Publications

Local coherence is essential for long-form text generation models. We identify two important aspects of local coherence within the visual storytelling task: (1) the model needs to represent re-occurrences of characters within the image sequence in order to mention them correctly in the story; (2) character representations should enable us to find instances of the same characters and distinguish different characters. In this paper, we propose a loss function inspired by a linguistic theory of coherence for self-supervised learning for image sequence representations. We further propose combining features from an object and a face detector to construct stronger character features. To evaluate input-output relevance that current reference-based metrics don{‚}t measure, we propose a character matching metric to check whether the models generate referring expressions correctly for characters in input image sequences. Experiments on a visual story generation dataset show that our proposed features and loss function are effective for generating more coherent and visually grounded stories.

• https://aclanthology.org/2023.findings-acl.603
• https://doi.org/10.18653/v1/2023.findings-acl.603

@inproceedings{hong-etal-2023-visual,
title = {Visual Coherence Loss for Coherent and Visually Grounded Story Generation},
author = {Xudong Hong and Vera Demberg and Asad Sayeed and Qiankun Zheng and Bernt Schiele},
editor = {Anna Rogers and Jordan Boyd-Graber and Naoaki Okazaki},
url = {https://aclanthology.org/2023.findings-acl.603},
doi = {https://doi.org/10.18653/v1/2023.findings-acl.603},
year = {2023},
date = {2023},
booktitle = {Findings of the Association for Computational Linguistics: ACL 2023},
pages = {9456-9470},
publisher = {Association for Computational Linguistics},
address = {Toronto, Canada},
abstract = {Local coherence is essential for long-form text generation models. We identify two important aspects of local coherence within the visual storytelling task: (1) the model needs to represent re-occurrences of characters within the image sequence in order to mention them correctly in the story; (2) character representations should enable us to find instances of the same characters and distinguish different characters. In this paper, we propose a loss function inspired by a linguistic theory of coherence for self-supervised learning for image sequence representations. We further propose combining features from an object and a face detector to construct stronger character features. To evaluate input-output relevance that current reference-based metrics don{'}t measure, we propose a character matching metric to check whether the models generate referring expressions correctly for characters in input image sequences. Experiments on a visual story generation dataset show that our proposed features and loss function are effective for generating more coherent and visually grounded stories.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

This thesis focuses on acquiring wide coverage script knowledge. Script knowledge constitutes a category of common sense knowledge that delineates the procedural aspects of daily activities, such as taking a train and going grocery shopping. It is believed to reside in human memory and is generally assumed by all conversational parties. Conversational utterances often omit details assumed to be known by listeners, who, in turn, comprehend these concise expressions based on their shared understanding, with common sense knowledge forming the basis. Common sense knowledge is indispensable for both the production and comprehension of conversation. As outlined in Chapters 2 and 3, Natural Language Processing (NLP) applications experience significant enhancements with access to script knowledge. Notably, various NLP tasks demonstrate substantial performance improvements when script knowledge is accessible, suggesting that these applications are not fully cognizant of script knowledge. However, acquiring high-quality script knowledge is costly, resulting in limited resources that cover only a few scenarios. Consequently, the practical utility of existing resources is constrained due to insufficient coverage of script knowledge. This thesis is dedicated to developing cost-effective methods for acquiring script knowledge to augment NLP applications and expand the coverage of explicit script knowledge. Previous resources have been generated through intricate manual annotation pipelines. In this work, we introduce automated methods to streamline the annotation process. Specifically, we propose a zero-shot script parser in Chapter 5. By leveraging representation learning, we extract script annotations from existing resources and employ this knowledge to automatically annotate texts from unknown scenarios. When applied to parallel descriptions of unknown scenarios, the acquired script knowledge proves adequate to support NLP applications, such as story generation (Chapter 6). In Chapter 7, we explore the potential of pretrained language models as a source of script knowledge.

• https://publikationen.sulb.uni-saarland.de/handle/20.500.11880/37341
• https://doi.org/10.22028/D291-41495

@phdthesis{Zhai_Diss_2023,
title = {Towards wider coverage script knowledge for NLP},
author = {Fangzhou Zhai},
url = {https://publikationen.sulb.uni-saarland.de/handle/20.500.11880/37341},
doi = {https://doi.org/10.22028/D291-41495},
year = {2023},
date = {2023},
school = {Saarland University},
publisher = {Saarl{\"a}ndische Universit{\"a}ts- und Landesbibliothek},
address = {Saarbruecken, Germany},
abstract = {This thesis focuses on acquiring wide coverage script knowledge. Script knowledge constitutes a category of common sense knowledge that delineates the procedural aspects of daily activities, such as taking a train and going grocery shopping. It is believed to reside in human memory and is generally assumed by all conversational parties. Conversational utterances often omit details assumed to be known by listeners, who, in turn, comprehend these concise expressions based on their shared understanding, with common sense knowledge forming the basis. Common sense knowledge is indispensable for both the production and comprehension of conversation. As outlined in Chapters 2 and 3, Natural Language Processing (NLP) applications experience significant enhancements with access to script knowledge. Notably, various NLP tasks demonstrate substantial performance improvements when script knowledge is accessible, suggesting that these applications are not fully cognizant of script knowledge. However, acquiring high-quality script knowledge is costly, resulting in limited resources that cover only a few scenarios. Consequently, the practical utility of existing resources is constrained due to insufficient coverage of script knowledge. This thesis is dedicated to developing cost-effective methods for acquiring script knowledge to augment NLP applications and expand the coverage of explicit script knowledge. Previous resources have been generated through intricate manual annotation pipelines. In this work, we introduce automated methods to streamline the annotation process. Specifically, we propose a zero-shot script parser in Chapter 5. By leveraging representation learning, we extract script annotations from existing resources and employ this knowledge to automatically annotate texts from unknown scenarios. When applied to parallel descriptions of unknown scenarios, the acquired script knowledge proves adequate to support NLP applications, such as story generation (Chapter 6). In Chapter 7, we explore the potential of pretrained language models as a source of script knowledge.},
pubstate = {published},
type = {phdthesis}
}

Copy BibTeX to Clipboard

• https://www.sciencedirect.com/science/article/pii/S037821662300098X
• https://doi.org/10.1016/j.pragma.2023.04.005

