Publications

Natural language generation (NLG) systems rely on corpora for both hand-crafted approaches in a traditional NLG architecture and for statistical end-to-end (learned) generation systems. Limitations in existing resources, however, make it difficult to develop systems which can vary the linguistic properties of an utterance as needed. For example, when users’ attention is split between a linguistic and a secondary task such as driving, a generation system may need to reduce the information density of an utterance to compensate for the reduction in user attention. We introduce a new corpus in the restaurant recommendation and comparison domain, collected in a paraphrasing paradigm, where subjects wrote texts targeting either a general audience or an elderly family member. This design resulted in a corpus of more than 5000 texts which exhibit a variety of lexical and syntactic choices and differ with respect to average word & sentence length and surprisal. The corpus includes two levels of meaning representation: flat ‘semantic stacks’ for propositional content and Rhetorical Structure Theory (RST) relations between these propositions.

• 2017-08_interspeech_extended-sparky-restaurant-corpus     (0.13MB)
• http://dx.doi.org/10.21437/Interspeech.2017-1555
• https://doi.org/10.21437/Interspeech.2017-1555

@inproceedings{Howcroft2017b,
title = {The Extended SPaRKy Restaurant Corpus: Designing a Corpus with Variable Information Density},
author = {David M. Howcroft and Dietrich Klakow and Vera Demberg},
url = {http://dx.doi.org/10.21437/Interspeech.2017-1555},
doi = {https://doi.org/10.21437/Interspeech.2017-1555},
year = {2017},
date = {2017-10-17},
booktitle = {Proc. Interspeech 2017},
pages = {3757-3761},
abstract = {Natural language generation (NLG) systems rely on corpora for both hand-crafted approaches in a traditional NLG architecture and for statistical end-to-end (learned) generation systems. Limitations in existing resources, however, make it difficult to develop systems which can vary the linguistic properties of an utterance as needed. For example, when users’ attention is split between a linguistic and a secondary task such as driving, a generation system may need to reduce the information density of an utterance to compensate for the reduction in user attention. We introduce a new corpus in the restaurant recommendation and comparison domain, collected in a paraphrasing paradigm, where subjects wrote texts targeting either a general audience or an elderly family member. This design resulted in a corpus of more than 5000 texts which exhibit a variety of lexical and syntactic choices and differ with respect to average word & sentence length and surprisal. The corpus includes two levels of meaning representation: flat ‘semantic stacks’ for propositional content and Rhetorical Structure Theory (RST) relations between these propositions.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Corpora of referring expressions elicited from human participants in a controlled environment are an important resource for research on automatic referring expression generation. We here present G-TUNA, a new corpus of referring expressions for German. Using the furniture stimuli set developed for the TUNA and D-TUNA corpora, our corpus extends on these corpora by providing data collected in a simulated driving dual-task setting, and additionally provides exact duration annotations for the spoken referring expressions. This corpus will hence allow researchers to analyze the interaction between referring expression length and speech rate, under conditions where the listener is under high vs. low cognitive load.

• http://www.aclweb.org/anthology/W17-3522
• https://doi.org/10.18653/v1/W17-3522

@inproceedings{W17-3522,
title = {G-TUNA: a corpus of referring expressions in German, including duration information},
author = {David M. Howcroft and Jorrig Vogels and Vera Demberg},
url = {http://www.aclweb.org/anthology/W17-3522},
doi = {https://doi.org/10.18653/v1/W17-3522},
year = {2017},
date = {2017},
booktitle = {Proceedings of the 10th International Conference on Natural Language Generation},
pages = {149-153},
publisher = {Association for Computational Linguistics},
address = {Santiago de Compostela, Spain},
abstract = {Corpora of referring expressions elicited from human participants in a controlled environment are an important resource for research on automatic referring expression generation. We here present G-TUNA, a new corpus of referring expressions for German. Using the furniture stimuli set developed for the TUNA and D-TUNA corpora, our corpus extends on these corpora by providing data collected in a simulated driving dual-task setting, and additionally provides exact duration annotations for the spoken referring expressions. This corpus will hence allow researchers to analyze the interaction between referring expression length and speech rate, under conditions where the listener is under high vs. low cognitive load.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

While previous research on readability has typically focused on document-level measures, recent work in areas such as natural language generation has pointed out the need of sentence-level readability measures. Much of psycholinguistics has focused for many years on processing measures that provide difficulty estimates on a word-by-word basis. However, these psycholinguistic measures have not yet been tested on sentence readability ranking tasks. In this paper, we use four psycholinguistic measures: idea density, surprisal, integration cost, and embedding depth to test whether these features are predictive of readability levels. We find that psycholinguistic features significantly improve performance by up to 3 percentage points over a standard document-level readability metric baseline.

