Refine
Year of publication
Document Type
- Doctoral Thesis (97) (remove)
Language
- English (97) (remove)
Keywords
- Satzverarbeitung (8)
- Spracherwerb (7)
- psycholinguistics (7)
- sentence processing (7)
- Informationsstruktur (6)
- Psycholinguistik (6)
- eye-tracking (6)
- information structure (6)
- prosody (6)
- Fokus (5)
Institute
- Department Linguistik (97) (remove)
This dissertation examines the integration of incongruent visual-scene and morphological-case information (“cues”) in building thematic-role representations of spoken relative clauses in German.
Addressing the mutual influence of visual and linguistic processing, the Coordinated Interplay Account (CIA) describes a mechanism in two steps supporting visuo-linguistic integration (Knoeferle & Crocker, 2006, Cog Sci). However, the outcomes and dynamics of integrating incongruent thematic-role representations from distinct sources have been investigated scarcely. Further, there is evidence that both second-language (L2) and older speakers may rely on non-syntactic cues relatively more than first-language (L1)/young speakers. Yet, the role of visual information for thematic-role comprehension has not been measured in L2 speakers, and only limitedly across the adult lifespan.
Thematically unambiguous canonically ordered (subject-extracted) and noncanonically ordered (object-extracted) spoken relative clauses in German (see 1a-b) were presented in isolation and alongside visual scenes conveying either the same (congruent) or the opposite (incongruent) thematic relations as the sentence did.
1 a Das ist der Koch, der die Braut verfolgt.
This is the.NOM cook who.NOM the.ACC bride follows
This is the cook who is following the bride.
b Das ist der Koch, den die Braut verfolgt.
This is the.NOM cook whom.ACC the.NOM bride follows
This is the cook whom the bride is following.
The relative contribution of each cue to thematic-role representations was assessed with agent identification. Accuracy and latency data were collected post-sentence from a sample of L1 and L2 speakers (Zona & Felser, 2023), and from a sample of L1 speakers from across the adult lifespan (Zona & Reifegerste, under review). In addition, the moment-by-moment dynamics of thematic-role assignment were investigated with mouse tracking in a young L1 sample (Zona, under review).
The following questions were addressed: (1) How do visual scenes influence thematic-role representations of canonical and noncanonical sentences? (2) How does reliance on visual-scene, case, and word-order cues vary in L1 and L2 speakers? (3) How does reliance on visual-scene, case, and word-order cues change across the lifespan?
The results showed reliable effects of incongruence of visually and linguistically conveyed thematic relations on thematic-role representations. Incongruent (vs. congruent) scenes yielded slower and less accurate responses to agent-identification probes presented post-sentence. The recently inspected agent was considered as the most likely agent ~300ms after trial onset, and the convergence of visual scenes and word order enabled comprehenders to assign thematic roles predictively.
L2 (vs. L1) participants relied more on word order overall. In response to noncanonical clauses presented with incongruent visual scenes, sensitivity to case predicted the size of incongruence effects better than L1-L2 grouping. These results suggest that the individual’s ability to exploit specific cues might predict their weighting.
Sensitivity to case was stable throughout the lifespan, while visual effects increased with increasing age and were modulated by individual interference-inhibition levels. Thus, age-related changes in comprehension may stem from stronger reliance on visually (vs. linguistically) conveyed meaning.
These patterns represent evidence for a recent-role preference – i.e., a tendency to re-assign visually conveyed thematic roles to the same referents in temporally coordinated utterances. The findings (i) extend the generalizability of CIA predictions across stimuli, tasks, populations, and measures of interest, (ii) contribute to specifying the outcomes and mechanisms of detecting and indexing incongruent representations within the CIA, and (iii) speak to current efforts to understand the sources of variability in sentence comprehension.
Background: Individuals with aphasia after stroke (IWA) often present with working memory (WM) deficits. Research investigating the relationship between WM and language abilities has led to the promising hypothesis that treatments of WM could lead to improvements in language, a phenomenon known as transfer. Although recent treatment protocols have been successful in improving WM, the evidence to date is scarce and the extent to which improvements in trained tasks of WM transfer to untrained memory tasks, spoken sentence comprehension, and functional communication is yet poorly understood.