@article{ryzhova-demberg-2023,
title = {Processing cost effects of atypicality inferences in a dual-task setup},
author = {Margarita Ryzhova and Vera Demberg},
url = {https://www.sciencedirect.com/science/article/pii/S037821662300098X},
doi = {https://doi.org/10.1016/j.pragma.2023.04.005},
year = {2023},
date = {2023},
journal = {Journal of Pragmatics},
pages = {47-80},
volume = {211},
abstract = {

},
pubstate = {published},
type = {article}
}

Copy BibTeX to Clipboard

There is a long-standing controversy around the question of whether referent predictability affects pronominalization: while there are good theoretical reasons for this prediction (e.g., Arnold, 2008), the experimental evidence has been rather mixed. We here report on three highly powered studies that manipulate a range of factors that have differed between previous studies, in order to determine more exactly under which conditions a predictability effect on pronominalization can be found. We use a constrained as well as a free reference task, and manipulate verb type, antecedent ambiguity, length of NP and whether the stimuli are presented within a story context or not. Our results find the story context to be the single important factor that allows to elicit an effect of predictability on pronoun choice, in line with (Rosa and Arnold, 2017; Weatherford and Arnold, 2021). We also propose a parametrization for a rational speech act model, that reconciles the findings between many of the experiments in the literature.

• https://europepmc.org/article/MED/37265576

@article{Demberg.etal23,
title = {A systematic evaluation of factors affecting referring expression choice in passage completion tasks},
author = {Vera Demberg and Ekaterina Kravtchenko and Jia Loy},
url = {https://europepmc.org/article/MED/37265576},
year = {2023},
date = {2023},
journal = {Journal of Memory and Language, 130, 104413},
abstract = {There is a long-standing controversy around the question of whether referent predictability affects pronominalization: while there are good theoretical reasons for this prediction (e.g., Arnold, 2008), the experimental evidence has been rather mixed. We here report on three highly powered studies that manipulate a range of factors that have differed between previous studies, in order to determine more exactly under which conditions a predictability effect on pronominalization can be found. We use a constrained as well as a free reference task, and manipulate verb type, antecedent ambiguity, length of NP and whether the stimuli are presented within a story context or not. Our results find the story context to be the single important factor that allows to elicit an effect of predictability on pronoun choice, in line with (Rosa and Arnold, 2017; Weatherford and Arnold, 2021). We also propose a parametrization for a rational speech act model, that reconciles the findings between many of the experiments in the literature.},
pubstate = {published},
type = {article}
}

Copy BibTeX to Clipboard

Most theories of pragmatics and language processing predict that speakers avoid excessive informational redundancy. Informationally redundant utterances are, however, quite common in natural dialogue. From a comprehension standpoint, it remains unclear how comprehenders interpret these utterances, and whether they make attempts to reconcile the ‚dips‘ in informational utility with expectations of ‚appropriate‘ or ‚rational‘ speaker informativity. We show that informationally redundant (overinformative) utterances can trigger pragmatic inferences that increase utterance utility in line with comprehender expectations. In a series of three studies, we look at utterances which refer to stereotyped event sequences describing common activities (scripts). When comprehenders encounter utterances describing events that can be easily inferred from prior context, they interpret them as signifying that the event conveys new, unstated information (i.e. an event otherwise assumed to be habitual, such as paying the cashier when shopping, is reinterpreted as non-habitual). We call these inferences atypicality inferences. Further, we show that the degree to which these atypicality inferences are triggered depends on the framing of the utterance. In the absence of an exclamation mark or a discourse marker indicating the speaker’s specific intent to communicate the given information, such inferences are far less likely to arise. Overall, the results demonstrate that excessive conceptual redundancy leads to comprehenders revising the conversational common ground, in an effort to accommodate unexpected dips in informational utility.

• 1-s2.0-S0010027722001470-main     (2.62MB)
• https://www.sciencedirect.com/science/article/pii/S0010027722001470
• https://doi.org/ 10.1016/j.cognition.2022.105159

@article{Kravtchenko_redundant_2022,
title = {Informationally redundant utterances elicit pragmatic inferences},
author = {Ekaterina Kravtchenko and Vera Demberg},
url = {https://www.sciencedirect.com/science/article/pii/S0010027722001470},
doi = {https://doi.org/ 10.1016/j.cognition.2022.105159},
year = {2022},
date = {2022},
journal = {Cognition},
pages = {105159},
volume = {225},
abstract = {Most theories of pragmatics and language processing predict that speakers avoid excessive informational redundancy. Informationally redundant utterances are, however, quite common in natural dialogue. From a comprehension standpoint, it remains unclear how comprehenders interpret these utterances, and whether they make attempts to reconcile the 'dips' in informational utility with expectations of 'appropriate' or 'rational' speaker informativity. We show that informationally redundant (overinformative) utterances can trigger pragmatic inferences that increase utterance utility in line with comprehender expectations. In a series of three studies, we look at utterances which refer to stereotyped event sequences describing common activities (scripts). When comprehenders encounter utterances describing events that can be easily inferred from prior context, they interpret them as signifying that the event conveys new, unstated information (i.e. an event otherwise assumed to be habitual, such as paying the cashier when shopping, is reinterpreted as non-habitual). We call these inferences atypicality inferences. Further, we show that the degree to which these atypicality inferences are triggered depends on the framing of the utterance. In the absence of an exclamation mark or a discourse marker indicating the speaker's specific intent to communicate the given information, such inferences are far less likely to arise. Overall, the results demonstrate that excessive conceptual redundancy leads to comprehenders revising the conversational common ground, in an effort to accommodate unexpected dips in informational utility.},
keywords = {Accommodation; Context-dependent implicatures; Experimental pragmatics; Psycholinguistics; Redundancy},
pubstate = {published},
type = {article}
}

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

Script knowledge (Schank and Abelson, 1977) is useful for a variety of NLP tasks. However, existing resources only cover a small number of activities, limiting its practical usefulness. In this work, we propose a zero-shot learning approach to script parsing, the task of tagging texts with scenario-specific event and participant types, which enables us to acquire script knowledge without domain-specific annotations. We (1) learn representations of potential event and participant mentions by promoting class consistency according to the annotated data; (2) perform clustering on the event /participant candidates from unannotated texts that belongs to an unseen scenario. The model achieves 68.1/74.4 average F1 for event / participant parsing, respectively, outperforming a previous CRF model that, in contrast, has access to scenario-specific supervision. We also evaluate the model by testing on a different corpus, where it achieved 55.5/54.0 average F1 for event / participant parsing.

