Publications

Recently neural network based approaches to knowledge-intensive NLP tasks, such as question answering, started to rely heavily on the combination of neural retrievers and readers. Retrieval is typically performed over a large textual knowledge base (KB) which requires significant memory and compute resources, especially when scaled up. On HotpotQA we systematically investigate reducing the size of the KB index by means of dimensionality (sparse random projections, PCA, autoencoders) and numerical precision reduction.
Our results show that PCA is an easy solution that requires very little data and is only slightly worse than autoencoders, which are less stable. All methods are sensitive to pre- and post-processing and data should always be centered and normalized both before and after dimension reduction. Finally, we show that it is possible to combine PCA with using 1bit per dimension. Overall we achieve (1) 100× compression with 75%, and (2) 24× compression with 92% original retrieval performance.

• Paper     (0.69MB)
• https://arxiv.org/abs/2204.02906

@inproceedings{Zouhar_2022_Base,
title = {Knowledge Base Index Compression via Dimensionality and Precision Reduction},
author = {Vil{\'e}m Zouhar and Marius Mosbach and Miaoran Zhang and Dietrich Klakow},
url = {https://arxiv.org/abs/2204.02906},
year = {2022},
date = {2022},
publisher = {Spa-NLP workshop at ACL 2022},
address = {22nd-27th May 2022 Dublin, Ireland},
abstract = {Recently neural network based approaches to knowledge-intensive NLP tasks, such as question answering, started to rely heavily on the combination of neural retrievers and readers. Retrieval is typically performed over a large textual knowledge base (KB) which requires significant memory and compute resources, especially when scaled up. On HotpotQA we systematically investigate reducing the size of the KB index by means of dimensionality (sparse random projections, PCA, autoencoders) and numerical precision reduction. Our results show that PCA is an easy solution that requires very little data and is only slightly worse than autoencoders, which are less stable. All methods are sensitive to pre- and post-processing and data should always be centered and normalized both before and after dimension reduction. Finally, we show that it is possible to combine PCA with using 1bit per dimension. Overall we achieve (1) 100× compression with 75%, and (2) 24× compression with 92% original retrieval performance.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Cross-lingual natural language processing relies on translation, either by humans or machines, at different levels, from translating training data to translating test sets. However, compared to original texts in the same language, translations possess distinct qualities referred to as translationese. Previous research has shown that these translation artifacts influence the performance of a variety of cross-lingual tasks. In this work, we propose a novel approach to reducing translationese by extending an established bias-removal technique. We use the Iterative Null-space Projection (INLP) algorithm, and show by measuring classification accuracy before and after debiasing, that translationese is reduced at both sentence and word level. We evaluate the utility of debiasing translationese on a natural language inference (NLI) task, and show that by reducing this bias, NLI accuracy improves. To the best of our knowledge, this is the first study to debias translationese as represented in latent embedding space.

• 2022.naacl-main.292     (1.11MB)
• https://aclanthology.org/2022.naacl-main.292/
• https://doi.org/10.18653/v1/2022.naacl-main.292

@inproceedings{Chowdhury_2022_Debiasing,
title = {Towards Debiasing Translation Artifacts},
author = {Koel Dutta Chowdhury and Rricha Jalota and Josef van Genabith and Cristina Espa{\~n}a-Bonet},
url = {https://aclanthology.org/2022.naacl-main.292/},
doi = {https://doi.org/10.18653/v1/2022.naacl-main.292},
year = {2022},
date = {2022},
booktitle = {Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies},
pages = {3983-3991},
publisher = {Association for Computational Linguistics},
address = {Seattle, United States, July 2022},
abstract = {Cross-lingual natural language processing relies on translation, either by humans or machines, at different levels, from translating training data to translating test sets. However, compared to original texts in the same language, translations possess distinct qualities referred to as translationese. Previous research has shown that these translation artifacts influence the performance of a variety of cross-lingual tasks. In this work, we propose a novel approach to reducing translationese by extending an established bias-removal technique. We use the Iterative Null-space Projection (INLP) algorithm, and show by measuring classification accuracy before and after debiasing, that translationese is reduced at both sentence and word level. We evaluate the utility of debiasing translationese on a natural language inference (NLI) task, and show that by reducing this bias, NLI accuracy improves. To the best of our knowledge, this is the first study to debias translationese as represented in latent embedding space.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Experimental pragmatics tells us that a metaphor conveys salient features of a vehicle and that highly typical featurestend to be salient. But can highly atypical features also be salient? When asking if John is loyal and hearing “John is afox”, will the hearer conclude that John is disloyal because loyalty is saliently atypical for a fox? This prediction followsfrom our RSA-based model of metaphor understanding which relies on gradient salience. Our behavioral experimentscorroborate the model’s predictions, providing evidence that high and low typicality are salient and result in high in-terpretation confidence and agreement, while average typicality is not salient and makes a metaphor confusing. Ourmodel implements the idea that other features of a vehicle, along with possible alternative vehicles, influence metaphorinterpretation. It produces a significantly better fit compared to an existing RSA model of metaphor understanding,supporting our predictions about the factors at play.

