistocleous-2013-textual-239887,
	title        = {Textual structure and modality in Thucydides’ military exhortations},
	booktitle    = {Thucydides Between History and Literature  Ed. by Tsakmakis, Antonis / Tamiolaki, Melina},
	author       = {Tsakmakis, Antonis and Themistocleous, Charalambos},
	year         = {2013},
	publisher    = {de Gruyter},
	address      = {Berlin},
	ISBN         = {9783110297751},
	pages        = {391--408},
}

@inProceedings{themistocleous-etal-2012-cypriot-239875,
	title        = {Cypriot Greek Lexicography: A Reverse Dictionary of Cypriot Greek},
	booktitle    = {Proceedings of the 15th EURALEX international Congress 7 – 11 August, 2012 Oslo. Edited by Ruth Vatvedt Fjeld and Julie Matilde Torjusen},
	author       = {Themistocleous, Charalambos and Katsoyannou, Marianna and Armosti, Spyros and Christodoulou, Kyriaki},
	year         = {2012},
	ISBN         = {9788230320952},
	pages        = {262--266},
}

@article{themistocleous-2014-modern-239904,
	title        = {Modern Greek Prosody. Using speech melody in communication (Prosodia tis Neas Ellinikis. I axiopoiisi tis melo- dias tis fonis stin epikoinonia)},
	journal      = {Stasinos},
	author       = {Themistocleous, Charalambos},
	year         = {2014},
	volume       = {6},
	pages        = {319--344},
}

@edited_book{fragaki-etal-2012-current-239877,
	title        = {Current Trends in Greek Linguistics},
	editor       = {Fragaki, Georgia and Georgakopoulos, Thanasis and Themistocleous, Charalambos},
	year         = {2012},
	publisher    = {Cambridge Scholars Publishing},
	address      = {Newcastle upon Tyne},
	ISBN         = {1-4438-4025-4},
}

@incollection{melissaropoulou-etal-2013-present-239888,
	title        = {The Present Perfect in Cypriot Greek revisited},
	booktitle    = {Language Variation – European Perspectives IV},
	editor       = {Peter Auer and Javier Caro Reina and Göz Kaufmann},
	author       = {Melissaropoulou, Dimitra and Themistocleous, Charalambos and Tsiplakou, Stavroula and Tsolakidis, Symeon},
	year         = {2013},
	publisher    = {John Benjamin's},
	address      = {Amsterdam / Philadelphia},
	ISBN         = {9789027234940},
	pages        = {159 – 172},
}

@article{themistocleous-2014-edge-231105,
	title        = {Edge-Tone Effects and Prosodic Domain Effects on Final Lengthening},
	journal      = {Linguistic Variation },
	author       = {Themistocleous, Charalambos},
	year         = {2014},
	volume       = {14},
	number       = {1},
	pages        = {129–160},
}

@inProceedings{botinis-etal-2007-multifactor-239903,
	title        = {Multifactor Analysis of Discourse Turn in Greek},
	abstract     = {The present article reports on an experimental study of turn structures in telephone conversations during Greek news broadcasts. Discourse segmentation was carried out based on turn constructional units (TCUs). Turn-taking and turn- leaving alternations of TCUs were analyzed in terms of speaker’s prosodic characteristics, syntactic structures and lexical discourse markers. The results indicate that the speaker's TCU tonal onset and TCU tonal offset along with global tonal variations, as well as word order are discourse correlates of turn-taking and turn-leaving.},
	booktitle    = {International Congress of Phonetic Sciences 2007 (ICPhS 2007). Saarbrucken, Germany},
	author       = {Botinis, Antonis and Bakakou-Orphanou, Aikaterini and Themistocleous, Charalambos},
	year         = {2007},
	address      = {Saarbrücken, Germany},
	pages        = {1341--1344},
}

@book{themistocleous-2011-prosody-239902,
	title        = {Prosody and Information Structure in Athenian and Cypriot Greek (Prosodia kai plirophoriaki domi stin Atheniaki kai kypriaki Ellinici)},
	author       = {Themistocleous, Charalambos},
	year         = {2011},
	publisher    = {University of Athens},
	address      = {Athens},
}

@inProceedings{themistocleous-2008-focus-239878,
	title        = {Focus Effects on Syllable Duration in Cypriot Greek},
	booktitle    = {Experimental Linguistics 2008 (ExLing 2008)},
	author       = {Themistocleous, Charalambos},
	year         = {2008},
	publisher    = {University of Athens},
	address      = {Athens},
	ISBN         = {978-960-466-020-9},
	pages        = {241--244},
}

@inProceedings{botinis-etal-2004-duration-239898,
	title        = {Duration correlates of stop consonants in Cypriot Greek},
	abstract     = {This is a production study of stop consonant durations as a function of voice, length, stress, syllable position, speech tempo and speaker’ s gender in Cypriot Greek. The results indicate that all six investigated factors have a significant effect on either total duration of the stops or one of their occlusion and burst parts.},
	booktitle    = {FONETIK 2004. The XVIIth Swedish Phonetics Conference May 26-28 2004},
	author       = {Botinis, Antonis and Christofi, Marios and Themistocleous, Charalambos and Kyprianou, Aggeliki},
	year         = {2004},
	publisher    = {Dept. of Linguistics, Stockholm University},
	address      = {Stockholm},
	ISBN         = {91-7265-901-7},
	pages        = {140--143},
}