• 2022.coling-1.356     (0.63MB)
• https://aclanthology.org/2022.coling-1.356

@inproceedings{zhai-etal-2022-zero,
title = {Zero-shot Script Parsing},
author = {Fangzhou Zhai and Vera Demberg and Alexander Koller},
url = {https://aclanthology.org/2022.coling-1.356},
year = {2022},
date = {2022},
booktitle = {Proceedings of the 29th International Conference on Computational Linguistics},
pages = {4049-4060},
publisher = {International Committee on Computational Linguistics},
address = {Gyeongju, Republic of Korea},
abstract = {Script knowledge (Schank and Abelson, 1977) is useful for a variety of NLP tasks. However, existing resources only cover a small number of activities, limiting its practical usefulness. In this work, we propose a zero-shot learning approach to script parsing, the task of tagging texts with scenario-specific event and participant types, which enables us to acquire script knowledge without domain-specific annotations. We (1) learn representations of potential event and participant mentions by promoting class consistency according to the annotated data; (2) perform clustering on the event /participant candidates from unannotated texts that belongs to an unseen scenario. The model achieves 68.1/74.4 average F1 for event / participant parsing, respectively, outperforming a previous CRF model that, in contrast, has access to scenario-specific supervision. We also evaluate the model by testing on a different corpus, where it achieved 55.5/54.0 average F1 for event / participant parsing.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Recipe texts are an idiosyncratic form of instructional language that pose unique challenges for automatic understanding. One challenge is that a cooking step in one recipe can be explained in another recipe in different words, at a different level of abstraction, or not at all. Previous work has annotated correspondences between recipe instructions at the sentence level, often glossing over important correspondences between cooking steps across recipes. We present a novel and fully-parsed English recipe corpus, ARA (Aligned Recipe Actions), which annotates correspondences between individual actions across similar recipes with the goal of capturing information implicit for accurate recipe understanding. We represent this information in the form of recipe graphs, and we train a neural model for predicting correspondences on ARA. We find that substantial gains in accuracy can be obtained by taking fine-grained structural information about the recipes into account.

• https://aclanthology.org/2021.emnlp-main.554

@inproceedings{donatelli21:align,
title = {Aligning Actions Across Recipe Graphs},
author = {Lucia Donatelli and Theresa Schmidt and Debanjali Biswas and Arne K{\"o}hn and Fangzhou Zhai and Alexander Koller},
url = {https://aclanthology.org/2021.emnlp-main.554},
year = {2021},
date = {2021},
booktitle = {Proceedings of EMNLP},
pages = {6930–6942},
abstract = {Recipe texts are an idiosyncratic form of instructional language that pose unique challenges for automatic understanding. One challenge is that a cooking step in one recipe can be explained in another recipe in different words, at a different level of abstraction, or not at all. Previous work has annotated correspondences between recipe instructions at the sentence level, often glossing over important correspondences between cooking steps across recipes. We present a novel and fully-parsed English recipe corpus, ARA (Aligned Recipe Actions), which annotates correspondences between individual actions across similar recipes with the goal of capturing information implicit for accurate recipe understanding. We represent this information in the form of recipe graphs, and we train a neural model for predicting correspondences on ARA. We find that substantial gains in accuracy can be obtained by taking fine-grained structural information about the recipes into account.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Scripts capture commonsense knowledge about everyday activities and their participants. Script knowledge proved useful in a number of NLP tasks, such as referent prediction, discourse classification, and story generation. A crucial step for the exploitation of script knowledge is script parsing, the task of tagging a text with the events and participants from a certain activity. This task is challenging: it requires information both about the ways events and participants are usually uttered in surface language as well as the order in which they occur in the world. We show how to do accurate script parsing with a hierarchical sequence model and transfer learning. Our model improves the state of the art of event parsing by over 16 points F-score and, for the first time, accurately tags script participants.

• https://aclanthology.org/2021.starsem-1.18
• https://doi.org/10.18653/v1/2021.starsem-1.18

@inproceedings{zhaiSkrjanecKoller2021,
title = {Script Parsing with Hierarchical Sequence Modelling},
author = {Fangzhou Zhai and Iza Skrjanec and Alexander Koller},
url = {https://aclanthology.org/2021.starsem-1.18},
doi = {https://doi.org/10.18653/v1/2021.starsem-1.18},
year = {2021},
date = {2021},
booktitle = {Proceedings of *SEM 2021: 10th Joint Conf. on Lexical and Computational Semantics},
pages = {195-201},
abstract = {Scripts capture commonsense knowledge about everyday activities and their participants. Script knowledge proved useful in a number of NLP tasks, such as referent prediction, discourse classification, and story generation. A crucial step for the exploitation of script knowledge is script parsing, the task of tagging a text with the events and participants from a certain activity. This task is challenging: it requires information both about the ways events and participants are usually uttered in surface language as well as the order in which they occur in the world. We show how to do accurate script parsing with a hierarchical sequence model and transfer learning. Our model improves the state of the art of event parsing by over 16 points F-score and, for the first time, accurately tags script participants.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

This thesis explores new methods of accounting for discourse-level linguistic phenomena, using computational modeling. When communicating, efficient speakers frequently choose to either omit, or otherwise reduce the length of their utterances wherever possible. Frameworks such as Uniform Information Density (UID) have argued that speakers preferentially reduce or omit those elements that are more predictable in context, and easier to recover. However, these frameworks have nothing to say about the effects of a linguistic choice on how a message is interpreted. I run 3 experiments which show that while UID posits no specific consequences to being „overinformative“ (including more information in an utterance than is necessary), in fact overinformativeness can trigger pragmatic inferences which alter comprehenders‘ background beliefs about the world. In this case, I show that the Rational Speech Act (RSA) model, which models back-and-forth pragmatic reasoning between speakers and comprehenders, predicts both efficiency-based utterance choices, as well as any consequent change in perceived meaning. I also provide evidence that it’s critical to model communication as a lossy process (which UID assumes), which allows the RSA model to account for phenomena that it otherwise is not able to. I further show that while UID predicts increased use of pronouns when referring to more contextually predictable referents, existing research does not unequivocally support this. I run 2 experiments which fail to show evidence that speakers use reduced expressions for predictable elements. In contrast to UID and similar frameworks, the RSA model can straightforwardly predict the results that have been observed to date. In the end, I argue that the RSA model is a highly attractive alternative for modeling speaker utterance choice at the discourse level. When it reflects communication as a lossy process, it is able to predict the same predictability-driven utterance reduction that UID does. However, by additionally modeling back-and-forth pragmatic reasoning, it successfully models utterance choice phenomena that simpler frameworks cannot account for.