• eScholarship UC item 7kq207zs     (0.07MB)
• https://escholarship.org/uc/item/7kq207zs

@inproceedings{Mayn_2022_of,
title = {Pragmatics of Metaphor Revisited: Modeling the Role of Degree and Salience in Metaphor Understanding},
author = {Alexandra Mayn and Vera Demberg},
url = {https://escholarship.org/uc/item/7kq207zs},
year = {2022},
date = {2022},
booktitle = {Proceedings of the Annual Meeting of the Cognitive Science Society, 43(43)},
pages = {3178ff.},
publisher = {CogSci2022},
abstract = {Experimental pragmatics tells us that a metaphor conveys salient features of a vehicle and that highly typical featurestend to be salient. But can highly atypical features also be salient? When asking if John is loyal and hearing “John is afox”, will the hearer conclude that John is disloyal because loyalty is saliently atypical for a fox? This prediction followsfrom our RSA-based model of metaphor understanding which relies on gradient salience. Our behavioral experimentscorroborate the model’s predictions, providing evidence that high and low typicality are salient and result in high in-terpretation confidence and agreement, while average typicality is not salient and makes a metaphor confusing. Ourmodel implements the idea that other features of a vehicle, along with possible alternative vehicles, influence metaphorinterpretation. It produces a significantly better fit compared to an existing RSA model of metaphor understanding,supporting our predictions about the factors at play.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Empirical studies have demonstrated that when comprehenders are faced with informationally redundant utterances, they may make pragmatic inferences (Kravtchenko & Demberg, 2015). Previous work has also shown that the strength of these inferences depends on prominence of the redundant utterance – if it is stressed prosodically, marked with an exclamation mark, or introduced with a discourse marker such as “Oh yeah”, atypicality inferences are stronger (Kravtchenko & Demberg, 2015, 2022; Ryzhova & Demberg, 2020). The goal of the present paper is to demonstrate how both the atypicality inference and the effect of prominence can be modelled using the rational speech act (RSA) framework. We show that atypicality inferences can be captured by introducing joint reasoning about the habituality of events, following Degen, Tessler, and Goodman (2015); Goodman and Frank (2016). However, we find that joint reasoning models principally cannot account for the effect of differences in utterance prominence. This is because prominence markers do not contribute to the truth-conditional meaning. We then proceed to demonstrate that leveraging a noisy channel model, which has previously been used to model low-level acoustic perception (Bergen & Goodman, 2015), can successfully account for the empirically observed patterns of utterance prominence.

• eScholarship UC item 7630p08b     (0.24MB)
• https://escholarship.org/uc/item/7630p08b

@article{Kravtchenko_2022_atypicality,
title = {Modeling atypicality inferences in pragmatic reasoning},
author = {Ekaterina Kravtchenko and Vera Demberg},
url = {https://escholarship.org/uc/item/7630p08b},
year = {2022},
date = {2022},
journal = {Proceedings of the Annual Meeting of the Cognitive Science Society},
pages = {1918-1924},
publisher = {CogSci 2022},
address = {Toronto, Canada},
volume = {44},
number = {44},
abstract = {Empirical studies have demonstrated that when comprehenders are faced with informationally redundant utterances, they may make pragmatic inferences (Kravtchenko & Demberg, 2015). Previous work has also shown that the strength of these inferences depends on prominence of the redundant utterance – if it is stressed prosodically, marked with an exclamation mark, or introduced with a discourse marker such as “Oh yeah”, atypicality inferences are stronger (Kravtchenko & Demberg, 2015, 2022; Ryzhova & Demberg, 2020). The goal of the present paper is to demonstrate how both the atypicality inference and the effect of prominence can be modelled using the rational speech act (RSA) framework. We show that atypicality inferences can be captured by introducing joint reasoning about the habituality of events, following Degen, Tessler, and Goodman (2015); Goodman and Frank (2016). However, we find that joint reasoning models principally cannot account for the effect of differences in utterance prominence. This is because prominence markers do not contribute to the truth-conditional meaning. We then proceed to demonstrate that leveraging a noisy channel model, which has previously been used to model low-level acoustic perception (Bergen & Goodman, 2015), can successfully account for the empirically observed patterns of utterance prominence.},
pubstate = {published},
type = {article}
}

Copy BibTeX to Clipboard

We present two comparable diachronic corpora of scientific English and German from the Late Modern Period (17th c.–19th c.) annotated with Universal Dependencies. We describe several steps of data pre-processing and evaluate the resulting parsing accuracy showing how our pre-processing steps significantly improve output quality. As a sanity check for the representativity of our data, we conduct a case study comparing previously gained insights on grammatical change in the scientific genre with our data. Our results reflect the often reported trend of English scientific discourse towards heavy noun phrases and a simplification of the sentence structure (Halliday, 1988; Halliday and Martin, 1993; Biber and Gray, 2011; Biber and Gray, 2016). We also show that this trend applies to German scientific discourse as well. The presented corpora are valuable resources suitable for the contrastive analysis of syntactic diachronic change in the scientific genre between 1650 and 1900. The presented pre-processing procedures and their evaluations are applicable to other languages and can be useful for a variety of Natural Language Processing tasks such as syntactic parsing.