• http://www.aclweb.org/anthology/E17-1090

@inproceedings{Howcroft2017,
title = {Psycholinguistic Models of Sentence Processing Improve Sentence Readability Ranking},
author = {David M. Howcroft and Vera Demberg},
url = {http://www.aclweb.org/anthology/E17-1090},
year = {2017},
date = {2017-10-17},
booktitle = {Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers},
pages = {958-968},
publisher = {Association for Computational Linguistics},
address = {Valencia, Spain},
abstract = {While previous research on readability has typically focused on document-level measures, recent work in areas such as natural language generation has pointed out the need of sentence-level readability measures. Much of psycholinguistics has focused for many years on processing measures that provide difficulty estimates on a word-by-word basis. However, these psycholinguistic measures have not yet been tested on sentence readability ranking tasks. In this paper, we use four psycholinguistic measures: idea density, surprisal, integration cost, and embedding depth to test whether these features are predictive of readability levels. We find that psycholinguistic features significantly improve performance by up to 3 percentage points over a standard document-level readability metric baseline.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Script knowledge plays a central role in text understanding and is relevant for a variety of downstream tasks. In this paper, we consider two recent datasets which provide a rich and general representation of script events in terms of paraphrase sets.

We introduce the task of mapping event mentions in narrative texts to such script event types, and present a model for this task that exploits rich linguistic representations as well as information on temporal ordering. The results of our experiments demonstrate that this complex task is indeed feasible.

• http://www.aclweb.org/anthology/S17-1016

@inproceedings{ostermann-EtAl:2017:starSEM,
title = {Aligning Script Events with Narrative Texts},
author = {Michael Roth and Stefan Thater andSimon Ostermann and Manfred Pinkal},
url = {http://www.aclweb.org/anthology/S17-1016},
year = {2017},
date = {2017-10-17},
booktitle = {Proceedings of the 6th Joint Conference on Lexical and Computational Semantics (*SEM 2017)},
publisher = {Association for Computational Linguistics},
address = {Vancouver, Canada},
abstract = {Script knowledge plays a central role in text understanding and is relevant for a variety of downstream tasks. In this paper, we consider two recent datasets which provide a rich and general representation of script events in terms of paraphrase sets. We introduce the task of mapping event mentions in narrative texts to such script event types, and present a model for this task that exploits rich linguistic representations as well as information on temporal ordering. The results of our experiments demonstrate that this complex task is indeed feasible.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Word embeddings are now a standard technique for inducing meaning representations for words. For getting good representations, it is important to take into account different senses of a word. In this paper, we propose a mixture model for learning multi-sense word embeddings.

Our model generalizes the previous works in that it allows to induce different weights of different senses of a word. The experimental results show that our model outperforms previous models on standard evaluation tasks.

• http://www.aclweb.org/anthology/S17-1015

@inproceedings{nguyen-EtAl:2017:starSEM,
title = {A Mixture Model for Learning Multi-Sense Word Embeddings},
author = {Dai Quoc Nguyen and Dat Quoc Nguyen and Ashutosh Modi and Stefan Thater and Manfred Pinkal},
url = {http://www.aclweb.org/anthology/S17-1015},
year = {2017},
date = {2017},
pages = {121-127},
publisher = {Association for Computational Linguistics},
address = {Vancouver, Canada},
abstract = {Word embeddings are now a standard technique for inducing meaning representations for words. For getting good representations, it is important to take into account different senses of a word. In this paper, we propose a mixture model for learning multi-sense word embeddings. Our model generalizes the previous works in that it allows to induce different weights of different senses of a word. The experimental results show that our model outperforms previous models on standard evaluation tasks.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

This paper presents an approach to the task of predicting an event description from a preceding sentence in a text. Our approach explores sequence-to-sequence learning using a bidirectional multi-layer recurrent neural network. Our approach substantially outperforms previous work in terms of the BLEU score on two datasets derived from WikiHow and DeScript respectively.

