Periodic Reporting for period 4 - MOR-PHON (Resolving Morpho-Phonological Alternation: Historical, Neurolinguistic, and Computational Approaches)
Reporting period: 2021-04-01 to 2022-09-30
The project MORPHON examined the way in which words are put together and how related words can differ across both history and languages. What distinguished our approach is that we combined expertise in (1) theoretical and typological linguistics, (2) brain-imaging methods, and (3) computational modelling. Our results have shed light on the existence and cross-linguistic incidence of morpho-phonological alternations, their diachronic profiles, their processing and mental representation. Rather than setting morpho-phonological alternations aside as irregularities of morphology (requiring individual listing and storing), we recognise certain kinds of them (stress shifts, feature changes, deletions, and tonal changes) as something universally to be expected in mental lexicons and as something the brains of speakers and listeners can easily handle.
In morpho-phonological alternations, the shapes of morphemes can differ between morphologically related word forms. For example, the morpho-phonologically related words revére – réverence differ with respect to where the main prominence falls on the word and how the stressed vowels differ. The second vowel in the verb 'revere' has a stressed [iː] (cf. 'bee') while the stressed vowel in the noun 'reverence' has [ɛ] (cf. 'bell'). The opaque phonological relationship between morphologically related forms has been a longstanding challenge in theoretical, historical, psycho- and neuro-linguistics, and computational linguistics alike. Morpho-phonological alternations of all kinds have been analysed across the languages of the world; but fundamental questions have remained controversial or indeed unasked:
▪ Why do they exist in the first place and why are they so widespread?
▪ How do they come about and what is their diachronic time-course?
▪ How are they represented in mental lexicons and how are they processed?
The position we advocated was that morpho-phonological variants are not listed and stored independently in the mental lexicon, but rather are mapped onto single abstract representations. This is a controversial position, and its defence required the systematic study of types of alternations and their histories, and precise hypotheses about the nature of mental representations. We examined different types of alternations throughout history across many languages (including Bengali, Dutch, English, German, Mandarin, Swedish) and focused on the internal structure of words under three broad areas:
A) Structure of affixation, morphological complexity and processing.
B) Morphophonological alternations (including metrical patterns) and role of affixation in loans in Germanic, keeping track of the abstract phonological representations
C) Compounding: conflicts between phonological and lexical word formation.
Address (URL) of the project's public website
http://brainlab.clp.ox.ac.uk/
In morpho-phonological alternations, the shapes of morphemes can differ between morphologically related word forms. For example, the morpho-phonologically related words revére – réverence differ with respect to where the main prominence falls on the word and how the stressed vowels differ. The second vowel in the verb 'revere' has a stressed [iː] (cf. 'bee') while the stressed vowel in the noun 'reverence' has [ɛ] (cf. 'bell'). The opaque phonological relationship between morphologically related forms has been a longstanding challenge in theoretical, historical, psycho- and neuro-linguistics, and computational linguistics alike. Morpho-phonological alternations of all kinds have been analysed across the languages of the world; but fundamental questions have remained controversial or indeed unasked:
▪ Why do they exist in the first place and why are they so widespread?
▪ How do they come about and what is their diachronic time-course?
▪ How are they represented in mental lexicons and how are they processed?
The position we advocated was that morpho-phonological variants are not listed and stored independently in the mental lexicon, but rather are mapped onto single abstract representations. This is a controversial position, and its defence required the systematic study of types of alternations and their histories, and precise hypotheses about the nature of mental representations. We examined different types of alternations throughout history across many languages (including Bengali, Dutch, English, German, Mandarin, Swedish) and focused on the internal structure of words under three broad areas:
A) Structure of affixation, morphological complexity and processing.
B) Morphophonological alternations (including metrical patterns) and role of affixation in loans in Germanic, keeping track of the abstract phonological representations
C) Compounding: conflicts between phonological and lexical word formation.
Address (URL) of the project's public website
http://brainlab.clp.ox.ac.uk/
Across experimental designs and in historical linguistics, we found that complexity in linguistic morphological structure plays a critical role in processing and change. Under (A), we experimentally examined morphological complexity in Bengali, English and German using EEG, behavioural, and eye-tracking techniques. Our results showed that morpheme complexity and constraints on morpheme sequences play a definite role in processing. In Bengali, processing asymmetries of affixed words in behavioural tasks were attributable to differences in inhibition from the phonological cohort and also to the salience of morpheme boundaries. Processing asymmetries (using an MMN paradigm) were also found for English ablaut verbs (e.g. get ~ got; sit ~ sat; *gef ~ *gof, *sif ~ *saf). Our EEG results showed that both for real words, as well as pseudowords, the phonological representations affected processing. Height mismatched in both directions for the sit ~ sat pair, while place was asymmetric in get ~ got pair, both of which were reflected in MMN patterns. Pseudowords behaved in the same way, indicating that the phonological representations of individual segments remained constant. Predicted symmetries in tonal dissimilation in Mandarin were also revealed in EEG patterns, and pitch accent tonal contrasts in Swedish showed that representations of the tones affected processing (revealed in eye-tracking) for monolinguals and bilinguals.