• 2022.lrec-1.514     (0.34MB)
• https://aclanthology.org/2022.lrec-1.514/

@inproceedings{krielke-etal-2022-tracing,
title = {Tracing Syntactic Change in the Scientific Genre: Two Universal Dependency-parsed Diachronic Corpora of Scientific English and German},
author = {Marie-Pauline Krielke and Luigi Talamo andM. Fawzi and J. Knappen},
url = {https://aclanthology.org/2022.lrec-1.514/},
year = {2022},
date = {2022},
publisher = {LREC 2022},
address = {Marseille, France},
abstract = {We present two comparable diachronic corpora of scientific English and German from the Late Modern Period (17th c.–19th c.) annotated with Universal Dependencies. We describe several steps of data pre-processing and evaluate the resulting parsing accuracy showing how our pre-processing steps significantly improve output quality. As a sanity check for the representativity of our data, we conduct a case study comparing previously gained insights on grammatical change in the scientific genre with our data. Our results reflect the often reported trend of English scientific discourse towards heavy noun phrases and a simplification of the sentence structure (Halliday, 1988; Halliday and Martin, 1993; Biber and Gray, 2011; Biber and Gray, 2016). We also show that this trend applies to German scientific discourse as well. The presented corpora are valuable resources suitable for the contrastive analysis of syntactic diachronic change in the scientific genre between 1650 and 1900. The presented pre-processing procedures and their evaluations are applicable to other languages and can be useful for a variety of Natural Language Processing tasks such as syntactic parsing.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

The prosodic word (PW) has been proposed as a planning unit in speech production (Levelt et al. [1999. A theory of lexical access in speech production. Behavioral and Brain Sciences, 22, 1–75]), supported by evidence that speech initiation time (RT) is faster for Dutch utterances with fewer PWs due to cliticisation (with the number of lexical words and syllables kept constant) (Wheeldon & Lahiri [1997. Prosodic units in speech production. Journal of Memory and Language, 37(3), 356–381. https://doi.org/10.1006/jmla.1997.2517], W&L). The present study examined prosodic cliticisation (and resulting RT) for a different set of potential clitics (articles, direct-object pronouns), in English, using a different response task (immediate reading aloud). W&L’s result of shorter RTs for fewer PWs was replicated for articles, but not for pronouns, suggesting a difference in cliticisation for these two function word types. However, a post-hoc analysis of the duration of the verb preceding the clitic suggests that both are cliticised. These findings highlight the importance of supplementing production latency measures with phonetic duration measures to understand different stages of language production during utterance planning.

• https://www.tandfonline.com/eprint/4K7DVYQIWRKITU3JCACY/full?target=10.1080/23273798.2022.2060517
• https://doi.org/10.1080/23273798.2022.2060517

@article{Yuen_of_2022,
title = {Planning of prosodic clitics in Australian English},
author = {Ivan Yuen and Katherine Demuth and Stefanie Shattuck-Hufnagel},
url = {https://www.tandfonline.com/eprint/4K7DVYQIWRKITU3JCACY/full?target=10.1080/23273798.2022.2060517},
doi = {https://doi.org/10.1080/23273798.2022.2060517},
year = {2022},
date = {2022-04-05},
journal = {Language, Cognition and Neuroscience},
pages = {1-6},
publisher = {Routledge},
abstract = {The prosodic word (PW) has been proposed as a planning unit in speech production (Levelt et al. [1999. A theory of lexical access in speech production. Behavioral and Brain Sciences, 22, 1–75]), supported by evidence that speech initiation time (RT) is faster for Dutch utterances with fewer PWs due to cliticisation (with the number of lexical words and syllables kept constant) (Wheeldon & Lahiri [1997. Prosodic units in speech production. Journal of Memory and Language, 37(3), 356–381. https://doi.org/10.1006/jmla.1997.2517], W&L). The present study examined prosodic cliticisation (and resulting RT) for a different set of potential clitics (articles, direct-object pronouns), in English, using a different response task (immediate reading aloud). W&L’s result of shorter RTs for fewer PWs was replicated for articles, but not for pronouns, suggesting a difference in cliticisation for these two function word types. However, a post-hoc analysis of the duration of the verb preceding the clitic suggests that both are cliticised. These findings highlight the importance of supplementing production latency measures with phonetic duration measures to understand different stages of language production during utterance planning.},
pubstate = {published},
type = {article}
}

Copy BibTeX to Clipboard

We present results of an eye tracking experiment which aimed at testing sentence comprehension in closely related Slavic languages. Since none of the participants were trained in translation studies or Slavic linguistics, the study illustrates effects of intercomprehension. The participants were exposed to auditory stimuli in Bulgarian, Czech, Polish, and Russian accompanied by a visual scene. The analysis of anticipatory eye movements has shown that native speakers of one Slavic language listening to sentences in another Slavic language, turn their attention to and begin fixating on the referent objects as soon as they identify a predicate. This experiment provides evidence for surprisal-based effects in intercomprehension.

• 2022_212_219     (0.08MB)
• https://www.essv.de/pdf/2022_212_219.pdf?id=1161

@inproceedings{Kudera/etal:2022a,
title = {Comprehension of closely related languages: A visual world eye tracking study},
author = {Jacek Kudera and Philip Georgis and Hasan Md Tusfiqur Alam and Bernd M{\"o}bius and Tania Avgustinova and Dietrich Klakow},
url = {https://www.essv.de/pdf/2022_212_219.pdf?id=1161},
year = {2022},
date = {2022},
booktitle = {Elektronische Sprachsignalverarbeitung 2022, Tagungsband der 33. Konferenz (Sønderborg)},
pages = {212-219},
abstract = {We present results of an eye tracking experiment which aimed at testing sentence comprehension in closely related Slavic languages. Since none of the participants were trained in translation studies or Slavic linguistics, the study illustrates effects of intercomprehension. The participants were exposed to auditory stimuli in Bulgarian, Czech, Polish, and Russian accompanied by a visual scene. The analysis of anticipatory eye movements has shown that native speakers of one Slavic language listening to sentences in another Slavic language, turn their attention to and begin fixating on the referent objects as soon as they identify a predicate. This experiment provides evidence for surprisal-based effects in intercomprehension.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Models of acoustic word embeddings (AWEs) learn to map variable-length spoken word segments onto fixed-dimensionality vector representations such that different acoustic exemplars of the same word are projected nearby in the embedding space. In addition to their speech technology applications, AWE models have been shown to predict human performance on a variety of auditory lexical processing tasks. Current AWE models are based on neural networks and trained in a bottom-up approach that integrates acoustic cues to build up a word representation given an acoustic or symbolic supervision signal. Therefore, these models do not leverage or capture high-level lexical knowledge during the learning process. In this paper, we propose a multi-task learning model that incorporates top-down lexical knowledge into the training procedure of AWEs. Our model learns a mapping between the acoustic input and a lexical representation that encodes high-level information such as word semantics in addition to bottom-up form-based supervision. We experiment with three languages and demonstrate that incorporating lexical knowledge improves the embedding space discriminability and encourages the model to better separate lexical categories.