---

Diese Arbeit erforscht neue Methoden, linguistische Phänomene auf Gesprächsebene per Computermodellierung zu erfassen. Effiziente Sprecher:innen entscheiden sich bei der Kommunikation häufig dazu, wenn immer es möglich ist, Äußerungen entweder ganz auszulassen oder aber ihre Länge zu reduzieren. Modelle wie Uniform Information Density (UID) argumentieren, dass Sprecher:innen vorzugsweise diejenigen Elemente auslassen, die im jeweiligen Kontext vorhersagbarer und einfacher wiederherzustellen sind. Allerdings sagen diese Modelle nichts über die Auswirkungen einer linguistischen Entscheidung bezüglich der Interpretation einer Nachricht aus. Ich führe drei Experimente durch, die zeigen, dass wenngleich UID keine spezifischen Auswirkungen von „Überinformation“ (einer Äußerung mehr Information als nötig geben) postuliert, Überinformationen doch pragmatische Schlussfolgerungen, die das gedankliche Weltmodell der Versteher:innen ändern können, auslöst. Für diesen Fall zeige ich, dass das Rational-Speech-Act-Modell (RSA), welches pragmatische Hin-und-Her-Schlussfolgerungen zwischen Sprecher:innen und Versteher:innen modelliert, sowohl effizienzbasierte Äußerungsauswahl als auch jegliche resultierende Verständnisänderung vorhersagt. Ich liefere auch Anhaltspunkte dafür, dass es entscheidend ist, Kommunikation als verlustbehafteten Prozess zu modellieren (wovon UID ausgeht), was es dem RSA-Modell erlaubt, Phänomene einzubeziehen, wozu es sonst nicht in der Lage wäre. Weiterhin zeige ich, dass obschon UID beim Bezug auf kontextuell vorhersagbarere Bezugswörter eine erhöhte Nutzung von Pronomen vorhersagt, dies von existierender Forschung nicht einstimmig gestützt wird. Ich führe zwei Experimente durch, die keine Anhaltspunkte dafür, dass Sprecher:innen reduzierte Ausdrücke für vorhersagbare Elemente verwenden, finden. Im Gegensatz zu UID und ähnlichen Modellen kann dass RSA-Modell direkt die bislang beobachteten Resultate vorhersagen. Schließlich lege ich dar, warum das RSA-Modell eine höchst attraktive Alternative zur Modellierung von Sprachäußerungsentscheidungen auf Gesprächsebene ist. Wenn es Kommunikation als einen verlustbehafteten Prozess widerspiegelt, kann es dieselbe vorhersagebasierte Äußerungsreduktion vorhersagen wie auch UID. Modelliert man jedoch zusätzlich pragmatische Hin-und-Her-Schlussfolgerungen, modelliert RSA erfolgreich Phänomene bei Äußerungsentscheidungen, die einfachere Modelle nicht abbilden können.

• https://publikationen.sulb.uni-saarland.de/handle/20.500.11880/33102
• https://doi.org/10.22028/D291-35858

@phdthesis{Kravtchenko_Diss_2021,
title = {Integrating pragmatic reasoning in an efficiency-based theory of utterance choice},
author = {Ekaterina Kravtchenko},
url = {https://publikationen.sulb.uni-saarland.de/handle/20.500.11880/33102},
doi = {https://doi.org/10.22028/D291-35858},
year = {2021},
date = {2021},
school = {Saarland University},
address = {Saarbruecken, Germany},
abstract = {This thesis explores new methods of accounting for discourse-level linguistic phenomena, using computational modeling. When communicating, efficient speakers frequently choose to either omit, or otherwise reduce the length of their utterances wherever possible. Frameworks such as Uniform Information Density (UID) have argued that speakers preferentially reduce or omit those elements that are more predictable in context, and easier to recover. However, these frameworks have nothing to say about the effects of a linguistic choice on how a message is interpreted. I run 3 experiments which show that while UID posits no specific consequences to being "overinformative" (including more information in an utterance than is necessary), in fact overinformativeness can trigger pragmatic inferences which alter comprehenders' background beliefs about the world. In this case, I show that the Rational Speech Act (RSA) model, which models back-and-forth pragmatic reasoning between speakers and comprehenders, predicts both efficiency-based utterance choices, as well as any consequent change in perceived meaning. I also provide evidence that it's critical to model communication as a lossy process (which UID assumes), which allows the RSA model to account for phenomena that it otherwise is not able to. I further show that while UID predicts increased use of pronouns when referring to more contextually predictable referents, existing research does not unequivocally support this. I run 2 experiments which fail to show evidence that speakers use reduced expressions for predictable elements. In contrast to UID and similar frameworks, the RSA model can straightforwardly predict the results that have been observed to date. In the end, I argue that the RSA model is a highly attractive alternative for modeling speaker utterance choice at the discourse level. When it reflects communication as a lossy process, it is able to predict the same predictability-driven utterance reduction that UID does. However, by additionally modeling back-and-forth pragmatic reasoning, it successfully models utterance choice phenomena that simpler frameworks cannot account for.