Aims: We aimed at (a) investigating whether WM can be improved through an adaptive n-back training in IWA (Study 1–3); (b) testing whether WM training leads to near transfer to unpracticed WM tasks (Study 1–3), and far transfer to spoken sentence comprehension (Study 1–3), functional communication (Study 2–3), and memory in daily life in IWA (Study 2–3); and (c) evaluating the methodological quality of existing WM treatments in IWA (Study 3). To address these goals, we conducted two empirical studies – a case-controls study with Hungarian speaking IWA (Study 1) and a multiple baseline study with German speaking IWA (Study 2) – and a systematic review (Study 3).
Methods: In Study 1 and 2 participants with chronic, post-stroke aphasia performed an adaptive, computerized n-back training. ‘Adaptivity’ was implemented by adjusting the tasks’ difficulty level according to the participants’ performance, ensuring that they always practiced at an optimal level of difficulty. To assess the specificity of transfer effects and to better understand the underlying mechanisms of transfer on spoken sentence comprehension, we included an outcome measure testing specific syntactic structures that have been proposed to involve WM processes (e.g., non-canonical structures with varying complexity).
Results: We detected a mixed pattern of training and transfer effects across individuals: five participants out of six significantly improved in the n-back training. Our most important finding is that all six participants improved significantly in spoken sentence comprehension (i.e., far transfer effects). In addition, we also found far transfer to functional communication (in two participants out of three in Study 2) and everyday memory functioning (in all three participants in Study 2), and near transfer to unpracticed n-back tasks (in four participants out of six). Pooled data analysis of Study 1 and 2 showed a significant negative relationship between initial spoken sentence comprehension and the amount of improvement in this ability, suggesting that the more severe the participants’ spoken sentence comprehension deficit was at the beginning of training, the more they improved after training. Taken together, we detected both near far and transfer effects in our studies, but the effects varied across participants. The systematic review evaluating the methodological quality of existing WM treatments in stroke IWA (Study 3) showed poor internal and external validity across the included 17 studies. Poor internal validity was mainly due to use of inappropriate design, lack of randomization of study phases, lack of blinding of participants and/or assessors, and insufficient sampling. Low external validity was mainly related to incomplete information on the setting, lack of use of appropriate analysis or justification for the suitability of the analysis procedure used, and lack of replication across participants and/or behaviors. Results in terms of WM, spoken sentence comprehension, and reading are promising, but further studies with more rigorous methodology and stronger experimental control are needed to determine the beneficial effects of WM intervention.
Conclusions: Results of the empirical studies suggest that WM can be improved with a computerized and adaptive WM training, and improvements can lead to transfer effects to spoken sentence comprehension and functional communication in some individuals with chronic post-stroke aphasia. The fact that improvements were not specific to certain syntactic structures (i.e., non-canonical complex sentences) in spoken sentence comprehension suggest that WM is not involved in the online, automatic processing of syntactic information (i.e., parsing and interpretation), but plays a more general role in the later stage of spoken sentence comprehension (i.e., post-interpretive comprehension). The individual differences in treatment outcomes call for future research to clarify how far these results are generalizable to the population level of IWA. Future studies are needed to identify a few mechanisms that may generalize to at least a subpopulation of IWA as well as to investigate baseline non-linguistic cognitive and language abilities that may play a role in transfer effects and the maintenance of such effects. These may require larger yet homogenous samples.