• abdullah22_interspeech     (5.75MB)
• https://www.isca-speech.org/archive/interspeech_2022/abdullah22_interspeech.html
• https://doi.org/10.21437/Interspeech.2022-626

@inproceedings{Abdullah/etal:2022a,
title = {Integrating form and meaning: A multi-task learning model for acoustic word embeddings},
author = {Badr M. Abdullah and Bernd M{\"o}bius and Dietrich Klakow},
url = {https://www.isca-speech.org/archive/interspeech_2022/abdullah22_interspeech.html},
doi = {https://doi.org/10.21437/Interspeech.2022-626},
year = {2022},
date = {2022},
booktitle = {Proceedings of Interspeech 2022},
pages = {1876-1880},
abstract = {Models of acoustic word embeddings (AWEs) learn to map variable-length spoken word segments onto fixed-dimensionality vector representations such that different acoustic exemplars of the same word are projected nearby in the embedding space. In addition to their speech technology applications, AWE models have been shown to predict human performance on a variety of auditory lexical processing tasks. Current AWE models are based on neural networks and trained in a bottom-up approach that integrates acoustic cues to build up a word representation given an acoustic or symbolic supervision signal. Therefore, these models do not leverage or capture high-level lexical knowledge during the learning process. In this paper, we propose a multi-task learning model that incorporates top-down lexical knowledge into the training procedure of AWEs. Our model learns a mapping between the acoustic input and a lexical representation that encodes high-level information such as word semantics in addition to bottom-up form-based supervision. We experiment with three languages and demonstrate that incorporating lexical knowledge improves the embedding space discriminability and encourages the model to better separate lexical categories.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

This study compares how American (US) and German (DE) listeners perceive emotional expressiveness from Amazon Alexa text-to-speech (TTS) and human voices. Participants heard identical stimuli, manipulated from an emotionally ‘neutral‘ production to three levels of increased happiness generated by resynthesis. Results show that, for both groups, ‘happiness‘ manipulations lead to higher ratings of emotional valence (i.e., more positive) for the human voice. Moreover, there was a difference across the groups in their perception of arousal (i.e., excitement): US listeners show higher ratings for human voices with manipulations, while DE listeners perceive the Alexa voice as sounding less ‘excited‘ overall. We discuss these findings in terms of theories of cross-cultural emotion perception and human-computer interaction.

• gessinger22_interspeech     (0.23MB)
• https://www.isca-speech.org/archive/interspeech_2022/gessinger22_interspeech.html
• https://doi.org/10.21437/Interspeech.2022-146

@inproceedings{Gessinger/etal:2022a,
title = {Cross-cultural comparison of gradient emotion perception: Human vs. Alexa TTS voices},
author = {Iona Gessinger and Michelle Cohn and Georgia Zellou and Bernd M{\"o}bius},
url = {https://www.isca-speech.org/archive/interspeech_2022/gessinger22_interspeech.html},
doi = {https://doi.org/10.21437/Interspeech.2022-146},
year = {2022},
date = {2022},
booktitle = {Proceedings of Interspeech 2022},
pages = {4970-4974},
abstract = {This study compares how American (US) and German (DE) listeners perceive emotional expressiveness from Amazon Alexa text-to-speech (TTS) and human voices. Participants heard identical stimuli, manipulated from an emotionally ‘neutral' production to three levels of increased happiness generated by resynthesis. Results show that, for both groups, ‘happiness' manipulations lead to higher ratings of emotional valence (i.e., more positive) for the human voice. Moreover, there was a difference across the groups in their perception of arousal (i.e., excitement): US listeners show higher ratings for human voices with manipulations, while DE listeners perceive the Alexa voice as sounding less ‘excited' overall. We discuss these findings in terms of theories of cross-cultural emotion perception and human-computer interaction.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

This introductory article for the Special Issue on Vocal Accommodation in Speech Communication provides an overview of prevailing theories of vocal accommodation and summarizes the ten papers in the collection. Communication Accommodation Theory focusses on social factors evoking accent convergence or divergence, while the Interactive Alignment Model proposes cognitive integration of perception and production as an automatic priming mechanism driving convergence language production. Recent research including most of the papers in this Special Issue indicates that a hybrid or interactive synergy model provides a more comprehensive account of observed patterns of phonetic convergence than purely automatic mechanisms. Some of the fundamental questions that this special collection aimed to cover concerned (1) the nature of vocal accommodation in terms of underlying mechanisms and social functions in human–human and human–computer interaction; (2) the effect of task-specific and talker-specific characteristics (gender, age, personality, linguistic and cultural background, role in interaction) on degree and direction of convergence towards human and computer interlocutors; (3) integration of articulatory, perceptual, neurocognitive, and/or multimodal data to the analysis of acoustic accommodation in interactive and non-interactive speech tasks; and (4) the contribution of short/long-term accommodation in human–human and human–computer interactions to the diffusion of linguistic innovation and ultimately language variation and change.