For (B), we considered the notions about the way in which analogical extensions and prosodic constraints interact even with loanwords. By focusing on medieval stages of English, Dutch and German, we found that metrical patterns in Germanic have consistently remained on the first syllable regardless of phonological alternations. Romance loans, however, caused havoc in the individual metrical systems, particularly due to affixed words. We have shown that, despite the large number of loans and substantial increase in complex words in the lexicons, the basic foot structure has remained trochaic. The critical difference is that English maintains its preference not to stress final syllables, making the main stress fall away from the right edge. In contrast, German and Dutch do allow final stressed syllables if they have the relevant weight. This suggests that all three languages have maintained their basic foot structure and presence or absence of heavy syllables at the right edge of words, despite the onslaught of loans.
For (C), our results supported the idea that compounding in English and Bengali carry similar lexical and morphological properties, but prosodically, the outputs can be very different. We also found evidence through naming tasks that visual concatenation (e.g. time zone vs. timeline) is not a reliable indicator of compounding in English: prosodic output can be very different from the visual input.
For (B), we considered the notions about the way in which analogical extensions and prosodic constraints interact even with loanwords. By focusing on medieval stages of English, Dutch and German, we found that metrical patterns in Germanic have consistently remained on the first syllable regardless of phonological alternations. Romance loans, however, caused havoc in the individual metrical systems, particularly due to affixed words. We have shown that, despite the large number of loans and substantial increase in complex words in the lexicons, the basic foot structure has remained trochaic. The critical difference is that English maintains its preference not to stress final syllables, making the main stress fall away from the right edge. In contrast, German and Dutch do allow final stressed syllables if they have the relevant weight. This suggests that all three languages have maintained their basic foot structure and presence or absence of heavy syllables at the right edge of words, despite the onslaught of loans.
For (C), our results supported the idea that compounding in English and Bengali carry similar lexical and morphological properties, but prosodically, the outputs can be very different. We also found evidence through naming tasks that visual concatenation (e.g. time zone vs. timeline) is not a reliable indicator of compounding in English: prosodic output can be very different from the visual input.
The evidence we have collected from all the different methodologies has resulted in a greater understanding of not only the changes in language over time but also the ideal methodological resources and tools by which to research this. Morphological processing combined with theoretical research on the historical developments have gone beyond what we originally planned. We have presented our work in more than 70 presentations including publications (articles, chapters and one edited volume) and presentations (conference and invited talks) summarising our findings.
In developing our speech recognition model, we identified an approach that results in an Automatic Speech Recognition (ASR) system that requires relatively little training, can be adapted flexibly to new languages (even those where training data in the form of annotated speech recordings may be sparse), and employs theoretical insights from the study of how the human brain processes speech to emulate this process on an app. Importantly, it aims to target those features that are essential to human understanding of speech and ignores or tolerates those that can vary across speakers or utterances.In all, the PI as well as the members of the team have given 21 talks (invited as well as at conferences where we submitted abstracts). Presentations were made at various linguistic conferences across the world including Toronto (Canada), Canberra (Australia), Amsterdam (The Netherlands), Berkeley (USA), New Delhi (India), Zurich (Switzerland) as well as at several universities across the UK.
In developing our speech recognition model, we identified an approach that results in an Automatic Speech Recognition (ASR) system that requires relatively little training, can be adapted flexibly to new languages (even those where training data in the form of annotated speech recordings may be sparse), and employs theoretical insights from the study of how the human brain processes speech to emulate this process on an app. Importantly, it aims to target those features that are essential to human understanding of speech and ignores or tolerates those that can vary across speakers or utterances.In all, the PI as well as the members of the team have given 21 talks (invited as well as at conferences where we submitted abstracts). Presentations were made at various linguistic conferences across the world including Toronto (Canada), Canberra (Australia), Amsterdam (The Netherlands), Berkeley (USA), New Delhi (India), Zurich (Switzerland) as well as at several universities across the UK.