Projects

Please cick on the respective entry to show detailed information. An overview of my DFG-funded project is available at gepris.

▶ Neural Language Generation Conditioned on Emotions (NLGCE, 2021-2024)

Intelligent agents as they are implemented in chat bots or assistant systems are successful in solving predefined tasks. However, they lack the capability to behave empathetically, which includes the estimation of the emotional state of the interlocutor and generating appropriate responses. While different research projects exist which focus on the analysis of emotions in human, including the analysis of body postures, textual information, or prosody, there is only a limited amount of work on the generation of emotionally appropriate responses. One reason is that emotions are on the fence between content and style and therefore challenging to be injected in content-wise conditioned responses, based on the dialogue setting or task to be achieved. Further, emotions can be expressed in different ways, including the description of an emotionally connotated event, a bodily reaction, an expression, or a direct report of a subjective feeling. This Ph.D.\ thesis project explores this novel challenge in natural language generation (NLG), namely to condition the generated text not only on the task, content, and dialogue context, but in addition formulate the generated text such that it is appropriate emotionally for the interlocutor. To address the full variety and complexity of emotion expressions, we will integrate psychological theories on event appraisal, emotion regulation and emotional components. We will therefore develop computational methods which will allow intelligent agents to react in the appropriate emotional tone. This will, for the first time, allow smart agents to be perceived as being empathic, not only based on prosody, but also based on content.

  • Role: Supervisor
  • Scholarship Grantee: Yarik Menchaca-Resendiz
  • Funding: DAAD
▶ Style Transfer of Psychological Concepts (STPC, 2021-)
▶ Computational Event Evaluation based on Appraisal Theories for Emotion Analysis (CEAT, 2021-2024)

Emotion analysis has typically been formulated as text classification task in which predefined emotion labels are assigned to textual units. The label set commonly follows the set of basic emotions as proposed by Ekman (Anger, Fear, Joy, Surprise, Sadness, Disgust) or Plutchik (adding Trust and Anticipation) or the valence-arousal-dominance model. This constitutes a gap between the state of research in psychology and computational linguistics, as the appraisal theories are widely accepted, but have not been used so far for emotion analysis in text. With CEAT, we fill this gap and develop computational models of the cognitive appraisal of events and, to a lesser degree, of bodily symptoms and action tendencies. To represent the cognitive appraisal, we build on top of Smith/Ellsworth’s (1985) work who show that the variables pleasantness, responsibility, certainty, attention, effort and situational control are sufficient to discriminate between a set of 15 emotions. In this project, we create two approaches to assign these appraisal dimensions to textual event descriptions, firstly by building on top of semantic parsing and secondly in a deep learning setting. Based on these dimensions, we then predict the emotion associated with the textual fragment. This will lead to models that can automatically assign an emotion to an event description, even if no emotion words or self reports of feeling are available.

  • Role: Principle Investigator
  • Other Project Members: Laura Oberlaender, Enrica Troiano, Maximilian Wegge
  • Funding: DFG
  • Selected papers describing the results: 1 2 3 4 5 6 7 8
▶ Automatic Fact-Checking of Biomedical Information in Social Media and Scientific Literature (FIBISS, 2021-2024)

Most research on methods and models for automatic fact checking, which can distinguish misinformation and desinformation from correct information, focus on the news domain. News, including those shared in social media spaces, are checked for their truthfulness.Such methods have not been developed for the biomedical domain yet. Challenges include the richness of (established) sources of information, the complexity of information, as well as the differences between the language of experts and medical laypeople.In this project, we develop information extraction systems for laypeople and expert language, map the extracted information onto each other and finally check their truthfulness, based on established sources.The project combines therefore methods from transfer learning, information extraction, and fact checking for the biomedical domain, especially in social media.

  • Role: Principle Investigator
  • Other Project Members: Amelie Wuehrl, Lara Grimminger, Lynn Greschner
  • Funding: DFG
  • Selected papers describing the results: 1 2 3 4 5

Previous

▶ Emotion and Argument in Digital Information Spaces (EmoArg, 2018-2022)
  • Role: Co-Principle Investigator
  • PIs: Kai Sassenberg (Leibniz-Institut für Wissensmedien) and Sebastian Padó (Uni Stuttgart)
  • Other Project Members: Enrica Troiano (Ph.D. student)
  • Funding: University of Stuttgart
  • Selected papers describing the results: 1 2 3 4

The goal of this project is to develop a better understanding of the impact of emotion on the perception of facts. This aspect is for instance relevant in the distribution of fake news. We further develop methods which are able to separate factual content of statements from the emotional connotation.

More information:

▶ Structured Multi-Domain Emotion Analysis from Text (SEAT, 2018-2021)

Emotion analysis in natural language processings aims at associating text with emotions, for instance with anger, fear, joy, surprise, disgust or sadness. This task extends sentiment analysis, which adds further qualitative value in applications, for instance in social media analysis, in the analysis of fictional stories or news articles.

Existing research has so far mainly focused on the association of text with specific emotion models from psychological research. The development of methods for detecting phrases in text which denote the emotion experiencer (the character or person who feels the emotion), the emotion theme (the cause of the development of an emotion) as well as the modifiers of an emotion (intensifiers and diminishers) has been neglected.

In this project, we aim at filling this gap. We will develop manually annotated corpora from different domains (news, novels, social media) in German and English. Based on these resources, we develop models which are able to automatically recognize and extract such information. We work on different levels: Firstly, we connect words with emotions (with distributional and lexical methods), including grammatical variants. Then, secondly, we analyze these mentions in context with modifiers, the feeler and the theme (cause) of the emotion. Thirdly, we model these information in context, i.e., beyond seperated mentions. All methods will be analyzed regarding their domain and language independence.

