Refine
Year of publication
- 2022 (48) (remove)
Document Type
- Article (42)
- Doctoral Thesis (3)
- Bachelor Thesis (1)
- Part of Periodical (1)
- Postprint (1)
Is part of the Bibliography
- yes (48)
Keywords
- German (3)
- alternatives (3)
- focus (3)
- Individual differences (2)
- N400 (2)
- Prosody (2)
- information structure (2)
- interference (2)
- processing (2)
- similarity-based interference (2)
Institute
- Department Linguistik (48) (remove)
Neural conversation models aim to predict appropriate contributions to a (given) conversation by using neural networks trained on dialogue data. A specific strand focuses on non-goal driven dialogues, first proposed by Ritter et al. (2011): They investigated the task of transforming an utterance into an appropriate reply. Then, this strand evolved into dialogue system approaches using long dialogue histories and additional background context. Contributing meaningful and appropriate to a conversation is a complex task, and therefore research in this area has been very diverse: Serban et al. (2016), for example, looked into utilizing variable length dialogue histories, Zhang et al. (2018) added additional context to the dialogue history, Wolf et al. (2019) proposed a model based on pre-trained Self-Attention neural networks (Vasvani et al., 2017), and Dinan et al. (2021) investigated safety issues of these approaches. This trend can be seen as a transformation from trying to somehow carry on a conversation to generating appropriate replies in a controlled and reliable way.
In this thesis, we first elaborate the meaning of appropriateness in the context of neural conversation models by drawing inspiration from the Cooperative Principle (Grice, 1975). We first define what an appropriate contribution has to be by operationalizing these maxims as demands on conversation models: being fluent, informative, consistent towards given context, coherent and following a social norm. Then, we identify different targets (or intervention points) to achieve the conversational appropriateness by investigating recent research in that field.
In this thesis, we investigate the aspect of consistency towards context in greater detail, being one aspect of our interpretation of appropriateness.
During the research, we developed a new context-based dialogue dataset (KOMODIS) that combines factual and opinionated context to dialogues. The KOMODIS
dataset is publicly available and we use the data in this thesis to gather new insights in context-augmented dialogue generation.
We further introduced a new way of encoding context within Self-Attention based neural networks. For that, we elaborate the issue of space complexity from knowledge graphs,
and propose a concise encoding strategy for structured context inspired from graph neural networks (Gilmer et al., 2017) to reduce the space complexity of the additional context. We discuss limitations of context-augmentation for neural conversation models, explore the characteristics of knowledge graphs, and explain how we create and augment knowledge graphs for our experiments.
Lastly, we analyzed the potential of reinforcement and transfer learning to improve context-consistency for neural conversation models. We find that current reward functions need to be more precise to enable the potential of reinforcement learning, and that sequential transfer learning can improve the subjective quality of generated dialogues.
The Final-over-Final Condition has emerged as a robust and explanatory generalization for a wide range of phenomena (Biberauer, Holmberg, and Roberts 2014, Sheehan et al. 2017). In this article, we argue that it also holds in another domain, nominalization. In languages that show overt nominalization of VPs, one word order is routinely unattested, namely, a head-initial VP with a suffixal nominalizer. This typological gap can be accounted for by the Final-over-Final Condition, if we allow it to hold within mixed extended projections. This view also makes correct predictions about agentive nominalizations and nominalized serial verb constructions.
Discourse Prominence and Antecedent MisRetrieval during Native and Non-Native Pronoun Resolution
(2022)
Previous studies on non-native (L2) anaphor resolution suggest that L2 comprehenders are guided more strongly by discourse-level cues compared to native (L1) comprehenders. Here we examine whether and how a grammatically inappropriate antecedent’s discourse status affects the likelihood of it being considered during L1 and L2 pronoun resolution. We used an interference paradigm to examine how the extrasentential discourse impacts the resolution of German object pronouns. In an eye-tracking-during-reading experiment we examined whether an elaborated local antecedent ruled out by binding Condition B would be mis-retrieved during pronoun resolution, and whether initially introducing this antecedent as the discourse topic would affect the chances of it being mis-retrieved. While both participant groups rejected the inappropriate antecedent in an offline questionnaire irrespective of its discourse prominence, their real-time processing patterns differed. L1 speakers initially mis-retrieved the inappropriate antecedent regardless of its contextual prominence. L1 Russian/L2 German speakers, in contrast, were affected by the antecedent’s discourse status, considering it only when it was discourse-new but not when it had previously been introduced as the discourse topic. Our findings show that L2 comprehenders are highly sensitive to discourse dynamics such as topic shifts, supporting the claim that discourse-level cues are more strongly weighted during L2 compared to L1 processing.