Since the BLEU score is not easy to interpret as a measure of event prediction, we complement our study with a second evaluation that exploits the rich linguistic annotation of gold paraphrase sets of events.

• I17-2007     (0.16MB)
• http://www.aclweb.org/anthology/I17-2007

@inproceedings{nguyen-EtAl:2017:I17-2,
title = {Sequence to Sequence Learning for Event Prediction},
author = {Dai Quoc Nguyen and Dat Quoc Nguyen and Cuong Xuan Chu and Stefan Thater and Manfred Pinkal},
url = {http://www.aclweb.org/anthology/I17-2007},
year = {2017},
date = {2017-10-17},
booktitle = {Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 2: Short Papers)},
pages = {37-42},
publisher = {Asian Federation of Natural Language Processing},
address = {Taipei, Taiwan},
abstract = {This paper presents an approach to the task of predicting an event description from a preceding sentence in a text. Our approach explores sequence-to-sequence learning using a bidirectional multi-layer recurrent neural network. Our approach substantially outperforms previous work in terms of the BLEU score on two datasets derived from WikiHow and DeScript respectively. Since the BLEU score is not easy to interpret as a measure of event prediction, we complement our study with a second evaluation that exploits the rich linguistic annotation of gold paraphrase sets of events.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

@inproceedings{Tourtourietal2017a,
title = {Overspecifications efficiently manage referential entropy in situated communication},
author = {Elli Tourtouri and Francesca Delogu and Matthew W. Crocker},
year = {2017},
date = {2017},
booktitle = {Paper presented at the 39th Annual Conference of the German Linguistic Society (DGfS)},
publisher = {Saarland University},
address = {Saarbruecken, Germany},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

• https://www.sciencedirect.com/science/article/pii/S0010027716303122
• https://doi.org/10.1016/j.cognition.2016.12.017

@article{Brouwer2017,
title = {Teasing apart coercion and surprisal: Evidence from ERPs and eye-movements},
author = {Francesca Delogu and Matthew W. Crocker and Heiner Drenhaus},
url = {https://www.sciencedirect.com/science/article/pii/S0010027716303122},
doi = {https://doi.org/10.1016/j.cognition.2016.12.017},
year = {2017},
date = {2017},
journal = {Cognition},
pages = {46-59},
volume = {161},
abstract = {

Copy BibTeX to Clipboard

The study of language is ultimately about meaning: how can meaning be constructed from linguistic signal, and how can it be represented? he human language comprehension system is highly eicient and accurate at atributing meaning to linguistic input. Hence, in trying to identify computational principles and representations for meaning construction, we should consider how these could be implemented at the neural level in the brain. Here, we introduce a framework for such a neural semantics. his framework ofers meaning representations that are neurally plausible (can be implemented in neural hardware), expressive (capture negation, quantiication, and modality), compositional (capture complex propositional meaning as the sum of its parts), graded (are probabilistic in nature), and inferential (allow for inferences beyond literal propositional content). Moreover, it is shown how these meaning representations can be constructed incrementally, on a word-by-word basis in a neurocomputational model of language processing.

• https://research.rug.nl/en/publications/from-semantics-to-dialectometry-festschrift-in-honor-of-john-nerb

@article{Brouwer2017b,
title = {Neural semantics},
author = {Harm Brouwer and Matthew W. Crocker and Noortje Venhuizen},
url = {https://research.rug.nl/en/publications/from-semantics-to-dialectometry-festschrift-in-honor-of-john-nerb},
year = {2017},
date = {2017},
journal = {From Semantics to Dialectometry: Festschrift in Honour of John Nerbonne},
pages = {75-83},
abstract = {The study of language is ultimately about meaning: how can meaning be constructed from linguistic signal, and how can it be represented? he human language comprehension system is highly eicient and accurate at atributing meaning to linguistic input. Hence, in trying to identify computational principles and representations for meaning construction, we should consider how these could be implemented at the neural level in the brain. Here, we introduce a framework for such a neural semantics. his framework ofers meaning representations that are neurally plausible (can be implemented in neural hardware), expressive (capture negation, quantiication, and modality), compositional (capture complex propositional meaning as the sum of its parts), graded (are probabilistic in nature), and inferential (allow for inferences beyond literal propositional content). Moreover, it is shown how these meaning representations can be constructed incrementally, on a word-by-word basis in a neurocomputational model of language processing.},
pubstate = {published},
type = {article}
}

