Programme

Here is the list of posters that will be presented in the 72nd lunch:

Language models like the Bidirectional Encoder Representations from Transformers (BERT) have been established as the new state of the art in Natural Language Processing (NLP). These models use attention mechanisms, which despite their good performance, it is unclear how they affect the model's outcome. There have been some studies that either prove that,  not all attention heads are contributing much to the models' performance or that there is no correlation between attention weights and the model's outcome. On the other hand, some studies claim that attention provides some form of explanation for the model's performance. All these studies have either used attention mechanisms with deep neural networks other than BERT or analysed the attention weights of BERT on tasks other than text classification, mostly using sentiment analysis datasets like IMDB or SST2. In this paper, we examined the pattern of attention weights of fine-tuned BERT on text classification tasks on multiple cyberbullying-related datasets and provided an analysis of their relationship with its outcome. Our findings show that although fine-tuning does change the pattern of BERT's attention weights, there is no evidence that the attention weights of fine-tuned BERT play a direct role in its outcome.
In my presentation, a case of dialect variation in German will be investigated within a usage-based Construction Grammar framework. I will focus on one dialect area, namely the state of Vorarlberg in Austria. The construction in question is the dative+possessive construction, which is used to express a possessive relation. There are a number of other constructions that can be used for the same purpose; however, the construction in question is restricted to usage in dialect language and does not occur in standard German. Therefore, one question that will be addressed is which factors might have an influence on the choice of this variable. Language-internal factors that can be identified are information structure as well as the special socio-pragmatic meaning of the construction. Furthermore, since the dative+possessive construction occurs in almost all German dialects, there is a fair amount of variation regarding the construction itself. Within the defined dialect area, the construction seems to have developed some idiosyncratic features that lead to a mismatch between form and meaning. I argue that this is due to a neoanalysis of the possessive element as a constructional marker. By relating the construction in question to other possible variants in a network structure, analogical patterns can be discerned.
The term 'copredication' is commonly used to capture the phenomenon that we can use a single nominal to denote two or more categorically distinct kind of entities in the same sentence, like "The book is heavy and informative", "Alex drank and dropped the beer" or "The storm sunk my boat and my dreams". We propose that linguistic intuitions about the felicitousness of copredication statements are driven by phenomena related to the context-sensitive modulation of bodies of information on different levels of representation (words, sentences/situations, ...). We suggest that such an account can explain both the graded nature of felicitousness judgments and the context effects in the processing of co-predicative expressions. We underpin the account by a model of context-sensitive conceptual representations within the predictive processing framework.  
The main focus of this studyis the developments of the Middle Chinese sequence *ɣu-in 54 Yue dialects in Guangdong. Some examples of the reflexes of MC *ɣu-from several Yue dialects areshown in (1): (1)Middle Guangzhou Taishan Bao’an Maoming Chinese ‘to return’ *ɣuan [wan] [van] [wan] [ʋan] ‘lake’ *ɣu [wu] [vu] [fu] [fu] From (1), we can see that reflexes of *ɣu-can be [w-] or [v-] for both words or[f-](for ‘lake’)in combination with [w-] or [v-](for ‘to return’). The reflexes of MC *ɣu-sequence invites attention because (a) the previous dialect descriptions do not capture the developments of this sequence across all the dialects systematically;(b) the phonological history of *ɣu-in Yue remains descriptive in Chinese dialectology, i.e. listing correspondences between Middle Chinese and present-day Yue dialects instead of proposing sound changes that were involved and (c) previous studies do not explore numerous dialects all at once. The data used in these studies are from several Yue dialect surveys (Zhan & Cheung 1987, Zhan & Cheung 1994, Zhan & Cheung 1998, Shao 2016 and Beijing University Linguistics Faculty 1989). After assessing the survey data, I have identified 4 main patterns of reflexes: i) Guangzhou-, ii) Taishan-, iii) Maoming-and iv) Bao’an-type dialects. These patterns I have identified resemble the examples given in (1).