  • Role: Principle Investigator
  • Other Project Members: Laura Bostan, Evgeny Kim (Ph.D. students)
  • Funding: DFG
  • Selected papers describing the results: 1 2 3 4 5 6 7 8 9 10 11
▶ Comprehensive Modeling of Conversational Contributions in Prose Texts (QUOTE, 2017-2020)

In many kinds of prose texts, both literary or newswire texts, reported speech plays an important role as a source of information aboutcharacters, their attitudes, and their relationships. Going further,such information can aid in the analysis of patterns of behavior and theconstruction of social networks.While readers do not have any problem in assembling representations forcomplete situations from individual instances of reported speech, thisis still a challenging task for computers. Current state of the artmethods are generally organized as “pipelines” which start fromindividual instances of reported speech and proceed incrementally tomore global properties of the situation or characters. Since individualinstances of reported speech are often short and uninformative, apipeline procedure often causes prediction errors which cannot berectified in retrospect.In this project, we develop joint inference methods to model the variousaspects of reported speech (who is the speaker? the hearer? What is thecontent? What is the relationship between speaker and hearer?) togetherinstead of individually. The resulting joint model takes account of theinterdependencies between these aspects. Thus, information from thedifferent aspects can complement each other. The result of this part ofthe project is a solid starting place (in terms of natural languageprocessing methods) for the application of such methods for theautomatic analysis of reported speech in digital humanities and socialsciences.This algorithmic goal is complemented by a goal from corpus andcomputational linguistics, namely elucidating the relationship betweenreported speech and other aspects of semantic analysis. In particular,there appears to be a close relationship between reported speech and (asubset) of semantic roles. Yet, no comprehensive formal analysis hasbeen carried out so far. We will provide a linguistic characterizationof the relationship and exploit it algorithmically to further improvethe recognition of reported speech as discussed above. The results ofthis part of the project is the (at least partial) consolidation of twostrands of research that have largely been treated as independent sofar.

  • Role: Co-Principle Investigator
  • Principle Investigator: Sebastian Padó
  • Other Project Members: Sean Papay (Ph.D. student)
  • Funding: DFG
  • Selected papers describing the results: 1 2 3 4 5
▶ PSINK - Automatische Erstellung einer Wissensbasis zur Unterstuetzung der Translation von der praeklinischen Forschung in die klinische Anwendung bei Rueckenmarksverletzungen (PSINK, 2016-2020)
  • Role: Co-Principle Investigator and Proposal Author
  • Principle Investigator: Philipp Cimiano and Hans-Werner Mueller
  • Other Project Members: Nicole Brazda (Co-PI), Hendrik ter Horst, Veronica Estrada, Jessica Schira, Christian Ohmann
  • Former Project Members: Matthias Hartung
  • Funding: BMBF, i:DSem

We develop automatic information extraction methods to populate a database of preclinical experiments in the domain of spinal cord injuries.

Partners are:

More information:

▶ Center for Reflective Text Analysis (CRETA, 2016--2021)

The Chair of Theoretical Computational Linguistics is a partner in this BMBF-funded project in which we work on emotion analysis for literature analysis. Partners at University of Stuttgart are: Institut für Literaturwissenschaft / Germanistische Mediävistik, Institut für Visualisierung und Interaktive Systeme, Historisches Institut / Landesgeschichte, Institut für Sozialwissenschaften / Internationale Beziehungen und Europäische Integration, Institut für Maschinelle Sprachverarbeitung, Institut für Philosophie / Wissenschaftstheorie und Technikphilosophie, Institut für Literaturwissenschaft / Neuere Deutsche Literatur, Institut für Linguistik / Romanistik, Institut für Literaturwissenschaft / Digital Humanities, Stuttgart Research Centre for Text Studies

More information, including publications:

▶ Confidence Estimation for Biomedical Information Extraction (KABI, 2016--2018)
  • Role: Principle Investigator
  • Other project members: Camilo Thorne (Postdoc)
  • Funding: MWK Baden-Württemberg and University of Stuttgart

In the Life Sciences, most information is only available in free text in scientific publications. Automatic methods to extract such knowledge and to provide it in structured databases is challenged by a dilemma: Especially if potentially new information is detected in text, it is unclear if this information is actually correct or if it is wrongly extracted, for instance because the text is formulated in an uncommon way. In this project, methods will be developed which help to estimate the reliability of extracted information from biomedical publications.

▶ Recipe Classification (2016)
  • Role: Principle Investigator
  • Other project members: Christian Scheible, Hanna Kicherer
  • Funding: Chefkoch GmbH
▶ It's OWL (2013 – 2014)
  • Role: Postdoc in Project MMI
  • Funding: German Federal Ministry of Research and Education (BMBF)
▶ Sentiment Analysis for Distance Education Evaluation (SADE, 2014)
  • Role: Principle Investigator
  • Funding: Online Akademie GmbH
▶ +Spaces (2011 – 2012)
▶ Aneurist (2006 – 2010)
▶ Learning and Inference Platform (2006 – 2008)
  • Role: Ph.D. student
  • Funding: Fraunhofer and Max-Planck Societies
  • Other partners: MPG-FhG Collaboration Project with Fraunhofer FIRST, SCAI, IAIS, ITWM and Max-Planck Friedrich Miescher Laboratory, Institute for biological Cybernetics, Institute for Informatics, Institute for Molecular Genetics

  • Learning and Inference Platform