Copy BibTeX to Clipboard

Event-Related Potentials (ERPs)—stimulus-locked, scalp-recorded voltage fluctuations caused by post-synaptic neural activity—have proven invaluable to the study of language comprehension. Of interest in the ERP signal are systematic, reoccurring voltage fluctuations called components, which are taken to reflect the neural activity underlying specific computational operations carried out in given neuroanatomical networks (cf. Näätänen and Picton, 1987). For language processing, the N400 component and the P600 component are of particular salience (see Kutas et al., 2006, for a review). The typical approach to determining whether a target word in a sentence leads to differential modulation of these components, relative to a control word, is to look for effects on mean amplitude in predetermined time-windows on the respective ERP waveforms, e.g., 350–550 ms for the N400 component and 600–900 ms for the P600 component. The common mode of operation in psycholinguistics, then, is to tabulate the presence/absence of N400- and/or P600-effects across studies, and to use this categorical data to inform neurocognitive models that attribute specific functional roles to the N400 and P600 component (see Kuperberg, 2007; Bornkessel-Schlesewsky and Schlesewsky, 2008; Brouwer et al., 2012, for reviews).

Here, we assert that this Waveform-based Component Structure (WCS) approach to ERPs leads to inconsistent data patterns, and hence, misinforms neurocognitive models of the electrophysiology of language processing. The reason for this is that the WCS approach ignores the latent component structure underlying ERP waveforms (cf. Luck, 2005), thereby leading to conclusions about component structure that do not factor in spatiotemporal component overlap of the N400 and the P600. This becomes particularly problematic when spatiotemporal component overlap interacts with differential P600 modulations due to task demands (cf. Kolk et al., 2003). While the problem of spatiotemporal component overlap is generally acknowledged, and occasionally invoked to account for within-study inconsistencies in the data, its implications are often overlooked in psycholinguistic theorizing that aims to integrate findings across studies. We believe WCS-centric theorizing to be the single largest reason for the lack of convergence regarding the processes underlying the N400 and the P600, thereby seriously hindering the advancement of neurocognitive theories and models of language processing.

• https://www.frontiersin.org/articles/10.3389/fpsyg.2017.01327/full
• https://doi.org/10.3389/fpsyg.2017.01327

@article{Brouwer2017,
title = {On the proper treatment of the N400 and P600 in Language comprehension},
author = {Harm Brouwer and Matthew W. Crocker},
url = {https://www.frontiersin.org/articles/10.3389/fpsyg.2017.01327/full},
doi = {https://doi.org/10.3389/fpsyg.2017.01327},
year = {2017},
date = {2017},
journal = {Frontiers in Psychology},
volume = {8},
abstract = {

Event-Related Potentials (ERPs)—stimulus-locked, scalp-recorded voltage fluctuations caused by post-synaptic neural activity—have proven invaluable to the study of language comprehension. Of interest in the ERP signal are systematic, reoccurring voltage fluctuations called components, which are taken to reflect the neural activity underlying specific computational operations carried out in given neuroanatomical networks (cf. N{\"a}{\"a}t{\"a}nen and Picton, 1987). For language processing, the N400 component and the P600 component are of particular salience (see Kutas et al., 2006, for a review). The typical approach to determining whether a target word in a sentence leads to differential modulation of these components, relative to a control word, is to look for effects on mean amplitude in predetermined time-windows on the respective ERP waveforms, e.g., 350–550 ms for the N400 component and 600–900 ms for the P600 component. The common mode of operation in psycholinguistics, then, is to tabulate the presence/absence of N400- and/or P600-effects across studies, and to use this categorical data to inform neurocognitive models that attribute specific functional roles to the N400 and P600 component (see Kuperberg, 2007; Bornkessel-Schlesewsky and Schlesewsky, 2008; Brouwer et al., 2012, for reviews).