Meaning and alternatives
(2022)
Alternatives and competition in language are pervasive at all levels of linguistic analysis. More specifically, alternatives have been argued to play a prominent role in an ever-growing class of phenomena in the investigation of natural language meaning. In this article, we focus on scalar implicatures, as they are arguably the most paradigmatic case of an alternative-based phenomenon. We first review the main challenge for theories of alternatives, the so-called symmetry problem, and we briefly discuss how it has shaped the different approaches to alternatives. We then turn to two more recent challenges concerning scalar diversity and the inferences of sentences with multiple scalars. Finally, we describe several related alternative-based phenomena and recent conceptual approaches to alternatives. As we discuss, while important progress has been made, much more work is needed both on the theoretical side and on understanding the empirical landscape better.
The aim of this dissertation was to conduct a larger-scale cross-linguistic empirical investigation of similarity-based interference effects in sentence comprehension.
Interference studies can offer valuable insights into the mechanisms that are involved in long-distance dependency completion.
Many studies have investigated similarity-based interference effects, showing that syntactic and semantic information are employed during long-distance dependency formation (e.g., Arnett & Wagers, 2017; Cunnings & Sturt, 2018; Van Dyke, 2007, Van Dyke & Lewis, 2003; Van Dyke & McElree, 2011). Nevertheless, there are some important open questions in the interference literature that are critical to our understanding of the constraints involved in dependency resolution.
The first research question concerns the relative timing of syntactic and semantic interference in online sentence comprehension. Only few interference studies have investigated this question, and, to date, there is not enough data to draw conclusions with regard to their time course (Van Dyke, 2007; Van Dyke & McElree, 2011).
Our first cross-linguistic study explores the relative timing of syntactic and semantic interference in two eye-tracking reading experiments that implement the study design used in Van Dyke (2007). The first experiment tests English sentences. The second, larger-sample experiment investigates the two interference types in German.
Overall, the data suggest that syntactic and semantic interference can arise simultaneously during retrieval.
The second research question concerns a special case of semantic interference: We investigate whether cue-based retrieval interference can be caused by semantically similar items which are not embedded in a syntactic structure.
This second interference study builds on a landmark study by Van Dyke & McElree (2006). The study design used in their study is unique in that it is able to pin down the source of interference as a consequence of cue overload during retrieval, when semantic retrieval cues do not uniquely match the retrieval target. Unlike most other interference studies, this design is able to rule out encoding interference as an alternative explanation. Encoding accounts postulate that it is not cue overload at the retrieval site but the erroneous encoding of similar linguistic items in memory that leads to interference (Lewandowsky et al., 2008; Oberauer & Kliegl, 2006). While Van Dyke & McElree (2006) reported cue-based retrieval interference from sentence-external distractors, the evidence for this effect was weak. A subsequent study did not show interference of this type (Van Dyke et al., 2014). Given these inconclusive findings, further research is necessary to investigate semantic cue-based retrieval interference.
The second study in this dissertation provides a larger-scale cross-linguistic investigation of cue-based retrieval interference from sentence-external items. Three larger-sample eye-tracking studies in English, German, and Russian tested cue-based interference in the online processing of filler-gap dependencies. This study further extends the previous research by investigating interference in each language under varying task demands (Logačev & Vasishth, 2016; Swets et al., 2008).
Overall, we see some very modest support for proactive cue-based retrieval interference in English. Unexpectedly, this was observed only under a low task demand. In German and Russian, there is some evidence against the interference effect. It is possible that interference is attenuated in languages with richer case marking.
In sum, the cross-linguistic experiments on the time course of syntactic and semantic interference from sentence-internal distractors support existing evidence of syntactic and semantic interference during sentence comprehension. Our data further show that both types of interference effects can arise simultaneously. Our cross-linguistic experiments investigating semantic cue-based retrieval interference from sentence-external distractors suggest that this type of interference may arise only in specific linguistic contexts.
Dynamical models make specific assumptions about cognitive processes that generate human behavior. In data assimilation, these models are tested against timeordered data. Recent progress on Bayesian data assimilation demonstrates that this approach combines the strengths of statistical modeling of individual differences with the those of dynamical cognitive models.