@incollection{themistocleous-2011-cypriot-239891,
	title        = {Cypriot Greek Nuclear Pitch Accents},
	abstract     = {This study examines the Cypriot Greek nuclear pitch accents. Specifically, it examines the tonal representation of narrow information focus, broad focus, contrastive focus, and contrastive topic. For this purpose, an experiment has been designed in which four distinct contexts have been devised in order to elicit utterances with different information structure categories. The experiment examined whether the information structure categories are associated with categorically distinct nuclear pitch accents. The results showed no effect of the information structure categories on the realization of the nuclear pitch accents. However, secondary cues from segmental duration distinguish narrow foci from broad focus. These results, which underline a division of labor between phonological structure and meaning interpretation, are difficult to accommodate in models that directly associate meaning to tonal categories.},
	booktitle    = {In Z. Gavriilidou, A. Efthymiou, E. Thomadaki & P. Kambakis-Vougiouklis (eds), 2012, Selected papers of the 10th ICGL},
	author       = {Themistocleous, Charalambos},
	year         = {2011},
	publisher    = {Democritus University of Thrace},
	address      = {Komotini, Greece},
	pages        = {796 – 805},
}

@inProceedings{themistocleous-etal-2012-cypriot-232924,
	title        = {Cypriot Greek Lexicography: An online lexical database},
	abstract     = {This article presents an online dictionary environment, with enhanced sorting and searching functionalities and a text to speech feature, for hearing the pronunciation of the words. The online dictionary environment has been developed as part of the ‘Syntychies’ research program. ‘Syntychies’ online environment is a pioneering web-service for Greek dialectal lexicography and it is the first of its kind for Cypriot Greek.},
	booktitle    = {Proceedings of the 15th EURALEX International Congress. Eds. Ruth Vatvedt Fjeld and Julie Matilde Torjusen 7-11 August 2012. Oslo},
	author       = {Themistocleous, Charalambos and Katsoyannou, Marianna and Armosti, Spyros and Christodoulou, Kyriaki},
	year         = {2012},
	publisher    = {Department of Linguistics and Scandinavian Studies, University of Oslo},
	address      = {Oslo},
	ISBN         = {9788230320952},
	pages        = {889--891},
}

@incollection{themistocleous-2012-meaning-232920,
	title        = {Meaning and Form in the Tonal Representation},
	booktitle    = {Current Trends in Greek Linguistics / editor(s): Georgia Fragaki, Thanasis Georgakopoulos and Charalambos Themistocleous },
	author       = {Themistocleous, Charalambos},
	year         = {2012},
	publisher    = {Cambridge Scholars Publishing},
	address      = {Newcastle upon Tyne},
	ISBN         = {978-1-4438-4025-5},
	pages        = {271--289},
}

@inProceedings{efstathopoulou-themistocleous-2010-comparative-231101,
	title        = {A comparative study of Greek and English VOTs produced by Cypriot Greeks and Greek Canadians},
	abstract     = {This study examined the voice onset time (VOT) of English/Greek voiceless stops [p t k] produced by speakers of Cypriot Greek (henceforth CG) using comparative data from Greek/English bilinguals living in the Greater Vancouver area, Canada (henceforth Greek-Canadians, GrC). The purpose of this study was twofold: First, it examined CG stop consonants, including VOT measurements (c.f. Klatt, 1973, 1975; Lisker & Abramson, 1964, 1967). Second, it was a comparative research of the differences concerning VOTs of stop consonants of Greek and English tokens, between a late situation of diglossia of Greek-Canadians and the Greek/English diglossia of CG speakers (c.f Efstathopoulou, 2006, 2007). Five Cypriot Greek speakers uttered three voiceless stops [p t k] preceding five vowels [a e i o u] in initially stressed syllables in disyllabic CVCV words. Factors such as age, level of education were also taken into account. The analysis of the Greek stops yielded significant differences in VOTs among the Greek varieties studied, attributed to the greater sociolinguistic differences of the speakers (c.f. Ferguson, 1959, 1996; Tsiplakou et al. 2003). Furthermore, the production of English stops resulted in significant differences in the examined bilingual populations. These differences were attributed to sociolinguistic differences among the speakers, to the internal grammar of each variety and to the degree of exposure to English native stops that the speakers of each Greek variety had.},
	booktitle    = {5th Athens Postgraduate Conference of the Faculty of Philology National and Kapodistrian University of Athens 29-31 May 2009},
	author       = {Efstathopoulou, Pagona-Niki and Themistocleous, Charalambos},
	year         = {2010},
	publisher    = {University of Athens},
	address      = {Athens},
	ISBN         = {978-960-466-056-8},
}