Here, we assert that this Waveform-based Component Structure (WCS) approach to ERPs leads to inconsistent data patterns, and hence, misinforms neurocognitive models of the electrophysiology of language processing. The reason for this is that the WCS approach ignores the latent component structure underlying ERP waveforms (cf. Luck, 2005), thereby leading to conclusions about component structure that do not factor in spatiotemporal component overlap of the N400 and the P600. This becomes particularly problematic when spatiotemporal component overlap interacts with differential P600 modulations due to task demands (cf. Kolk et al., 2003). While the problem of spatiotemporal component overlap is generally acknowledged, and occasionally invoked to account for within-study inconsistencies in the data, its implications are often overlooked in psycholinguistic theorizing that aims to integrate findings across studies. We believe WCS-centric theorizing to be the single largest reason for the lack of convergence regarding the processes underlying the N400 and the P600, thereby seriously hindering the advancement of neurocognitive theories and models of language processing.

Copy BibTeX to Clipboard

• https://www.researchgate.net/publication/320988782_The_Influence_of_Script_Knowledge_on_Language_Processing_Evidence_from_ERPs

@miscellaneous{Rabs2017,
title = {The influence of script knowledge on language processing: Evidence from ERPs},
author = {Elisabeth Rabs and Heiner Drenhaus and Francesca Delogu and Matthew W. Crocker},
url = {https://www.researchgate.net/publication/320988782_The_Influence_of_Script_Knowledge_on_Language_Processing_Evidence_from_ERPs},
year = {2017},
date = {2017},
publisher = {23rd AMLaP Conference},
address = {Lancaster, UK},
abstract = {

Copy BibTeX to Clipboard

• https://www.researchgate.net/publication/319543522_The_influence_of_lexical_priming_versus_event_knowledge_on_the_N400_and_the_P600

@miscellaneous{Delogu2017b,
title = {The influence of lexical priming versus event knowledge on the N400 and the P600},
author = {Francesca Delogu and Harm Brouwer and Matthew W. Crocker},
url = {https://www.researchgate.net/publication/319543522_The_influence_of_lexical_priming_versus_event_knowledge_on_the_N400_and_the_P600},
year = {2017},
date = {2017},
publisher = {23rd AMLaP Conference},
address = {Lancaster, UK},
abstract = {

Copy BibTeX to Clipboard

A common reason for errors in coreference resolution is the lack of semantic information to help determine the compatibility between mentions referring to the same entity. Distributed representations, which have been shown successful in encoding relatedness between words, could potentially be a good source of such knowledge. Moreover, being obtained in an unsupervised manner, they could help address data sparsity issues in labeled training data at a small cost. In this work we investigate whether and to what extend features derived from word embeddings can be successfully used for supervised coreference resolution. We experiment with several word embedding models, and several different types of embeddingbased features, including embedding cluster and cosine similarity-based features. Our evaluations show improvements in the performance of a supervised state-of-theart coreference system.

• https://aclanthology.org/R17-1088/
• https://doi.org/10.26615/978-954-452-049-6_088

@inproceedings{simova:2017,
title = {Word Embeddings as Features for Supervised Coreference Resolution},
author = {Iliana Simova and Hans Uszkoreit},
url = {https://aclanthology.org/R17-1088/},
doi = {https://doi.org/10.26615/978-954-452-049-6_088},
year = {2017},
date = {2017},
booktitle = {Proceedings of Recent Advances in Natural Language Processing},
pages = {686-693},
publisher = {INCOMA Ltd.},
address = {Varna, Bulgaria},
abstract = {A common reason for errors in coreference resolution is the lack of semantic information to help determine the compatibility between mentions referring to the same entity. Distributed representations, which have been shown successful in encoding relatedness between words, could potentially be a good source of such knowledge. Moreover, being obtained in an unsupervised manner, they could help address data sparsity issues in labeled training data at a small cost. In this work we investigate whether and to what extend features derived from word embeddings can be successfully used for supervised coreference resolution. We experiment with several word embedding models, and several different types of embeddingbased features, including embedding cluster and cosine similarity-based features. Our evaluations show improvements in the performance of a supervised state-of-theart coreference system.},
keywords = {B5, sfb 1102},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

The MaryTTS system is a modular text-to-speech (TTS) system which has been developed for nearly 20 years. This paper describes the MaryTTS entry for the Blizzard Challenge 2017. In contrast to last year’s MaryTTS system, based on a unit selection baseline using the latest stable MaryTTS version, the basis for this year’s system is a new, experimental version with a completely redesigned architecture.