• pardo_etal_jphon-si2022     (0.36MB)
• https://www.coli.uni-saarland.de/~moebius/documents/pardo_etal_jphon-si2022.pdf

@article{Pardo_etal22,
title = {Special issue: Vocal accommodation in speech communication},
author = {Jennifer Pardo and Elisa Pellegrino and Volker Dellwo and Bernd M{\"o}bius},
url = {https://www.coli.uni-saarland.de/~moebius/documents/pardo_etal_jphon-si2022.pdf},
year = {2022},
date = {2022},
journal = {Journal of Phonetics},
pages = {paper 101196},
volume = {95, 1-9},
abstract = {This introductory article for the Special Issue on Vocal Accommodation in Speech Communication provides an overview of prevailing theories of vocal accommodation and summarizes the ten papers in the collection. Communication Accommodation Theory focusses on social factors evoking accent convergence or divergence, while the Interactive Alignment Model proposes cognitive integration of perception and production as an automatic priming mechanism driving convergence language production. Recent research including most of the papers in this Special Issue indicates that a hybrid or interactive synergy model provides a more comprehensive account of observed patterns of phonetic convergence than purely automatic mechanisms. Some of the fundamental questions that this special collection aimed to cover concerned (1) the nature of vocal accommodation in terms of underlying mechanisms and social functions in human–human and human–computer interaction; (2) the effect of task-specific and talker-specific characteristics (gender, age, personality, linguistic and cultural background, role in interaction) on degree and direction of convergence towards human and computer interlocutors; (3) integration of articulatory, perceptual, neurocognitive, and/or multimodal data to the analysis of acoustic accommodation in interactive and non-interactive speech tasks; and (4) the contribution of short/long-term accommodation in human–human and human–computer interactions to the diffusion of linguistic innovation and ultimately language variation and change.},
pubstate = {published},
type = {article}
}

Copy BibTeX to Clipboard

Planning models are usually defined in lifted, i.e. first order formalisms, while most solvers need (variable-free) grounded representations. Though techniques for grounding prune unnecessary parts of the model, grounding might – nevertheless – be prohibitively expensive in terms of runtime. To overcome this issue, there has been renewed interest in solving planning problems based on the lifted representation in the last years. While these approaches are based on (heuristic) search, we present an encoding of lifted classical planning in propositional logic and use SAT solvers to solve it. Our evaluation shows that our approach is competitive with the heuristic search-based approaches in satisficing planning and outperforms them in a (length-)optimal setting.

• https://ojs.aaai.org/index.php/ICAPS/article/view/19794

@inproceedings{HoellerB22,
title = {Encoding Lifted Classical Planning in Propositional Logic},
author = {Daniel H{\"o}ller and Gregor Behnke},
url = {https://ojs.aaai.org/index.php/ICAPS/article/view/19794},
year = {2022},
date = {2022},
booktitle = {Proceedings of the 32nd International Conference on Automated Planning and Scheduling (ICAPS)},
pages = {134-144},
publisher = {AAAI Press},
abstract = {Planning models are usually defined in lifted, i.e. first order formalisms, while most solvers need (variable-free) grounded representations. Though techniques for grounding prune unnecessary parts of the model, grounding might – nevertheless – be prohibitively expensive in terms of runtime. To overcome this issue, there has been renewed interest in solving planning problems based on the lifted representation in the last years. While these approaches are based on (heuristic) search, we present an encoding of lifted classical planning in propositional logic and use SAT solvers to solve it. Our evaluation shows that our approach is competitive with the heuristic search-based approaches in satisficing planning and outperforms them in a (length-)optimal setting.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Obtaining linguistic annotation from novice crowdworkers is far from trivial. A case in point is the annotation of discourse relations, which is a complicated task. Recent methods have obtained promising results by extracting relation labels from either discourse connectives (DCs) or question-answer (QA) pairs that participants provide. The current contribution studies the effect of worker selection and training on the agreement on implicit relation labels between workers and gold labels, for both the DC and the QA method. In Study 1, workers were not specifically selected or trained, and the results show that there is much room for improvement. Study 2 shows that a combination of selection and training does lead to improved results, but the method is cost- and time-intensive. Study 3 shows that a selection-only approach is a viable alternative; it results in annotations of comparable quality compared to annotations from trained participants. The results generalized over both the DC and QA method and therefore indicate that a selection-only approach could also be effective for other crowdsourced discourse annotation tasks.

• https://aclanthology.org/2022.lrec-1.231/

@inproceedings{ Scholmanet-al22-3,
title = {Design Choices in Crowdsourcing Discourse Relation Annotations: The Effect of Worker Selection and Training},
author = {Merel Scholman and Valentina Pyatkin and Frances Pik Yu Yung and Ido Dagan and Reut Tsarfaty and Vera Demberg},
url = {https://aclanthology.org/2022.lrec-1.231/},
year = {2022},
date = {2022},
booktitle = {Proceedings of the Thirteenth Language Resources and Evaluation Conference, Marseille, France},
pages = {2148–2156},
publisher = {European Language Resources Association},
abstract = {Obtaining linguistic annotation from novice crowdworkers is far from trivial. A case in point is the annotation of discourse relations, which is a complicated task. Recent methods have obtained promising results by extracting relation labels from either discourse connectives (DCs) or question-answer (QA) pairs that participants provide. The current contribution studies the effect of worker selection and training on the agreement on implicit relation labels between workers and gold labels, for both the DC and the QA method. In Study 1, workers were not specifically selected or trained, and the results show that there is much room for improvement. Study 2 shows that a combination of selection and training does lead to improved results, but the method is cost- and time-intensive. Study 3 shows that a selection-only approach is a viable alternative; it results in annotations of comparable quality compared to annotations from trained participants. The results generalized over both the DC and QA method and therefore indicate that a selection-only approach could also be effective for other crowdsourced discourse annotation tasks.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