Breaking down complexity
(2015)
The unbounded expressive capacity of human language cannot boil down to an infinite list of sentences stored in a finite brain. Our linguistic knowledge is rather grounded around a rule-based universal syntactic computation—called Merge—which takes categorized units in input (e.g. this and ship), and generates structures by binding words recursively into more complex hierarchies of any length (e.g. this ship; this ship sinks…). Here we present data from different fMRI datasets probing the cortical implementation of this fundamental process. We first pushed complexity down to a three-word level, to explore how Merge creates minimally hierarchical phrases and sentences. We then moved to the most fundamental two-word level, to directly assess the universal invariant nature of Merge, when no additive mechanisms are involved. Our most general finding is that Merge as the basic syntactic operation is primarily performed by confined area, namely BA 44 in the IFG. Activity reduces to its most ventral-anterior portion at the most fundamental level, following fine-grained sub-anatomical parcellation proposed for the region. The deep frontal operculum/anterior-dorsal insula (FOP/adINS), a phylogenetically older and less specialized region, rather appears to support word-accumulation processing in which the categorical information of the word is first accessed based on its lexical status, and then maintained on hold before further processing takes place. The present data confirm the general notion of BA 44 being activated as a function of complex structural hierarchy, but they go beyond this view by proposing that structural sensitivity in BA 44 is already appreciated at the lowest levels of complexity during which minimal phrase-structures are build up, and syntactic Merge is assessed. Further, they call for a redefinition of BA 44 from multimodal area to a macro-region with internal localizable functional profiles
Successful sentence comprehension requires the comprehender to correctly figure out who did what to whom. For example, in the sentence John kicked the ball, the comprehender has to figure out who did the action of kicking and what was being kicked. This process of identifying and connecting the syntactically-related words in a sentence is called dependency completion. What are the cognitive constraints that determine dependency completion? A widely-accepted theory is cue-based retrieval. The theory maintains that dependency completion is driven by a content-addressable search for the co-dependents in memory. The cue-based retrieval explains a wide range of empirical data from several constructions including subject-verb agreement, subject-verb non-agreement, plausibility mismatch configurations, and negative polarity items.
However, there are two major empirical challenges to the theory: (i) Grammatical sentences’ data from subject-verb number agreement dependencies, where the theory predicts a slowdown at the verb in sentences like the key to the cabinet was rusty compared to the key to the cabinets was rusty, but the data are inconsistent with this prediction; and, (ii) Data from antecedent-reflexive dependencies, where a facilitation in reading times is predicted at the reflexive in the bodybuilder who worked with the trainers injured themselves vs. the bodybuilder who worked with the trainer injured themselves, but the data do not show a facilitatory effect.
The work presented in this dissertation is dedicated to building a more general theory of dependency completion that can account for the above two datasets without losing the original empirical coverage of the cue-based retrieval assumption. In two journal articles, I present computational modeling work that addresses the above two empirical challenges.
To explain the grammatical sentences’ data from subject-verb number agreement dependencies, I propose a new model that assumes that the cue-based retrieval operates on a probabilistically distorted representation of nouns in memory (Article I). This hybrid distortion-plus-retrieval model was compared against the existing candidate models using data from 17 studies on subject-verb number agreement in 4 languages. I find that the hybrid model outperforms the existing models of number agreement processing suggesting that the cue-based retrieval theory must incorporate a feature distortion assumption.
To account for the absence of facilitatory effect in antecedent-reflexive dependencies, I propose an individual difference model, which was built within the cue-based retrieval framework (Article II). The model assumes that individuals may differ in how strongly they weigh a syntactic cue over a number cue. The model was fitted to data from two studies on antecedent-reflexive dependencies, and the participant-level cue-weighting was estimated. We find that one-fourth of the participants, in both studies, weigh the syntactic cue higher than the number cue in processing reflexive dependencies and the remaining participants weigh the two cues equally. The result indicates that the absence of predicted facilitatory effect at the level of grouped data is driven by some, not all, participants who weigh syntactic cues higher than the number cue. More generally, the result demonstrates that the assumption of differential cue weighting is important for a theory of dependency completion processes. This differential cue weighting idea was independently supported by a modeling study on subject-verb non-agreement dependencies (Article III).
Overall, the cue-based retrieval, which is a general theory of dependency completion, needs to incorporate two new assumptions: (i) the nouns stored in memory can undergo probabilistic feature distortion, and (ii) the linguistic cues used for retrieval can be weighted differentially. This is the cumulative result of the modeling work presented in this dissertation.