• MaryTTS_Blizzard2017     (0.2MB)
• http://mary.dfki.de/documentation/publications/index.html

@inproceedings{LeMaguer2017BC,
title = {The "Uprooted" MaryTTS Entry for the Blizzard Challenge 2017},
author = {S{\'e}bastien Le Maguer and Ingmar Steiner},
url = {http://mary.dfki.de/documentation/publications/index.html},
year = {2017},
date = {2017},
booktitle = {Blizzard Challenge},
address = {Stockholm, Sweden},
abstract = {The MaryTTS system is a modular text-to-speech (TTS) system which has been developed for nearly 20 years. This paper describes the MaryTTS entry for the Blizzard Challenge 2017. In contrast to last year’s MaryTTS system, based on a unit selection baseline using the latest stable MaryTTS version, the basis for this year’s system is a new, experimental version with a completely redesigned architecture.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

To shed light on the question whether humans converge phonetically to synthesized speech, a shadowing experiment was conducted using three different types of stimuli – natural speaker, diphone synthesis, and HMM synthesis. Three segment-level phonetic features of German that are well-known to vary across native speakers were examined. The first feature triggered convergence in roughly one third of the cases for all stimulus types. The second feature showed generally a small amount of convergence, which may be due to the nature of the feature itself. Still the effect was strongest for the natural stimuli, followed by the HMM stimuli and weakest for the diphone stimuli. The effect of the third feature was clearly observable for the natural stimuli and less pronounced in the synthetic stimuli. This is presumably a result of the partly insufficient perceptibility of this target feature in the synthetic stimuli and demonstrates the necessity of gaining fine-grained control over the synthesis output, should it be intended to implement capabilities of phonetic convergence on the segmental level in spoken dialogue systems

• 1433     (0.31MB)
• https://publikationen.sulb.uni-saarland.de/handle/20.500.11880/29623

@inproceedings{Gessinger2017IS,
title = {Shadowing Synthesized Speech - Segmental Analysis of Phonetic Convergence},
author = {Iona Gessinger and Eran Raveh and S{\'e}bastien Le Maguer and Bernd M{\"o}bius and Ingmar Steiner},
url = {https://publikationen.sulb.uni-saarland.de/handle/20.500.11880/29623},
year = {2017},
date = {2017},
booktitle = {Interspeech},
pages = {3797-3801},
address = {Stockholm, Sweden},
abstract = {To shed light on the question whether humans converge phonetically to synthesized speech, a shadowing experiment was conducted using three different types of stimuli – natural speaker, diphone synthesis, and HMM synthesis. Three segment-level phonetic features of German that are well-known to vary across native speakers were examined. The first feature triggered convergence in roughly one third of the cases for all stimulus types. The second feature showed generally a small amount of convergence, which may be due to the nature of the feature itself. Still the effect was strongest for the natural stimuli, followed by the HMM stimuli and weakest for the diphone stimuli. The effect of the third feature was clearly observable for the natural stimuli and less pronounced in the synthetic stimuli. This is presumably a result of the partly insufficient perceptibility of this target feature in the synthetic stimuli and demonstrates the necessity of gaining fine-grained control over the synthesis output, should it be intended to implement capabilities of phonetic convergence on the segmental level in spoken dialogue systems},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

We present an end-to-end text-to-speech (TTS) synthesis system that generates audio and synchronized tongue motion directly from text. This is achieved by adapting a statistical shape space model of the tongue surface to an articulatory speech corpus and training a speech synthesis system directly on the tongue model parameter weights. We focus our analysis on the application of two standard methodologies, based on Hidden Markov Models (HMMs) and Deep Neural Networks (DNNs), respectively, to train both acoustic models and the tongue model parameter weights. We evaluate both methodologies at every step by comparing the predicted articulatory movements against the reference data. The results show that even with less than 2h of data, DNNs already outperform HMMs.