We present DiscoGeM, a crowdsourced corpus of 6,505 implicit discourse relations from three genres: political speech, literature, and encyclopedic texts. Each instance was annotated by 10 crowd workers. Various label aggregation methods were explored to evaluate how to obtain a label that best captures the meaning inferred by the crowd annotators. The results show that a significant proportion of discourse relations in DiscoGeM are ambiguous and can express multiple relation senses. Probability distribution labels better capture these interpretations than single labels. Further, the results emphasize that text genre crucially affects the distribution of discourse relations, suggesting that genre should be included as a factor in automatic relation classification. We make available the newly created DiscoGeM corpus, as well as the dataset with all annotator-level labels. Both the corpus and the dataset can facilitate a multitude of applications and research purposes, for example to function as training data to improve the performance of automatic discourse relation parsers, as well as facilitate research into non-connective signals of discourse relations.

• https://aclanthology.org/2022.lrec-1.351/

@article{Scholman_et-al22.2,
title = {DiscoGeM: A Crowdsourced Corpus of Genre-Mixed Implicit Discourse Relations},
author = {Merel Scholman and Tianai Dong and Frances Pik Yu Yung and Vera Demberg},
url = {https://aclanthology.org/2022.lrec-1.351/},
year = {2022},
date = {2022},
journal = {Proceedings of the 13th International Conference on Language Resources and Evaluation (LREC 22), Marseille, France},
pages = {3281-3290},
abstract = {We present DiscoGeM, a crowdsourced corpus of 6,505 implicit discourse relations from three genres: political speech, literature, and encyclopedic texts. Each instance was annotated by 10 crowd workers. Various label aggregation methods were explored to evaluate how to obtain a label that best captures the meaning inferred by the crowd annotators. The results show that a significant proportion of discourse relations in DiscoGeM are ambiguous and can express multiple relation senses. Probability distribution labels better capture these interpretations than single labels. Further, the results emphasize that text genre crucially affects the distribution of discourse relations, suggesting that genre should be included as a factor in automatic relation classification. We make available the newly created DiscoGeM corpus, as well as the dataset with all annotator-level labels. Both the corpus and the dataset can facilitate a multitude of applications and research purposes, for example to function as training data to improve the performance of automatic discourse relation parsers, as well as facilitate research into non-connective signals of discourse relations.},
pubstate = {published},
type = {article}
}

Copy BibTeX to Clipboard

A central issue in linguistics concerns the relationship between theories and evidence in data. We investigate this issue in the field of discourse coherence, and particularly the study of coherence relations such as causal and contrastive. Proposed inventories of coherence relations differ greatly in the type and number of proposed relations. Such proposals are often validated by focusing on either the descriptive adequacy (researcher’s intuitions on textual interpretations) or the cognitive plausibility of distinctions (empirical research on cognition). We argue that both are important, and note that the concept of cognitive plausibility is in need of a concrete definition and quantifiable operationalization. This contribution focuses on how the criterion of cognitive plausibility can be operationalized and presents a systematic validation approach to evaluate discourse frameworks. This is done by detailing how various sources of evidence can be used to support or falsify distinctions between coherence relational labels. Finally, we present methodological issues regarding verification and falsification that are of importance to all discourse researchers studying the relationship between theory and data.

• https://journals.openedition.org/discours/12075

@article{Scholman_etal22,
title = {Descriptively adequate and cognitively plausible? Validating distinctions between types of coherence relations},
author = {Merel Scholman and Vera Demberg and Ted J. M. Sanders},
url = {https://journals.openedition.org/discours/12075},
year = {2022},
date = {2022},
journal = {Discours},
pages = {1-30a},
volume = {30},
abstract = {A central issue in linguistics concerns the relationship between theories and evidence in data. We investigate this issue in the field of discourse coherence, and particularly the study of coherence relations such as causal and contrastive. Proposed inventories of coherence relations differ greatly in the type and number of proposed relations. Such proposals are often validated by focusing on either the descriptive adequacy (researcher’s intuitions on textual interpretations) or the cognitive plausibility of distinctions (empirical research on cognition). We argue that both are important, and note that the concept of cognitive plausibility is in need of a concrete definition and quantifiable operationalization. This contribution focuses on how the criterion of cognitive plausibility can be operationalized and presents a systematic validation approach to evaluate discourse frameworks. This is done by detailing how various sources of evidence can be used to support or falsify distinctions between coherence relational labels. Finally, we present methodological issues regarding verification and falsification that are of importance to all discourse researchers studying the relationship between theory and data.},
pubstate = {published},
type = {article}
}

Copy BibTeX to Clipboard

In many linguistic fields requiring annotated data, multiple interpretations of a single item are possible. Multi-label annotations more accurately reflect this possibility. However, allowing for multi-label annotations also affects the chance that two coders agree with each other. Calculating inter-coder agreement for multi-label datasets is therefore not trivial. In the current contribution, we evaluate different metrics for calculating agreement on multi-label annotations: agreement on the intersection of annotated labels, an augmented version of Cohen’s Kappa, and precision, recall and F1. We propose a bootstrapping method to obtain chance agreement for each measure, which allows us to obtain an adjusted agreement coefficient that is more interpretable. We demonstrate how various measures affect estimates of agreement on simulated datasets and present a case study of discourse relation annotations. We also show how the proportion of double labels, and the entropy of the label distribution, influences the measures outlined above and how a bootstrapped adjusted agreement can make agreement measures more comparable across datasets in multi-label scenarios.