Native-speakerism, an ideology promoting native English speakers as perfect speech models for learners, affects English-language teachers around the world—whether they are native speakers (NSs) or non-native speakers (NNSs), whether they are foreign or local to their locales. These dichotomies—NS/NNS and foreign/local—pigeon-hole teachers into role identities not always suitable for each classroom’s individual context. However, from a translanguaging perspective, a native speaking English teacher (NEST) utilizes their linguistic repertoire to include the entirety of that teacher’s linguistic resources, consciously or subconsciously, in order to accommodate their speech for their students’ proficiency levels and the languages common between them. While some amount of native-speakerist idealism is reinforced by institutional policies (e.g., team-teaching schemes and English-only classroom rules), institutional support for translanguaging spaces is covert at most. This study thus explores the relationship between translanguaging and native-speakerism and how these concepts manifest and interact in Japanese EFL classrooms. Through interviews, questionnaires, and classroom observations, three American NESTs in Japan navigate their respective teacher and NS identities. The results suggest that the coexistence of native-speakerism and translanguaging practices within one classroom creates tension in an individual teacher’s role identity. The participants’ beliefs about their jobs do not always match up with their goals nor with their actions in the classroom. Through the frameworks of grammar of culture and the Dynamic Systems Model of Role Identity, native-speakerism acts as a barrier that prevents the teachers’ identities from achieving a stable system state, and the creation of a translanguaging space is thus seen as a cultural processor and competing practice against native-speakerist ideals.
Fine grained named entity typing is the task of assigning, for example, the type “person/politician" to the named entity "Barack Obama". So far, there are no openly available systems that perform this task for languages other than English. We propose annotation projection as a method to create non-English training data for existing typing models and discuss different sources of noise in this process using the example of English and German.
Summarization systems face the core challenge of identifying and selecting important information. In this paper, we tackle the problem of content selection in unsupervised extractive summarization of long, structured documents. We introduce a wide range of heuristics that leverage cognitive representations and organization of content units in order to rank sentences. Our experiments show that summarizers leveraging cognitive structures and the organization of the document are capable of extracting more summary-worthy content units, compared to heuristics that use frequency counts.
Slavic languages are known to lack what English-language scholarship calls a ‘definite article’. Yet there exists at least one Slavic language with this feature: Bulgarian. The Bulgarian marker of definiteness has an affixal nature, i.e. it is attached to the end of the word:  (1) Vizh-dam            kalendar-Ø see-1.SG              calendar.M-INDF ‘I see a calendar.’  (2) Vizh-dam            kalendar-a. see-1.SG              calendar.M-DEF.M ‘I see the calendar.’  (3) Vizh-dam            kniga-ta. see-1.SG              book.F-DEF.F ‘I see the book.’  (4) Vizh-dam            bashta-ta. see-1.SG              father.M-DEF.F ‘I see the father.’  We can see that there are different forms of the definite morpheme depending on the gender of the word it is attached to. However, in (4) we can see that a masculine noun (‘bashta’) takes the feminine definite morpheme. Why is this the case? If an adjective is inserted in the NP, the morpheme assumes the ‘proper’ gender:  (5) Vizh-dam             stari-ya                 bashta. see-1.SG              old-DEF.M           father ‘I see the old father.’  In my work, I tackle these issues as well as others by trying to offer a holistic account of what I call the Bulgarian Definite ending (DE). I contrast its morphology and usage with the well-established accounts of the English definite article (DA) in attempt to gain new theoretical insight. The findings of this work have implications for theoretical and applied research including phono-syntactic theory and English language teaching. 
The question of vowel compression in Spanish is much debated. Through a dialectal approach, I present evidence that phonologically, prosodically and phonetically constrained coda-driven vowel compression occurs in Spanish, namely Altiplateau Mexican Spanish.  