97 research outputs found
Using Deep Neural Networks to Learn Syntactic Agreement
No full textWe consider the extent to which different deep neural network (DNN) configurations can learn syntactic relations, by taking up Linzen et al.’s (2016) work on subject-verb agreement with LSTM RNNs. We test their methods on a much larger corpus than they used (a ⇠24 million example part of the WaCky corpus, instead of their ⇠1.35 million example corpus, both drawn from Wikipedia). We experiment with several different DNN architectures (LSTM RNNs, GRUs, and CNNs), and alternative parameter settings for these systems (vocabulary size, training to test ratio, number of layers, memory size, drop out rate, and lexical embedding dimension size). We also try out our own unsupervised DNN language model. Our results are broadly compatible with those that Linzen et al. report. However, we discovered some interesting, and in some cases, surprising features of DNNs and language models in their performance of the agreement learning task. In particular, we found that DNNs require large vocabularies to form substantive lexical embeddings in order to learn structural patterns. This finding has interesting consequences for our understanding of the way in which DNNs represent syntactic information. It suggests that DNNs learn syntactic patterns more efficiently through rich lexical embeddings, with semantic as well as syntactic cues, than from training on lexically impoverished strings that highlight structural patterns
How much harder are hard garden path sentences than easy ones?
No full textThe advent of broad-coverage computational models of human sentence processing has made it possible to derive quantitative predictions for empirical phenomena of longstanding interest in psycholinguistics; one such case is the disambiguation difficulty in temporarily ambiguous sentences (garden-path sentences). Adequate evaluation of the accuracy of such quantitative predictions requires going beyond the classic binary distinction between "hard" and "easy" garden path sentences. It requires precise quantitative measurements of processing difficulty and statistical analyses that focus on more than just statistical significance. We evaluate how well a particular specification of surprisal theory predicts data from a self-paced reading study designed to estimate the magnitude of the disambiguation difficulty in two temporarily ambiguous sentence types (NP/Z and NP/S ambiguities). Using Bayesian analysis we conclude that our specification of surprisal theory cannot account for the observed NP/Z garden path effects. We have insufficient evidence to draw conclusions about whether it can account for the NP/S garden path effects
How much harder are hard garden path sentences than easy ones?
No full textThe advent of broad-coverage computational models of human sentence processing has made it possible to derive quantitative predictions for empirical phenomena of longstanding interest in psycholinguistics; one such case is the disambiguation difficulty in temporarily ambiguous sentences (garden-path sentences). Adequate evaluation of the accuracy of such quantitative predictions requires going beyond the classic binary distinction between "hard" and "easy" garden path sentences. It requires precise quantitative measurements of processing difficulty and statistical analyses that focus on more than just statistical significance. We evaluate how well a particular specification of surprisal theory predicts data from a self-paced reading study designed to estimate the magnitude of the disambiguation difficulty in two temporarily ambiguous sentence types (NP/Z and NP/S ambiguities). Using Bayesian analysis we conclude that our specification of surprisal theory cannot account for the observed NP/Z garden path effects. We have insufficient evidence to draw conclusions about whether it can account for the NP/S garden path effects
How much harder are hard garden path sentences than easy ones?