---

Diese Arbeit erforscht neue Methoden, linguistische Ph{\"a}nomene auf Gespr{\"a}chsebene per Computermodellierung zu erfassen. Effiziente Sprecher:innen entscheiden sich bei der Kommunikation h{\"a}ufig dazu, wenn immer es m{\"o}glich ist, {\"A}u{\ss}erungen entweder ganz auszulassen oder aber ihre L{\"a}nge zu reduzieren. Modelle wie Uniform Information Density (UID) argumentieren, dass Sprecher:innen vorzugsweise diejenigen Elemente auslassen, die im jeweiligen Kontext vorhersagbarer und einfacher wiederherzustellen sind. Allerdings sagen diese Modelle nichts {\"u}ber die Auswirkungen einer linguistischen Entscheidung bez{\"u}glich der Interpretation einer Nachricht aus. Ich f{\"u}hre drei Experimente durch, die zeigen, dass wenngleich UID keine spezifischen Auswirkungen von "{\"U}berinformation" (einer {\"A}u{\ss}erung mehr Information als n{\"o}tig geben) postuliert, {\"U}berinformationen doch pragmatische Schlussfolgerungen, die das gedankliche Weltmodell der Versteher:innen {\"a}ndern k{\"o}nnen, ausl{\"o}st. F{\"u}r diesen Fall zeige ich, dass das Rational-Speech-Act-Modell (RSA), welches pragmatische Hin-und-Her-Schlussfolgerungen zwischen Sprecher:innen und Versteher:innen modelliert, sowohl effizienzbasierte {\"A}u{\ss}erungsauswahl als auch jegliche resultierende Verst{\"a}ndnis{\"a}nderung vorhersagt. Ich liefere auch Anhaltspunkte daf{\"u}r, dass es entscheidend ist, Kommunikation als verlustbehafteten Prozess zu modellieren (wovon UID ausgeht), was es dem RSA-Modell erlaubt, Ph{\"a}nomene einzubeziehen, wozu es sonst nicht in der Lage w{\"a}re. Weiterhin zeige ich, dass obschon UID beim Bezug auf kontextuell vorhersagbarere Bezugsw{\"o}rter eine erh{\"o}hte Nutzung von Pronomen vorhersagt, dies von existierender Forschung nicht einstimmig gest{\"u}tzt wird. Ich f{\"u}hre zwei Experimente durch, die keine Anhaltspunkte daf{\"u}r, dass Sprecher:innen reduzierte Ausdr{\"u}cke f{\"u}r vorhersagbare Elemente verwenden, finden. Im Gegensatz zu UID und {\"a}hnlichen Modellen kann dass RSA-Modell direkt die bislang beobachteten Resultate vorhersagen. Schlie{\ss}lich lege ich dar, warum das RSA-Modell eine h{\"o}chst attraktive Alternative zur Modellierung von Sprach{\"a}u{\ss}erungsentscheidungen auf Gespr{\"a}chsebene ist. Wenn es Kommunikation als einen verlustbehafteten Prozess widerspiegelt, kann es dieselbe vorhersagebasierte {\"A}u{\ss}erungsreduktion vorhersagen wie auch UID. Modelliert man jedoch zus{\"a}tzlich pragmatische Hin-und-Her-Schlussfolgerungen, modelliert RSA erfolgreich Ph{\"a}nomene bei {\"A}u{\ss}erungsentscheidungen, die einfachere Modelle nicht abbilden k{\"o}nnen.},
pubstate = {published},
type = {phdthesis}
}

Copy BibTeX to Clipboard

There is now a well-established literature showing that people anticipate upcoming concepts and words during language processing. Commonsense knowledge about typical event sequences and verbal selectional preferences can contribute to anticipating what will be mentioned next. We here investigate how temporal discourse connectives (before, after), which signal event ordering along a temporal dimension, modulate predictions for upcoming discourse referents. Our study analyses anticipatory gaze in the visual world and supports the idea that script knowledge, temporal connectives (before eating → menu, appetizer), and the verb’s selectional preferences (order → appetizer) jointly contribute to shaping rapid prediction of event participants.

• https://doi.org/10.1080/0163853X.2021.1930807
• https://doi.org/10.1080/0163853X.2021.1930807

@article{zarcone2021script,
title = {Interaction of script knowledge and temporal discourse cues in a visual world study},
author = {Alessandra Zarcone and Vera Demberg},
url = {https://doi.org/10.1080/0163853X.2021.1930807},
doi = {https://doi.org/10.1080/0163853X.2021.1930807},
year = {2021},
date = {2021-07-26},
journal = {Discourse Processes},
pages = {1-16},
publisher = {Routledge},
abstract = {There is now a well-established literature showing that people anticipate upcoming concepts and words during language processing. Commonsense knowledge about typical event sequences and verbal selectional preferences can contribute to anticipating what will be mentioned next. We here investigate how temporal discourse connectives (before, after), which signal event ordering along a temporal dimension, modulate predictions for upcoming discourse referents. Our study analyses anticipatory gaze in the visual world and supports the idea that script knowledge, temporal connectives (before eating → menu, appetizer), and the verb’s selectional preferences (order → appetizer) jointly contribute to shaping rapid prediction of event participants.},
pubstate = {published},
type = {article}
}

Copy BibTeX to Clipboard

The Uniform Information Density hypothesis (UID) predicts that lexical choice between long and short word forms depends on the predictability of the referent in context, and recent studies have shown such an effect of predictability on lexical choice during online production. We here set out to test whether the UID predictions hold up in a related setting, but different language (German) and different phenomenon, namely the choice between compounds (e.g. Badewanne / bathtub) or their base forms (Wanne / tub). Our study is consistent with the UID: we find that participants choose the shorter base form more often in predictive contexts, showing an active tendency to be information-theoretically efficient.

• https://escholarship.org/uc/item/3w6451rz

@inproceedings{Zarcone2021,
title = {A bathtub by any other name: the reduction of german compounds in predictive contexts},
author = {Alessandra Zarcone and Vera Demberg},
url = {https://escholarship.org/uc/item/3w6451rz},
year = {2021},
date = {2021},
booktitle = {Proceedings of the Annual Meeting of the Cognitive Science Society},
abstract = {The Uniform Information Density hypothesis (UID) predicts that lexical choice between long and short word forms depends on the predictability of the referent in context, and recent studies have shown such an effect of predictability on lexical choice during online production. We here set out to test whether the UID predictions hold up in a related setting, but different language (German) and different phenomenon, namely the choice between compounds (e.g. Badewanne / bathtub) or their base forms (Wanne / tub). Our study is consistent with the UID: we find that participants choose the shorter base form more often in predictive contexts, showing an active tendency to be information-theoretically efficient.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

A long-standing question in language understanding is whether pragmatic inferences are effortful or whether they happen seamlessly without measurable cognitive effort. We here measure the strength of particularized pragmatic inferences in a setting with high vs. low cognitive load. Cognitive load is induced by a secondary dot tracking task.