Vowel compression is the phonological process in which vowels are shortened relative to syllable structure, i.e. the greater number of segments, the more extreme the shortening. Maddieson’s (1985) Closed Syllable Shortening principle claimed that compression was coda-driven in all languages, i.e. vowels in closed syllables were shorter than those in open syllables. grande (big, SINGULAR) [ˈɡran̪.de] grandes (big, PLURAL) [ˈɡran̪.dĕs] Recent study however has questioned this. Katz (2012) showed that, in English, durational differences between vowels in open and closed syllables were negligible and that onset and coda complexity had no effect. In Aldrich & Simonet (2019), onset complexity, not coda complexity, was responsible for Spanish vowel compression. In both studies, onset-induced compression was consistent across all speakers whilst coda-induced compression was speaker-specific. Particular to the present study is the following quote: "We are aware of no published finding suggesting that, in Spanish, coda presence (or complexity) drives compensatory vowel shortening" (Aldrich & Simonet, 2019:268) Although this may be true for the Aldrich & Simonet’s aggregate findings, dialect specific analysis shows that coda-induced compression exists in Spanish. This paper therefore analyses Unstressed Vowel Reduction in Altiplateau Mexican Spanish (a variety of Spanish local to Mexico’s central highlands, i.e. Mexico City and Toluca) and shows the following: Unstressed mid and low vowels (/e, o, a/) centralise and shorten. Compression is phonologically constrained, occurring in word-final, post-tonic closed syllables. Although complex onsets result in shorter vowels than simplex onsets, these results are negligible when compared to the effect of coda. The phonetic environment influences reduction: compression is most perceivable when the coda is /s/. Prosodic effects are noted: compression is more extreme in prosodic domain-final positions, i.e. utterance-final, where complete vowel elision is possible. Contrary to recent study, this analysis shows that phonologically, phonetically and prosodically constrained coda-induced compression occurs in Spanish and that dialectal-specific phonologies exist. We link these phonologies to a wider rhythmic debate: these dialect-specific phonologies within Spanish allow dialects to behave in a way typically associated with stress-timed languages where reduction effects are common. Ultimately, this supports the claim that dichotomies between syllable- and stress-timed languages are not as fixed as previously claimed and that languages, and varieties within them, exist along a rhythm continuum (Aldrich & Simonet, 2019; Arvaniti, 2012; Duaer, 1982). In order to substantiate these claims further, it is necessary to examine sound changes across varieties of Spanish as well as across other rhythmically distinct languages, i.e. Portuguese. Nonetheless, this small-scale study is the first step in showing the existence of dialect-specific phonologies and rhythms within a language.
Tonal alignment is the timing of f0-contours in relation to speech segments. Perceptual studies indicate minimum perceptible differences in alignment may be about 50 ms (House 1999), while production data indicate category means may differ only 40 ms when tonal alignment is used for phonemic contrast, e.g. in Shilluk (Remijsen & Ayoker 2014). How can listeners perceive such small differences in phonetic implementation? How perceptually sensitive to differences in tonal alignment are listeners? We hypothesized that Shilluk listeners would be very sensitive to differences in tonal alignment, and demonstrate categorical perception, while listeners of control language Spanish would demonstrate continuous perception. A first forced-choice discrimination experiment contained pairs of synthesized stimuli with falling contours on the same CVC nonsense word. Each stimulus within a pair diverged equally from a given anchor point, at the start of the vowel (0 ms), slightly into the vowel (30 ms), or towards the end (80 ms) of the vowel with a duration of 150 ms. Crucially, the middle anchor point is at the estimated phonemic boundary between the contrastively aligned falling tonemes in Shilluk. Experiment 1 shows, firstly, listeners are not very sensitive to differences in tonal alignment: only gaps as large as 80 ms could be reliably perceived. Secondly, perceptual sensitivity to differences in tonal alignment is greater later in the nucleus/vowel. On average, Spanish and Shilluk listeners both perceive differences in tonal alignment continuously. These results were unexpected for the Shilluk listeners: the presence of contrastive tonal alignment in their L1 phonology suggests it would be sensitive and categorical. A second forced-choice discrimination experiment was conducted to test whether the presence of a nucleus-coda boundary would yield a different perception pattern for L1 Shilluk participants. The nonsense word in Experiment 2 had a short vowel of 70 ms, resulting in the late anchor point at 80 ms occurring in the coda. While the anchor point at 80 ms into the rhyme had been most sensitive when it occurred during the nucleus/vowel in Experiment 1, it was least sensitive when it extends into the coda. The present study clearly demonstrates that the presence of the nucleus-coda boundary changes the perceptibility of the stimulus pairs of the late anchor point condition. These findings suggest that defining tonal alignment simply in relation to the onset of the vowel does not accurately reflect the way it is perceived. References House, D. (1999). Perception of pitch and tonal timing: implications for mechanisms of tonogenesis. ICPhS14 Proceedings ’99, San Francisco, 1823–1826. Remijsen, B., & Ayoker, O. G. (2014). Contrastive tonal alignment in falling contours in Shilluk. Phonology, 31(3), 435–462. https://doi.org/10.1017/S0952675714000219


Supported by


Contact

Email: languagelunch@ed.ac.uk