• https://aclanthology.org/2022.coling-1.322/

@inproceedings{Marchaletal22-2,
title = {Establishing annotation quality in multi-label annotations},
author = {Marian Marchal and Merel Scholman and Frances Pik Yu Yung and Vera Demberg},
url = {https://aclanthology.org/2022.coling-1.322/},
year = {2022},
date = {2022},
booktitle = {Proceedings of the 29th International Conference on Computational Linguistic (COLING)},
pages = {3659–3668},
abstract = {In many linguistic fields requiring annotated data, multiple interpretations of a single item are possible. Multi-label annotations more accurately reflect this possibility. However, allowing for multi-label annotations also affects the chance that two coders agree with each other. Calculating inter-coder agreement for multi-label datasets is therefore not trivial. In the current contribution, we evaluate different metrics for calculating agreement on multi-label annotations: agreement on the intersection of annotated labels, an augmented version of Cohen’s Kappa, and precision, recall and F1. We propose a bootstrapping method to obtain chance agreement for each measure, which allows us to obtain an adjusted agreement coefficient that is more interpretable. We demonstrate how various measures affect estimates of agreement on simulated datasets and present a case study of discourse relation annotations. We also show how the proportion of double labels, and the entropy of the label distribution, influences the measures outlined above and how a bootstrapped adjusted agreement can make agreement measures more comparable across datasets in multi-label scenarios.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

It is generally assumed that readers draw on their background knowledge to make inferences about information that is left implicit in the text. However, readers may differ in how much background knowledge they have, which may impact their text understanding. The present study investigates the role of domain knowledge in discourse relation interpretation, in order to examine how readers with high vs. low domain knowledge differ in their discourse relation inferences. We compare interpretations of experts from the field of economics and biomedical sciences in scientific biomedical texts as well as more easily accessible economic texts. The results show that high-knowledge readers from the biomedical domain are better at inferring the correct relation interpretation in biomedical texts compared to low-knowledge readers, but such an effect was not found for the economic domain. The results also suggest that, in the absence of domain knowledge, readers exploit linguistic signals other than connectives to infer the discourse relation, but domain knowledge is sometimes required to exploit these cues. The study provides insight into the impact of domain knowledge on discourse relation inferencing and how readers interpret discourse relations when they lack the required domain knowledge.

• DD_MarchalScholmanDemberg_2109     (0.25MB)
• https://journals.uic.edu/ojs/index.php/dad/article/view/12343/10711

@article{Marchaletal22,
title = {The effect of domain knowledge on discourse relation inferences: Relation marking and interpretation strategies},
author = {Marian Marchal and Merel Scholman and Vera Demberg},
url = {https://journals.uic.edu/ojs/index.php/dad/article/view/12343/10711},
year = {2022},
date = {2022},
journal = {Dialogue & Discourse},
pages = {49-78},
volume = {13},
number = {(2)},
abstract = {It is generally assumed that readers draw on their background knowledge to make inferences about information that is left implicit in the text. However, readers may differ in how much background knowledge they have, which may impact their text understanding. The present study investigates the role of domain knowledge in discourse relation interpretation, in order to examine how readers with high vs. low domain knowledge differ in their discourse relation inferences. We compare interpretations of experts from the field of economics and biomedical sciences in scientific biomedical texts as well as more easily accessible economic texts. The results show that high-knowledge readers from the biomedical domain are better at inferring the correct relation interpretation in biomedical texts compared to low-knowledge readers, but such an effect was not found for the economic domain. The results also suggest that, in the absence of domain knowledge, readers exploit linguistic signals other than connectives to infer the discourse relation, but domain knowledge is sometimes required to exploit these cues. The study provides insight into the impact of domain knowledge on discourse relation inferencing and how readers interpret discourse relations when they lack the required domain knowledge.},
pubstate = {published},
type = {article}
}

Copy BibTeX to Clipboard

The present study investigates L2 prosodic realizations in the readings of two groups of Bulgarian informants: (a) with L2 German, and (b) with L2 English. Each group consisted of ten female learners, who read the fable “The North Wind and the Sun” in their L1 and in the respective L2. We also recorded two groups of female native speakers of the target languages as controls. The following durational parameters were obtained: mean accented syllable duration, accented/naccented duration ratio, speaking rate. With respect to F0 parameters, mean, median, minimum, maximum, span in semitones, and standard deviations per IP were measured. Additionally, we calculated the number of accented and unaccented syllables, IPs and pauses in each reading. Statistical analyses show that the two groups differ in their use of F0. Both groups use higher standard deviation and level in their L2, whereas the ‘German group’ use higher pitch span as well. The number of accented syllables, IPs and pauses is also higher in L2. Regarding duration, both groups use slower articulation rate. The accented/unaccented syllable duration ratio is lower in L2 for the ‘English group’. We also provide original data on speaking rate in Bulgarian from an information theoretical perspective.