If this type of pragmatic inference comes at no cognitive processing cost, inferences should be similarly strong in both the high and the low load condition. If they are effortful, we expect a smaller effect size in the dual tasking condition. Our results show that participants who have difficulty in dual tasking (as evidenced by incorrect answers to comprehension questions) exhibit a smaller pragmatic effect when they were distracted with a secondary task in comparison to the single task condition. This finding supports the idea that pragmatic inferences are effortful.

• cogsci20_proceedings_ryzhova     (0.8MB)
• https://www.semanticscholar.org/paper/Processing-particularized-pragmatic-inferences-load-Ryzhova-Demberg/a5b8d4c72590eaaf965d91d8fafa2495f680313d

@inproceedings{Ryzhova2020,
title = {Processing particularized pragmatic inferences under load},
author = {Margarita Ryzhova and Vera Demberg},
url = {https://www.semanticscholar.org/paper/Processing-particularized-pragmatic-inferences-load-Ryzhova-Demberg/a5b8d4c72590eaaf965d91d8fafa2495f680313d},
year = {2020},
date = {2020-10-17},
booktitle = {Proceedings of the 42nd Annual Meeting of the Cognitive Science Society (CogSci 2020)},
abstract = {A long-standing question in language understanding is whether pragmatic inferences are effortful or whether they happen seamlessly without measurable cognitive effort. We here measure the strength of particularized pragmatic inferences in a setting with high vs. low cognitive load. Cognitive load is induced by a secondary dot tracking task. If this type of pragmatic inference comes at no cognitive processing cost, inferences should be similarly strong in both the high and the low load condition. If they are effortful, we expect a smaller effect size in the dual tasking condition. Our results show that participants who have difficulty in dual tasking (as evidenced by incorrect answers to comprehension questions) exhibit a smaller pragmatic effect when they were distracted with a secondary task in comparison to the single task condition. This finding supports the idea that pragmatic inferences are effortful.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Automatically generated stories need to be not only coherent, but also interesting. Apart from realizing a story line, the text also needs to include rich details to engage the readers. We propose a model that features two different generation components: an outliner, which proceeds the main story line to realize global coherence; a detailer, which supplies relevant details to the story in a locally coherent manner. Human evaluations show our model substantially improves the informativeness of generated text while retaining its coherence, outperforming various baselines.

• https://www.aclweb.org/anthology/2020.coling-main.212
• https://doi.org/10.18653/v1/2020.coling-main.212

@inproceedings{zhai-etal-2020-story,
title = {Story Generation with Rich Details},
author = {Fangzhou Zhai and Vera Demberg and Alexander Koller},
url = {https://www.aclweb.org/anthology/2020.coling-main.212},
doi = {https://doi.org/10.18653/v1/2020.coling-main.212},
year = {2020},
date = {2020},
booktitle = {Proceedings of the 28th International Conference on Computational Linguistics (CoLing 2020)},
pages = {2346-2351},
publisher = {International Committee on Computational Linguistics},
address = {Barcelona, Spain (Online)},
abstract = {Automatically generated stories need to be not only coherent, but also interesting. Apart from realizing a story line, the text also needs to include rich details to engage the readers. We propose a model that features two different generation components: an outliner, which proceeds the main story line to realize global coherence; a detailer, which supplies relevant details to the story in a locally coherent manner. Human evaluations show our model substantially improves the informativeness of generated text while retaining its coherence, outperforming various baselines.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

While people process text, they make frequent use of information that is assumed to be common ground and left implicit in the text. One important type of such commonsense knowledge is script knowledge, which is the knowledge about the events and participants in everyday activities such as visiting a restaurant. Due to its implicitness, it is hard for machines to exploit such script knowledge for natural language processing (NLP). This dissertation addresses the role of script knowledge in a central field of NLP, natural language understanding (NLU). In the first part of this thesis, we address script parsing. The idea of script parsing is to align event and participant mentions in a text with an underlying script representation. This makes it possible for a system to leverage script knowledge for downstream tasks. We develop the first script parsing model for events that can be trained on a large scale on crowdsourced script data. The model is implemented as a linear-chain conditional random field and trained on sequences of short event descriptions, implicitly exploiting the inherent event ordering information. We show that this ordering information plays a crucial role for script parsing. Our model provides an important first step towards facilitating the use of script knowledge for NLU. In the second part of the thesis, we move our focus to an actual application in the area of NLU, i.e. machine comprehension. For the first time, we provide data sets for the systematic evaluation of the contribution of script knowledge for machine comprehension. We create MCScript, a corpus of narrations about everyday activities and questions on the texts. By collecting questions based on a scenario rather than a text, we aimed at creating challenging questions which require script knowledge for finding the correct answer. Based on the findings of a shared task carried out with the data set, which indicated that script knowledge is not relevant for good performance on our corpus, we revised the data collection process and created a second version of the data set.

• http://nbn-resolving.de/urn:nbn:de:bsz:291--ds-313016
• https://doi.org/10.22028/D291-31301