The dissertation makes an important theoretical contribution: Sentence comprehension in humans is driven by a mechanism that assumes cue-based retrieval, probabilistic feature distortion, and differential cue weighting. This insight is theoretically important because there is some independent support for these three assumptions in sentence processing and the broader memory literature. The modeling work presented here is also methodologically important because for the first time, it demonstrates (i) how the complex models of sentence processing can be evaluated using data from multiple studies simultaneously, without oversimplifying the models, and (ii) how the inferences drawn from the individual-level behavior can be used in theory development.
In this thesis, I develop a theoretical implementation of prosodic reconstruction and apply it to the empirical domain of German sentences in which part of a focus or contrastive topic is fronted.
Prosodic reconstruction refers to the idea that sentences involving syntactic movement show prosodic parallels with corresponding simpler structures without movement. I propose to model this recurrent observation by ordering syntax-prosody mapping before copy deletion.
In order to account for the partial fronting data, the idea is extended to the mapping between prosody and information structure. This assumption helps to explain why object-initial sentences containing a broad focus or broad contrastive topic show similar prosodic and interpretative restrictions as sentences with canonical word order.
The empirical adequacy of the model is tested against a set of gradient acceptability judgments.
Early sensitivity to prosodic phrase boundary cues: Behavioral evidence from German-learning infants
(2023)
This dissertation seeks to shed light on the relation of phrasal prosody and developmental speech perception in German-learning infants. Three independent empirical studies explore the role of acoustic correlates of major prosodic boundaries, specifically pitch change, final lengthening, and pause, in infant boundary perception. Moreover, it was examined whether the sensitivity to prosodic phrase boundary markings changes during the first year of life as a result of perceptual attunement to the ambient language (Aslin & Pisoni, 1980).
Using the headturn preference procedure six- and eight-month-old monolingual German-learning infants were tested on their discrimination of two different prosodic groupings of the same list of coordinated names either with or without an internal IPB after the second name, that is, [Moni und Lilli] [und Manu] or [Moni und Lilli und Manu]. The boundary marking was systematically varied with respect to single prosodic cues or specific cue combinations.
Results revealed that six- and eight-month-old German-learning infants successfully detect the internal prosodic boundary when it is signaled by all the three main boundary cues pitch change, final lengthening, and pause. For eight-, but not for six-month-olds, the combination of pitch change and final lengthening, without the occurrence of a pause, is sufficient. This mirrors an adult-like perception by eight-months (Holzgrefe-Lang et al., 2016). Six-month-olds detect a prosodic phrase boundary signaled by final lengthening and pause. The findings suggest a developmental change in German prosodic boundary cue perception from a strong reliance on the pause cue at six months to a differentiated sensitivity to the more subtle cues pitch change and final lengthening at eight months. Neither for six- nor for eight-month-olds the occurrence of pitch change or final lengthening as single cues is sufficient, similar to what has been observed for adult speakers of German (Holzgrefe-Lang et al., 2016).
The present dissertation provides new scientific knowledge on infants’ sensitivity to individual prosodic phrase boundary cues in the first year of life. Methodologically, the studies are pathbreaking since they used exactly the same stimulus materials – phonologically thoroughly controlled lists of names – that have also been used with adults (Holzgrefe-Lang et al., 2016) and with infants in a neurophysiological paradigm (Holzgrefe-Lang, Wellmann, Höhle, & Wartenburger, 2018), allowing for comparisons across age (six/ eight months and adults) and method (behavioral vs. neurophysiological methods). Moreover, materials are suited to be transferred to other languages allowing for a crosslinguistic comparison. Taken together with a study with similar French materials (van Ommen et al., 2020) the observed change in sensitivity in German-learning infants can be interpreted as a language-specific one, from an initial language-general processing mechanism that primarily focuses on the presence of pauses to a language-specific processing that takes into account prosodic properties available in the ambient language. The developmental pattern is discussed as an interplay of acoustic salience, prosodic typology (prosodic regularity) and cue reliability.