No full textThe advent of broad-coverage computational models of human sentence processing has made it possible to derive quantitative predictions for empirical phenomena of longstanding interest in psycholinguistics; one such case is the disambiguation difficulty in temporarily ambiguous sentences (garden-path sentences). Adequate evaluation of the accuracy of such quantitative predictions requires going beyond the classic binary distinction between "hard" and "easy" garden path sentences. It requires precise quantitative measurements of processing difficulty and statistical analyses that focus on more than just statistical significance. We evaluate how well a particular specification of surprisal theory predicts data from a self-paced reading study designed to estimate the magnitude of the disambiguation difficulty in two temporarily ambiguous sentence types (NP/Z and NP/S ambiguities). Using Bayesian analysis we conclude that our specification of surprisal theory cannot account for the observed NP/Z garden path effects. We have insufficient evidence to draw conclusions about whether it can account for the NP/S garden path effects
Recital de Joseph Matza en el Club Belalcázar de Bogotá
No full text- Sonata en La (G. F. Haendel)
- Chacona (J. S. Bach)
- Concierto Op. 35 en Re (P. I. Tschaikowski)
- Berceuse (Antonio María Valencia)
- Romanza (Antonio María Valencia)
- Moto perpetuo (Novacek)
- La fille aux cheveux de lin (C. Debussy)
- Danza de lasMedellín, Biblioteca Luis Echavarría Villegas, Sala de Patrimonio Documental, Colección Programas de manoBogotá. Club Belalcázar de Bogot
Regulation der ING5-Funktion durch Cyclin-abhängige Kinasen
No full textThe inhibitor of growth (ING) family consists of five members, ING1-ING5. These proteins are involved in different cellular processes, including transcription, replication and apoptosis. There have been reports suggesting that some members of the ING family are deregulated in different tumors. In support, reduced ING1 protein levels have been observed in tumors, suggesting a function as tumor suppressor. There are two different functions postulated for ING5. On the one hand ING5 binds to trimethylated lysine 4 of histone H3 (H3K4me3) in vitro and recruits the histone acetyltransferases HBO1 and MOZ/MORF and consequently seems to be involved in gene transcription. On the other hand it is also known that ING5 binds MCM2, thereby interacting with origins of replication. HBO1 is also involved in these complexes. Together these findings suggest that ING5 has functions in replication, some may overlap with activities relevant for transcription. Furthermore it has been postulated that some of these activities depend on the ability of ING5 to interact with the tumor suppressor p53. The aim of this thesis was to investigate in more detail the function and regulation of ING5. In particular the control of ING5 function by phosphorylation should be addressed and clarified. In preliminary work we identified ING5 as a novel substrate of Cyclin E/CDK2. Developing and applying phospho-specific antibody allowed me to validate the ING5 phoshorylation at Threonin 152 both in vitro and in cells. Moreover I found that this phosphorylation is regulated during the cell cycle and accumulates in the S- and G2-phases of the cell cycle. Knockdown experiments were performed to clarify the biological relevance of ING5. The reduction of ING5 protein levels induced apoptosis. This is in contrast to previous findings demonstrating that the knockdown of ING1 induces proliferation. Surprisingly the induction of apoptosis in response to the ING5 knockdown was p53 independent, despite the fact that ING5 interacts, albeit weakly, with p53. A number of links suggested at least a functional interaction of ING5 with the oncoprotein Myc. This included the correlation of Myc chromatin binding with H3K4me3 mark, the interaction of Myc with and phosphorylation by Cyclin E/CDK2, and the ability of Myc to induce apoptosis independently of p53. Therefore a possible connection between ING5 and Myc was analyzed in the regulation of apoptosis. I was able to demonstrate a direct interaction between ING5 and Myc. In addition I found that ING5 regulates Myc-mediated apoptosis. Furthermore the correlation of these proteins in transcription was investigated and ING5 binding at promoters of Myc target genes was identified. These findings suggest that ING5 functions as a cofactor for the regulation of Myc target genes, possibly those involved in the control of apoptosis
Analysing the potential of seq-to-seq models for incremental interpretation in task-oriented dialogue
No full textWe investigate how encoder-decoder models trained on a synthetic dataset of task-oriented dialogues process disfluencies, such as hesitations and self-corrections. We find that, contrary to earlier results, disfluencies have very little impact on the task success of seq-to-seq models with attention. Using visualisations and diagnostic classifiers, we analyse the representations that are incrementally built by the model, and discover that models develop little to no awareness of the structure of disfluencies. However, adding disfluencies to the data appears to help the model create clearer representations overall, as evidenced by the attention patterns the different models exhibit
- …