In this paper we examine the effect of uncertainty on readers’ predictions about meaning. In particular, we were interested in how uncertainty might influence the likelihood of committing to a specific sentence meaning. We conducted two event-related potential (ERP) experiments using particle verbs such as turn down and manipulated uncertainty by constraining the context such that readers could be either highly certain about the identity of a distant verb particle, such as turn the bed […] down, or less certain due to competing particles, such as turn the music […] up/down. The study was conducted in German, where verb particles appear clause-finally and may be separated from the verb by a large amount of material. We hypothesised that this separation would encourage readers to predict the particle, and that high certainty would make prediction of a specific particle more likely than lower certainty. If a specific particle was predicted, this would reflect a strong commitment to sentence meaning that should incur a higher processing cost if the prediction is wrong. If a specific particle was less likely to be predicted, commitment should be weaker and the processing cost of a wrong prediction lower. If true, this could suggest that uncertainty discourages predictions via an unacceptable cost-benefit ratio. However, given the clear predictions made by the literature, it was surprisingly unclear whether the uncertainty manipulation affected the two ERP components studied, the N400 and the PNP. Bayes factor analyses showed that evidence for our a priori hypothesised effect sizes was inconclusive, although there was decisive evidence against a priori hypothesised effect sizes larger than 1μV for the N400 and larger than 3μV for the PNP. We attribute the inconclusive finding to the properties of verb-particle dependencies that differ from the verb-noun dependencies in which the N400 and PNP are often studied.
In 2019 the Journal of Memory and Language instituted an open data and code policy; this policy requires that, as a rule, code and data be released at the latest upon publication. How effective is this policy? We compared 59 papers published before, and 59 papers published after, the policy took effect. After the policy was in place, the rate of data sharing increased by more than 50%. We further looked at whether papers published under the open data policy were reproducible, in the sense that the published results should be possible to regenerate given the data, and given the code, when code was provided. For 8 out of the 59 papers, data sets were inaccessible. The reproducibility rate ranged from 34% to 56%, depending on the reproducibility criteria. The strongest predictor of whether an attempt to reproduce would be successful is the presence of the analysis code: it increases the probability of reproducing reported results by almost 40%. We propose two simple steps that can increase the reproducibility of published papers: share the analysis code, and attempt to reproduce one's own analysis using only the shared materials.
Intuitively, strongly constraining contexts should lead to stronger probabilistic representations of sentences in memory. Encountering unexpected words could therefore be expected to trigger costlier shifts in these representations than expected words. However, psycholinguistic measures commonly used to study probabilistic processing, such as the N400 event-related potential (ERP) component, are sensitive to word predictability but not to contextual constraint. Some research suggests that constraint-related processing cost may be measurable via an ERP positivity following the N400, known as the anterior post-N400 positivity (PNP). The PNP is argued to reflect update of a sentence representation and to be distinct from the posterior P600, which reflects conflict detection and reanalysis. However, constraint-related PNP findings are inconsistent. We sought to conceptually replicate Federmeier et al. (2007) and Kuperberg et al. (2020), who observed that the PNP, but not the N400 or the P600, was affected by constraint at unexpected but plausible words. Using a pre-registered design and statistical approach maximising power, we demonstrated a dissociated effect of predictability and constraint: strong evidence for predictability but not constraint in the N400 window, and strong evidence for constraint but not predictability in the later window. However, the constraint effect was consistent with a P600 and not a PNP, suggesting increased conflict between a strong representation and unexpected input rather than greater update of the representation. We conclude that either a simple strong/weak constraint design is not always sufficient to elicit the PNP, or that previous PNP constraint findings could be an artifact of smaller sample size.
What is the processing cost of being garden-pathed by a temporary syntactic ambiguity? We argue that comparing average reading times in garden-path versus non-garden-path sentences is not enough to answer this question. Trial-level contaminants such as inattention, the fact that garden pathing may occur non-deterministically in the ambiguous condition, and "triage" (rejecting the sentence without reanalysis; Fodor & Inoue, 2000) lead to systematic underestimates of the true cost of garden pathing. Furthermore, the "pure" garden-path effect due to encountering an unexpected word needs to be separated from the additional cost of syntactic reanalysis. To get more realistic estimates for the individual processing costs of garden pathing and syntactic reanalysis, we implement a novel computational model that includes trial-level contaminants as probabilistically occurring latent cognitive processes. The model shows a good predictive fit to existing reading time and judgment data. Furthermore, the latent-process approach captures differences between noun phrase/zero complement (NP/Z) garden-path sentences and semantically biased reduced relative clause (RRC) garden-path sentences: The NP/Z garden path occurs nearly deterministically but can be mostly eliminated by adding a comma. By contrast, the RRC garden path occurs with a lower probability, but disambiguation via semantic plausibility is not always effective.