This study presents new insights into null subjects, topic drop and the interpretation of topic-dropped elements. Besides providing an empirical data survey, it offers explanations to well-known problems, e.g. syncretisms in the context of null-subject licensing or the marginality of dropping an element which carries oblique case. The book constitutes a valuable source for both empirically and theoretically interested (generative) linguists.
Production of regular and non-regular verbs : evidence for a lexical entry complexity account
(2010)
The incredible productivity and creativity of language depends on two fundamental resources: a mental lexicon and a mental grammar. Rules of grammar enable us to produce and understand complex phrases we have not encountered before and at the same time constrain the computation of complex expressions. The concepts of the mental lexicon and mental grammar have been thoroughly tested by comparing the use of regular versus non-regular word forms. Regular verbs (e.g. walk-walked) are computed using a suffixation rule in a neural system for grammatical processing; non-regular verbs (run-ran) are retrieved from associative memory. The role of regularity has only been explored for the past tense, where regularity is overtly visible. To explore the representation and encoding of regularity as well as the inflectional processes involved in the production of regular and non-regular verbs, this dissertation investigated three groups of German verbs: regular, irregular and hybrid verbs. Hybrid verbs in German have completely regular conjugation in the present tense and irregular conjugation in the past tense. Articulation latencies were measured while participants named pictures of actions, producing the 3rd person singular of regular, hybrid, and irregular verbs in present and past tense. Studying the production of German verbs in past and present tense, this dissertation explored the complexity of lexical entries as a decisive factor in the production of verbs.
Distributed optimality
(2001)
In this thesis I propose a synthesis (Distributed Optimality, DO) between Optimality Theory (OT, Prince & Smolensky, 1993) and a morphological framework in a genuine derivational tradition, namely Distributed Morphology (DM) as developed by Halle & Marantz (1993). By carrying over the apparatus of OT to DM, phenomena which are captured in DM by language-specific rules or features of lexical entries, are given a more principled account in the terms of ranked universal constraints. On the other hand, also the DM part makes two contributions, namely strong locality and impoverishment. The first gives rise to a simple formal interpretation of DO, while the latter is shown to be indispensable in any theoretically satisfying account of agreement morphology. The empirical basis of the work is given by the complex agreement morphology of genetically different languages. Theoretical focus is mainly on two areas: First, so-called direction marking which is shown to be preferably treated in terms of constraints on feature realization. Second, the effects of precedence constraints which are claimed to regulate the status of agreement affixes as prefixes or suffixes and their respective order. A universal typology for the order of agreement categories by means of OT-constraints is proposed.
The individual’s mental lexicon comprises all known words as well related infor-mation on semantics, orthography and phonology. Moreover, entries connect due to simi-larities in these language domains building a large network structure. The access to lexical information is crucial for processing of words and sentences. Thus, a lack of information in-hibits the retrieval and can cause language processing difficulties. Hence, the composition of the mental lexicon is essential for language skills and its assessment is a central topic of lin-guistic and educational research.
In early childhood, measurement of the mental lexicon is uncomplicated, for example through parental questionnaires or the analysis of speech samples. However, with growing content the measurement becomes more challenging: With more and more words in the mental lexicon, the inclusion of all possible known words into a test or questionnaire be-comes impossible. That is why there is a lack of methods to assess the mental lexicon for school children and adults. For the same reason, there are only few findings on the courses of lexical development during school years as well as its specific effect on other language skills. This dissertation is supposed to close this gap by pursuing two major goals: First, I wanted to develop a method to assess lexical features, namely lexicon size and lexical struc-ture, for children of different age groups. Second, I aimed to describe the results of this method in terms of lexical development of size and structure. Findings were intended to help understanding mechanisms of lexical acquisition and inform theories on vocabulary growth.