@phdthesis{Ostermann2019,
title = {Script knowledge for natural language understanding},
author = {Simon Ostermann},
url = {http://nbn-resolving.de/urn:nbn:de:bsz:291--ds-313016},
doi = {https://doi.org/10.22028/D291-31301},
year = {2019},
date = {2019},
school = {Saarland University},
address = {Saarbruecken, Germany},
abstract = {While people process text, they make frequent use of information that is assumed to be common ground and left implicit in the text. One important type of such commonsense knowledge is script knowledge, which is the knowledge about the events and participants in everyday activities such as visiting a restaurant. Due to its implicitness, it is hard for machines to exploit such script knowledge for natural language processing (NLP). This dissertation addresses the role of script knowledge in a central field of NLP, natural language understanding (NLU). In the first part of this thesis, we address script parsing. The idea of script parsing is to align event and participant mentions in a text with an underlying script representation. This makes it possible for a system to leverage script knowledge for downstream tasks. We develop the first script parsing model for events that can be trained on a large scale on crowdsourced script data. The model is implemented as a linear-chain conditional random field and trained on sequences of short event descriptions, implicitly exploiting the inherent event ordering information. We show that this ordering information plays a crucial role for script parsing. Our model provides an important first step towards facilitating the use of script knowledge for NLU. In the second part of the thesis, we move our focus to an actual application in the area of NLU, i.e. machine comprehension. For the first time, we provide data sets for the systematic evaluation of the contribution of script knowledge for machine comprehension. We create MCScript, a corpus of narrations about everyday activities and questions on the texts. By collecting questions based on a scenario rather than a text, we aimed at creating challenging questions which require script knowledge for finding the correct answer. Based on the findings of a shared task carried out with the data set, which indicated that script knowledge is not relevant for good performance on our corpus, we revised the data collection process and created a second version of the data set.},
pubstate = {published},
type = {phdthesis}
}

Copy BibTeX to Clipboard

According to Grice’s (1975) theory of pragmatics, people tend to omit basic information when participating in a conversation (or writing a narrative) under the assumption that left out details are already known or can be inferred from commonsense knowledge by the hearer (or reader). Writing and understanding of texts makes particular use of a specific kind of common-sense knowledge, referred to as script knowledge. Schank and Abelson (1977) proposed Scripts as a model of human knowledge represented in memory that stores the frequent habitual activities, called scenarios, (e.g. eating in a fast food restaurant, etc.), and the different courses of action in those routines. This thesis addresses measures to provide a sound empirical basis for high-quality script models. We work on three key areas related to script modeling: script knowledge acquisition, script induction and script identification in text. We extend the existing repository of script knowledge bases in two different ways. First, we crowdsource a corpus of 40 scenarios with 100 event sequence descriptions (ESDs) each, thus going beyond the size of previous script collections. Second, the corpus is enriched with partial alignments of ESDs, done by human annotators. The crowdsourced partial alignments are used as prior knowledge to guide the semi-supervised script-induction algorithm proposed in this dissertation. We further present a semi-supervised clustering approach to induce script structure from crowdsourced descriptions of event sequences by grouping event descriptions into paraphrase sets and inducing their temporal order. The proposed semi-supervised clustering model better handles order variation in scripts and extends script representation formalism, Temporal Script graphs, by incorporating „arbitrary order“ equivalence classes in order to allow for the flexible event order inherent in scripts. In the third part of this dissertation, we introduce the task of scenario detection, in which we identify references to scripts in narrative texts. We curate a benchmark dataset of annotated narrative texts, with segments labeled according to the scripts they instantiate. The dataset is the first of its kind. The analysis of the annotation shows that one can identify scenario references in text with reasonable reliability. Subsequently, we proposes a benchmark model that automatically segments and identifies text fragments referring to given scenarios. The proposed model achieved promising results, and therefore opens up research on script parsing and wide coverage script acquisition.

• http://nbn-resolving.de/urn:nbn:de:bsz:291--ds-301634
• https://doi.org/http://dx.doi.org/10.22028/D291-30163

@phdthesis{Wanzare2019,
title = {Script acquisition: a crowdsourcing and text mining approach},
author = {Lilian Diana Awuor Wanzare},
url = {http://nbn-resolving.de/urn:nbn:de:bsz:291--ds-301634},
doi = {https://doi.org/http://dx.doi.org/10.22028/D291-30163},
year = {2019},
date = {2019},
school = {Saarland University},
address = {Saarbruecken, Germany},
abstract = {According to Grice’s (1975) theory of pragmatics, people tend to omit basic information when participating in a conversation (or writing a narrative) under the assumption that left out details are already known or can be inferred from commonsense knowledge by the hearer (or reader). Writing and understanding of texts makes particular use of a specific kind of common-sense knowledge, referred to as script knowledge. Schank and Abelson (1977) proposed Scripts as a model of human knowledge represented in memory that stores the frequent habitual activities, called scenarios, (e.g. eating in a fast food restaurant, etc.), and the different courses of action in those routines. This thesis addresses measures to provide a sound empirical basis for high-quality script models. We work on three key areas related to script modeling: script knowledge acquisition, script induction and script identification in text. We extend the existing repository of script knowledge bases in two different ways. First, we crowdsource a corpus of 40 scenarios with 100 event sequence descriptions (ESDs) each, thus going beyond the size of previous script collections. Second, the corpus is enriched with partial alignments of ESDs, done by human annotators. The crowdsourced partial alignments are used as prior knowledge to guide the semi-supervised script-induction algorithm proposed in this dissertation. We further present a semi-supervised clustering approach to induce script structure from crowdsourced descriptions of event sequences by grouping event descriptions into paraphrase sets and inducing their temporal order. The proposed semi-supervised clustering model better handles order variation in scripts and extends script representation formalism, Temporal Script graphs, by incorporating "arbitrary order" equivalence classes in order to allow for the flexible event order inherent in scripts. In the third part of this dissertation, we introduce the task of scenario detection, in which we identify references to scripts in narrative texts. We curate a benchmark dataset of annotated narrative texts, with segments labeled according to the scripts they instantiate. The dataset is the first of its kind. The analysis of the annotation shows that one can identify scenario references in text with reasonable reliability. Subsequently, we proposes a benchmark model that automatically segments and identifies text fragments referring to given scenarios. The proposed model achieved promising results, and therefore opens up research on script parsing and wide coverage script acquisition.},
pubstate = {published},
type = {phdthesis}
}

Copy BibTeX to Clipboard

We introduce MCScript2.0, a machine comprehension corpus for the end-to-end evaluation of script knowledge. MCScript2.0 contains approx. 20,000 questions on approx. 3,500 texts, crowdsourced based on a new collection process that results in challenging questions. Half of the questions cannot be answered from the reading texts, but require the use of commonsense and, in particular, script knowledge. We give a thorough analysis of our corpus and show that while the task is not challenging to humans, existing machine comprehension models fail to perform well on the data, even if they make use of a commonsense knowledge base.