• https://www.isca-speech.org/archive/speechprosody_2022/andreeva22_speechprosody.html
• https://doi.org/10.21437/SpeechProsody.2022-155

@inproceedings{andreeva_2022_speechprosody,
title = {The influence of L1 prosody on Bulgarian-accented German and English},
author = {Bistra Andreeva and Snezhina Dimitrova},
url = {https://www.isca-speech.org/archive/speechprosody_2022/andreeva22_speechprosody.html},
doi = {https://doi.org/10.21437/SpeechProsody.2022-155},
year = {2022},
date = {2022},
booktitle = {Proc. Speech Prosody 2022},
pages = {764-768},
address = {Lisbon},
abstract = {The present study investigates L2 prosodic realizations in the readings of two groups of Bulgarian informants: (a) with L2 German, and (b) with L2 English. Each group consisted of ten female learners, who read the fable “The North Wind and the Sun” in their L1 and in the respective L2. We also recorded two groups of female native speakers of the target languages as controls. The following durational parameters were obtained: mean accented syllable duration, accented/naccented duration ratio, speaking rate. With respect to F0 parameters, mean, median, minimum, maximum, span in semitones, and standard deviations per IP were measured. Additionally, we calculated the number of accented and unaccented syllables, IPs and pauses in each reading. Statistical analyses show that the two groups differ in their use of F0. Both groups use higher standard deviation and level in their L2, whereas the ‘German group’ use higher pitch span as well. The number of accented syllables, IPs and pauses is also higher in L2. Regarding duration, both groups use slower articulation rate. The accented/unaccented syllable duration ratio is lower in L2 for the ‘English group’. We also provide original data on speaking rate in Bulgarian from an information theoretical perspective.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Cross-linguistic evidence suggests that syllables in predictable contexts have shorter duration than in unpredictable contexts. However, it is not clear if predictability uniformly affects phonetic cues of a phonological feature in a segment. The current study explored the effect of syllable-based predictability on the durational correlates of the phonological stop voicing contrast in German, viz. voice onset time (VOT) and closure duration (CD), using data in Ibrahim et al. [1]. The target stop consonants /b, p, d, k/ occurred in stressed CV syllables in polysyllabic words embedded in a sentence, with either voiced or voiceless preceding contexts. The syllable occurred in either a low or a high predictable condition, which was based on a syllable-level trigram language model. We measured VOT and CD of the target consonants (voiced vs. voiceless). Our results showed an interaction effect of predictability and the voicing status of the target consonants on VOT, but a uniform effect on closure duration. This interaction effect on a primary cue like VOT indicates a selective effect of predictability on VOT, but not on CD. This suggests that the effect of predictability is sensitive to the phonological relevance of a language-specific phonetic cue.

• ibrahim22_speechprosody     (3.62MB)
• https://www.isca-speech.org/archive/pdfs/speechprosody_2022/ibrahim22_speechprosody.pdf
• https://doi.org/10.21437/SpeechProsody.2022-136

@inproceedings{ibrahim_2022_speechprosody,
title = {The effect of predictability on German stop voicing is phonologically selective},
author = {Omnia Ibrahim and Ivan Yuen and Bistra Andreeva and Bernd M{\"o}bius},
url = {https://www.isca-speech.org/archive/pdfs/speechprosody_2022/ibrahim22_speechprosody.pdf},
doi = {https://doi.org/10.21437/SpeechProsody.2022-136},
year = {2022},
date = {2022},
booktitle = {Proc. Speech Prosody 2022},
pages = {669-673},
address = {Lisbon},
abstract = {Cross-linguistic evidence suggests that syllables in predictable contexts have shorter duration than in unpredictable contexts. However, it is not clear if predictability uniformly affects phonetic cues of a phonological feature in a segment. The current study explored the effect of syllable-based predictability on the durational correlates of the phonological stop voicing contrast in German, viz. voice onset time (VOT) and closure duration (CD), using data in Ibrahim et al. [1]. The target stop consonants /b, p, d, k/ occurred in stressed CV syllables in polysyllabic words embedded in a sentence, with either voiced or voiceless preceding contexts. The syllable occurred in either a low or a high predictable condition, which was based on a syllable-level trigram language model. We measured VOT and CD of the target consonants (voiced vs. voiceless). Our results showed an interaction effect of predictability and the voicing status of the target consonants on VOT, but a uniform effect on closure duration. This interaction effect on a primary cue like VOT indicates a selective effect of predictability on VOT, but not on CD. This suggests that the effect of predictability is sensitive to the phonological relevance of a language-specific phonetic cue.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

• 8Talamo     (0.53MB)
• https://www.italian-journal-linguistics.com/app/uploads/2023/01/8-Talamo.pdf
• https://doi.org/10.26346/1120-2726-197

@article{article,
title = {A new methodology for an old problem: A corpus-based typology of adnominal word order in European languages},
author = {Luigi Talamo and Annemarie Verkerk},
url = {https://www.italian-journal-linguistics.com/app/uploads/2023/01/8-Talamo.pdf},
doi = {https://doi.org/10.26346/1120-2726-197},
year = {2022},
date = {2022},
journal = {Italian Journal of Linguistics},
pages = {171-226},
volume = {34},
abstract = {

Copy BibTeX to Clipboard

• https://aclanthology.org/2022.emnlp-main.133

@inproceedings{espana-bonet-barron-cedeno-2022-undesired,
title = {The (Undesired) Attenuation of Human Biases by Multilinguality},
author = {Cristina Espa{\~n}a-Bonet and Alberto Barrón-Cede{\~n}o},
url = {https://aclanthology.org/2022.emnlp-main.133},
year = {2022},
date = {2022},
booktitle = {Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing},
pages = {2056–2077},
publisher = {Association for Computational Linguistics},
address = {Online and Abu Dhabi, UAE, Dec 2022},
abstract = {

Copy BibTeX to Clipboard