The approach is based on the dictionary method where a sample of words out of a dictionary is tested and results are projected on the whole dictionary to determine an indi-vidual’s lexicon size. In the present study, the childLex corpus, a written language corpus for children in German, served as the basis for lexicon size estimation. The corpus is assumed to comprise all words children attending primary school could know. Testing a sample of words out of the corpus enables projection of the results on the whole corpus. For this purpose, a vocabulary test based on the corpus was developed. Afterwards, test performance of virtual participants was simulated by drawing different lexicon sizes from the corpus and comparing whether the test items were included in the lexicon or not. This allowed determination of the relation between test performance and total lexicon size and thus could be transferred to a sample of real participants. Besides lexicon size, lexical content could be approximated with this approach and analyzed in terms of lexical structure.
To pursue the presented aims and establish the sampling method, I conducted three consecutive studies. Study 1 includes the development of a vocabulary test based on the childLex corpus. The testing was based on the yes/no format and included three versions for different age groups. The validation grounded on the Rasch Model shows that it is a valid instrument to measure vocabulary for primary school children in German. In Study 2, I estab-lished the method to estimate lexicon sizes and present results on lexical development dur-ing primary school. Plausible results demonstrate that lexical growth follows a quadratic function starting with about 6,000 words at the beginning of school and about 73,000 words on average for young adults. Moreover, the study revealed large interindividual differences. Study 3 focused on the analysis of network structures and their development in the mental lexicon due to orthographic similarities. It demonstrates that networks possess small-word characteristics and decrease in interconnectivity with age.
Taken together, this dissertation provides an innovative approach for the assessment and description of the development of the mental lexicon from primary school onwards. The studies determine recent results on lexical acquisition in different age groups that were miss-ing before. They impressively show the importance of this period and display the existence of extensive interindividual differences in lexical development. One central aim of future research needs to address the causes and prevention of these differences. In addition, the application of the method for further research (e.g. the adaptation for other target groups) and teaching purposes (e.g. adaptation of texts for different target groups) appears to be promising.
Infants' lexical processing is modulated by featural manipulations made to words, suggesting that early lexical representations are sufficiently specified to establish a match with the corresponding label. However, the precise degree of detail in early words requires further investigation due to equivocal findings. We studied this question by assessing children’s sensitivity to the degree of featural manipulation (Chapters 2 and 3), and sensitivity to the featural makeup of homorganic and heterorganic consonant clusters (Chapter 4). Gradient sensitivity on the one hand and sensitivity to homorganicity on the other hand would suggest that lexical processing makes use of sub-phonemic information, which in turn would indicate that early words contain sub-phonemic detail. The studies presented in this thesis assess children’s sensitivity to sub-phonemic detail using minimally demanding online paradigms suitable for infants: single-picture pupillometry and intermodal preferential looking. Such paradigms have the potential to uncover lexical knowledge that may be masked otherwise due to cognitive limitations. The study reported in Chapter 2 obtained a differential response in pupil dilation to the degree of featural manipulation, a result consistent with gradient sensitivity. The study reported in Chapter 3 obtained a differential response in proportion of looking time and pupil dilation to the degree of featural manipulation, a result again consistent with gradient sensitivity. The study reported in Chapter 4 obtained a differential response to the manipulation of homorganic and heterorganic consonant clusters, a result consistent with sensitivity to homorganicity. These results suggest that infants' lexical representations are not only specific, but also detailed to the extent that they contain sub-phonemic information.
Does it have to be trees? : Data-driven dependency parsing with incomplete and noisy training data
(2011)
We present a novel approach to training data-driven dependency parsers on incomplete annotations. Our parsers are simple modifications of two well-known dependency parsers, the transition-based Malt parser and the graph-based MST parser. While previous work on parsing with incomplete data has typically couched the task in frameworks of unsupervised or semi-supervised machine learning, we essentially treat it as a supervised problem. In particular, we propose what we call agnostic parsers which hide all fragmentation in the training data from their supervised components. We present experimental results with training data that was obtained by means of annotation projection. Annotation projection is a resource-lean technique which allows us to transfer annotations from one language to another within a parallel corpus. However, the output tends to be noisy and incomplete due to cross-lingual non-parallelism and error-prone word alignments. This makes the projected annotations a suitable test bed for our fragment parsers. Our results show that (i) dependency parsers trained on large amounts of projected annotations achieve higher accuracy than the direct projections, and that (ii) our agnostic fragment parsers perform roughly on a par with the original parsers which are trained only on strictly filtered, complete trees. Finally, (iii) when our fragment parsers are trained on artificially fragmented but otherwise gold standard dependencies, the performance loss is moderate even with up to 50% of all edges removed.