• 0936     (0.26MB)
• https://www.isca-speech.org/archive/interspeech_2017/maguer17_interspeech.html
• https://doi.org/10.21437/Interspeech.2017-936

@inproceedings{LeMaguer2017IS,
title = {An HMM/DNN comparison for synchronized text-to-speech and tongue motion synthesis},
author = {S{\'e}bastien Le Maguer and Ingmar Steiner and Alexander Hewer},
url = {https://www.isca-speech.org/archive/interspeech_2017/maguer17_interspeech.html},
doi = {https://doi.org/10.21437/Interspeech.2017-936},
year = {2017},
date = {2017},
booktitle = {Proc. Interspeech 2017},
pages = {239-243},
address = {Stockholm, Sweden},
abstract = {We present an end-to-end text-to-speech (TTS) synthesis system that generates audio and synchronized tongue motion directly from text. This is achieved by adapting a statistical shape space model of the tongue surface to an articulatory speech corpus and training a speech synthesis system directly on the tongue model parameter weights. We focus our analysis on the application of two standard methodologies, based on Hidden Markov Models (HMMs) and Deep Neural Networks (DNNs), respectively, to train both acoustic models and the tongue model parameter weights. We evaluate both methodologies at every step by comparing the predicted articulatory movements against the reference data. The results show that even with less than 2h of data, DNNs already outperform HMMs.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

• https://www.researchgate.net/publication/320979082_The_P600_-_not_the_N400_-_indexes_semantic_integration

@inproceedings{Delogu2017c,
title = {The P600 - not the N400 - indexes semantic integration},
author = {Francesca Delogu and Harm Brouwer and Matthew W. Crocker},
url = {https://www.researchgate.net/publication/320979082_The_P600_-_not_the_N400_-_indexes_semantic_integration},
year = {2017},
date = {2017},
publisher = {9th Annual Meeting of the Society for the Neurobiology of Language (SNL)},
address = {Baltimore, US},
abstract = {

Copy BibTeX to Clipboard

Training large vocabulary Neural Network Language Models (NNLMs) is a difficult task due to the explicit requirement of the output layer normalization, which typically involves the evaluation of the full softmax function over the complete vocabulary. This paper proposes a Batch Noise Contrastive Estimation (B-NCE) approach to alleviate this problem. This is achieved by reducing the vocabulary, at each time step, to the target words in the batch and then replacing the softmax by the noise contrastive estimation approach, where these words play the role of targets and noise samples at the same time. In doing so, the proposed approach can be fully formulated and implemented using optimal dense matrix operations. Applying B-NCE to train different NNLMs on the Large Text Compression Benchmark (LTCB) and the One Billion Word Benchmark (OBWB) shows a significant reduction of the training time with no noticeable degradation of the models performance. This paper also presents a new baseline comparative study of different standard NNLMs on the large OBWB on a single Titan-X GPU.

• https://arxiv.org/abs/1708.05997

@inproceedings{Oualil2017,
title = {A batch noise contrastive estimation approach for training large vocabulary language models},
author = {Youssef Oualil and Dietrich Klakow},
url = {https://arxiv.org/abs/1708.05997},
year = {2017},
date = {2017},
publisher = {18th Annual Conference of the International Speech Communication Association (INTERSPEECH)},
abstract = {Training large vocabulary Neural Network Language Models (NNLMs) is a difficult task due to the explicit requirement of the output layer normalization, which typically involves the evaluation of the full softmax function over the complete vocabulary. This paper proposes a Batch Noise Contrastive Estimation (B-NCE) approach to alleviate this problem. This is achieved by reducing the vocabulary, at each time step, to the target words in the batch and then replacing the softmax by the noise contrastive estimation approach, where these words play the role of targets and noise samples at the same time. In doing so, the proposed approach can be fully formulated and implemented using optimal dense matrix operations. Applying B-NCE to train different NNLMs on the Large Text Compression Benchmark (LTCB) and the One Billion Word Benchmark (OBWB) shows a significant reduction of the training time with no noticeable degradation of the models performance. This paper also presents a new baseline comparative study of different standard NNLMs on the large OBWB on a single Titan-X GPU.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