• https://www.aclweb.org/anthology/S19-1012

@inproceedings{ostermann2019mcscript2,
title = {MCScript2.0: A Machine Comprehension Corpus Focused on Script Events and Participants},
author = {Simon Ostermann and Michael Roth and Manfred Pinkal},
url = {https://www.aclweb.org/anthology/S19-1012},
year = {2019},
date = {2019-10-17},
booktitle = {Proceedings of the Eighth Joint Conference on Lexical and Computational Semantics (* SEM 2019)},
pages = {103-117},
abstract = {We introduce MCScript2.0, a machine comprehension corpus for the end-to-end evaluation of script knowledge. MCScript2.0 contains approx. 20,000 questions on approx. 3,500 texts, crowdsourced based on a new collection process that results in challenging questions. Half of the questions cannot be answered from the reading texts, but require the use of commonsense and, in particular, script knowledge. We give a thorough analysis of our corpus and show that while the task is not challenging to humans, existing machine comprehension models fail to perform well on the data, even if they make use of a commonsense knowledge base.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Automatically generating globally coherent stories is a challenging problem. Neural text generation models have been shown to perform well at generating fluent sentences from data, but they usually fail to keep track of the overall coherence of the story after a couple of sentences. Existing work that incorporates a text planning module succeeded in generating recipes and dialogues, but appears quite data-demanding. We propose a novel story generation approach that generates globally coherent stories from a fairly small corpus. The model exploits a symbolic text planning module to produce text plans, thus reducing the demand of data; a neural surface realization module then generates fluent text conditioned on the text plan. Human evaluation showed that our model outperforms various baselines by a wide margin and generates stories which are fluent as well as globally coherent.

• https://aclanthology.org/W19-3404
• https://doi.org/10.18653/v1/W19-3404

@inproceedings{Fangzhou2019,
title = {A Hybrid Model for Globally Coherent Story Generation},
author = {Fangzhou Zhai and Vera Demberg and Pavel Shkadzko and Wei Shi and Asad Sayeed},
url = {https://aclanthology.org/W19-3404},
doi = {https://doi.org/10.18653/v1/W19-3404},
year = {2019},
date = {2019},
booktitle = {Proceedings of the Second Workshop on Storytelling},
pages = {34-45},
publisher = {Association for Computational Linguistics},
address = {Florence, Italy},
abstract = {Automatically generating globally coherent stories is a challenging problem. Neural text generation models have been shown to perform well at generating fluent sentences from data, but they usually fail to keep track of the overall coherence of the story after a couple of sentences. Existing work that incorporates a text planning module succeeded in generating recipes and dialogues, but appears quite data-demanding. We propose a novel story generation approach that generates globally coherent stories from a fairly small corpus. The model exploits a symbolic text planning module to produce text plans, thus reducing the demand of data; a neural surface realization module then generates fluent text conditioned on the text plan. Human evaluation showed that our model outperforms various baselines by a wide margin and generates stories which are fluent as well as globally coherent.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Human world knowledge contains information about prototypical events and their participants and locations. In this paper, we train the first models using multi-task learning that can both predict missing event participants and also perform semantic role classification based on semantic plausibility. Our best-performing model is an improvement over the previous state-of-the-art on thematic fit modelling tasks. The event embeddings learned by the model can additionally be used effectively in an event similarity task, also outperforming the state-of-the-art.

• https://aclanthology.org/S18-2002
• https://doi.org/10.18653/v1/S18-2002

@inproceedings{Hong2018,
title = {Learning Distributed Event Representations with a Multi-Task Approach},
author = {Xudong Hong and Asad Sayeed and Vera Demberg},
url = {https://aclanthology.org/S18-2002},
doi = {https://doi.org/10.18653/v1/S18-2002},
year = {2018},
date = {2018},
booktitle = {Proceedings of the Seventh Joint Conference on Lexical and Computational Semantics},
pages = {11-21},
publisher = {Association for Computational Linguistics},
address = {New Orleans, USA},
abstract = {Human world knowledge contains information about prototypical events and their participants and locations. In this paper, we train the first models using multi-task learning that can both predict missing event participants and also perform semantic role classification based on semantic plausibility. Our best-performing model is an improvement over the previous state-of-the-art on thematic fit modelling tasks. The event embeddings learned by the model can additionally be used effectively in an event similarity task, also outperforming the state-of-the-art.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard

Commonsense knowledge as provided by scripts is crucially relevant for text understanding systems, providing a basis for commonsense inference. This paper considers a relevant subtask of script-based text understanding, the task of mapping event mentions in a text to script events. We focus on script representations where events are associated with paraphrase sets, i.e. sets of crowdsourced event descriptions. We provide a detailed annotation of event mention/description pairs with textual entailment types. We demonstrate that representing events in terms of paraphrase sets can massively improve the performance of text-to-script mapping systems. However, for a residual substantial fraction of cases, deeper inference is still required.

• https://www.semanticscholar.org/paper/Mapping-Texts-to-Scripts%3A-An-Entailment-Study-Ostermann-Seitz/7970ec54afb3d78d9f061a38db27d0bd19e215d5

@inproceedings{MCScriptb,
title = {Mapping Text to Scripts: An Entailment Study},
author = {Simon Ostermann and Hannah Seitz and Stefan Thater and Manfred Pinkal},
url = {https://www.semanticscholar.org/paper/Mapping-Texts-to-Scripts%3A-An-Entailment-Study-Ostermann-Seitz/7970ec54afb3d78d9f061a38db27d0bd19e215d5},
year = {2018},
date = {2018},
booktitle = {Proceedings of the 11th International Conference on Language Resources and Evaluation (LREC 2018)},
address = {Miyazaki, Japan},
abstract = {Commonsense knowledge as provided by scripts is crucially relevant for text understanding systems, providing a basis for commonsense inference. This paper considers a relevant subtask of script-based text understanding, the task of mapping event mentions in a text to script events. We focus on script representations where events are associated with paraphrase sets, i.e. sets of crowdsourced event descriptions. We provide a detailed annotation of event mention/description pairs with textual entailment types. We demonstrate that representing events in terms of paraphrase sets can massively improve the performance of text-to-script mapping systems. However, for a residual substantial fraction of cases, deeper inference is still required.},
pubstate = {published},
type = {inproceedings}
}

Copy BibTeX to Clipboard