This dissertation is concerned with the relation between qualitative phonological organization in the form of syllabic structure and continuous phonetics, that is, the spatial and temporal dimensions of vocal tract action that express syllabic structure. The main claim of the dissertation is twofold. First, we argue that syllabic organization exerts multiple effects on the spatio-temporal properties of the segments that partake in that organization. That is, there is no unique or privileged exponent of syllabic organization. Rather, syllabic organization is expressed in a pleiotropy of phonetic indices. Second, we claim that a better understanding of the relation between qualitative phonological organization and continuous phonetics is reached when one considers how the string of segments (over which the nature of the phonological organization is assessed) responds to perturbations (scaling of phonetic variables) of localized properties (such as durations) within that string. Specifically, variation in phonetic variables and more specifically prosodic variation is a crucial key to understanding the nature of the link between (phonological) syllabic organization and the phonetic spatio-temporal manifestation of that organization. The effects of prosodic variation on segmental properties and on the overlap between the segments, we argue, offer the right pathway to discover patterns related to syllabic organization. In our approach, to uncover evidence for global organization, the sequence of segments partaking in that organization as well as properties of these segments or their relations with one another must be somehow locally varied. The consequences of such variation on the rest of the sequence can then be used to unveil the span of organization. When local perturbations to segments or relations between adjacent segments have effects that ripple through the rest of the sequence, this is evidence that organization is global. If instead local perturbations stay local with no consequences for the rest of the whole, this indicates that organization is local.
The immense popularity of online communication services in the last decade has not only upended our lives (with news spreading like wildfire on the Web, presidents announcing their decisions on Twitter, and the outcome of political elections being determined on Facebook) but also dramatically increased the amount of data exchanged on these platforms. Therefore, if we wish to understand the needs of modern society better and want to protect it from new threats, we urgently need more robust, higher-quality natural language processing (NLP) applications that can recognize such necessities and menaces automatically, by analyzing uncensored texts. Unfortunately, most NLP programs today have been created for standard language, as we know it from newspapers, or, in the best case, adapted to the specifics of English social media.
This thesis reduces the existing deficit by entering the new frontier of German online communication and addressing one of its most prolific forms—users’ conversations on Twitter. In particular, it explores the ways and means by how people express their opinions on this service, examines current approaches to automatic mining of these feelings, and proposes novel methods, which outperform state-of-the-art techniques. For this purpose, I introduce a new corpus of German tweets that have been manually annotated with sentiments, their targets and holders, as well as lexical polarity items and their contextual modifiers. Using these data, I explore four major areas of sentiment research: (i) generation of sentiment lexicons, (ii) fine-grained opinion mining, (iii) message-level polarity classification, and (iv) discourse-aware sentiment analysis. In the first task, I compare three popular groups of lexicon generation methods: dictionary-, corpus-, and word-embedding–based ones, finding that dictionary-based systems generally yield better polarity lists than the last two groups. Apart from this, I propose a linear projection algorithm, whose results surpass many existing automatically-generated lexicons. Afterwords, in the second task, I examine two common approaches to automatic prediction of sentiment spans, their sources, and targets: conditional random fields (CRFs) and recurrent neural networks, obtaining higher scores with the former model and improving these results even further by redefining the structure of CRF graphs. When dealing with message-level polarity classification, I juxtapose three major sentiment paradigms: lexicon-, machine-learning–, and deep-learning–based systems, and try to unite the first and last of these method groups by introducing a bidirectional neural network with lexicon-based attention. Finally, in order to make the new classifier aware of microblogs' discourse structure, I let it separately analyze the elementary discourse units of each tweet and infer the overall polarity of a message from the scores of its EDUs with the help of two new approaches: latent-marginalized CRFs and Recursive Dirichlet Process.