This study investigated whether German vowels differ significantly from each other in mel-cepstral distortion (MCD) when they stand in different information density (ID) contexts. We hypothesized that vowels in the same ID contexts are more similar to each other than vowels that stand in different ID conditions. Read speech material from PhonDat2 of 16 German natives (m = 10, f = 6) was analyzed. Bi-phone and word language models were calculated based on DeWaC. To account for additional variability in the data, prosodic factors, as well as corpusspecific frequency values were also entered into the statistical models. Results showed that vowels in different ID conditions were significantly different in their MCD values. Unigram word probability and corpus-specific word frequency showed the expected effect on vowel similarity with a hierarchy between noncontrasting and contrasting conditions. However, these did not form a homogeneous group since there were group-internal significant differences. The largest distance can be found between vowels produced at fast speech rate, and between unstressed vowels.

• brandt_etal_is2017-1-1     (0.2MB)
• https://www.researchgate.net/publication/319185343_Mel-Cepstral_Distortion_of_German_Vowels_in_Different_Information_Density_Contexts

@inproceedings{Brandt/etal:2017,
title = {Mel-cepstral distortion of German vowels in different information density contexts},
author = {Erika Brandt and Frank Zimmerer and Bernd M{\"o}bius and Bistra Andreeva},
url = {https://www.researchgate.net/publication/319185343_Mel-Cepstral_Distortion_of_German_Vowels_in_Different_Information_Density_Contexts},
year = {2017},
date = {2017},
booktitle = {Proceedings of Interspeech},
address = {Stockholm, Sweden},
abstract = {This study investigated whether German vowels differ significantly from each other in mel-cepstral distortion (MCD) when they stand in different information density (ID) contexts. We hypothesized that vowels in the same ID contexts are more similar to each other than vowels that stand in different ID conditions. Read speech material from PhonDat2 of 16 German natives (m = 10, f = 6) was analyzed. Bi-phone and word language models were calculated based on DeWaC. To account for additional variability in the data, prosodic factors, as well as corpusspecific frequency values were also entered into the statistical models. Results showed that vowels in different ID conditions were significantly different in their MCD values. Unigram word probability and corpus-specific word frequency showed the expected effect on vowel similarity with a hierarchy between noncontrasting and contrasting conditions. However, these did not form a homogeneous group since there were group-internal significant differences. The largest distance can be found between vowels produced at fast speech rate, and between unstressed vowels.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

We present the Fragment Corpus (FraC), a corpus for the investigation of fragments (see Morgan 1973), i.e. incomplete sentences, in German. The corpus is a mixed-register corpus and consists of 17 different text types including written (newspaper texts, legal texts, blogs etc.) and spoken texts (dialogues, interviews, radio moderations etc.), as well as social media texts (tweets, sms, chats). Each subcorpus comprises approx. 2000 utterance units (including fragments, following the orthographic notion of sentence) which amounts to a total corpus size of 380K tokens. The data was taken from electronically available sources.

• paper194     (0.05MB)
• https://publikationen.sulb.uni-saarland.de/handle/20.500.11880/30290

@inproceedings{HorchReich:17,
title = {The Fragment Corpus},
author = {Eva Horch and Ingo Reich},
url = {https://publikationen.sulb.uni-saarland.de/handle/20.500.11880/30290},
year = {2017},
date = {2017},
booktitle = {Proceedings of the 9th International Corpus Linguistics Conference},
pages = {392-393},
address = {Birmingham, UK},
abstract = {We present the Fragment Corpus (FraC), a corpus for the investigation of fragments (see Morgan 1973), i.e. incomplete sentences, in German. The corpus is a mixed-register corpus and consists of 17 different text types including written (newspaper texts, legal texts, blogs etc.) and spoken texts (dialogues, interviews, radio moderations etc.), as well as social media texts (tweets, sms, chats). Each subcorpus comprises approx. 2000 utterance units (including fragments, following the orthographic notion of sentence) which amounts to a total corpus size of 380K tokens. The data was taken from electronically available sources.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard