Svetlana Kiritchenko, Ph.D.

National Research Council Canada

• Home
• CV
• Research
• Publications
• Resources

Research Projects:

• Ethics
  
  • Fairness in NLP
  • Explainable NLP
  • Ethical Challenges in Abusive Language Detection
• Sentiment and Emotion Analysis and Applications
  
  • Abusive Language Detection
  • Examining the State of Being Alone (SOLO)
  • WikiArt Emotions Dataset
  • Affect in Tweets
  • Stance
  • Best-Worst Scaling
  • Sentiment Analysis in Arabic
  • Personality Detection from Text
  • Automatic Sentiment Analysis Systems
  • Large-Scale Automatically Generated Sentiment Lexicons
• Medical Informatics
  
  • Social Media Mining for Health-Related Applications
  • Clinical Records Mining
  • Biomedical Information Extraction from Journal Articles
• Hierarchical Classification
• Semi-Supervised Learning
  
  • Email Classification with Co-Training

Ethics

• Fairness in NLP
  
  • Equity Evaluation Corpus: 8,640 English sentences carefully chosen to tease out biases towards certain races and genders. The sentences were constructed from templates using the pre-defined sets of noun phrases corresponding to males and females as well as first names associated with African American and European American males and females. We used the dataset to examine 219 automatic sentiment analysis systems that took part in a recent shared task, SemEval-2018 Task 1 'Affect in Tweets'. We found that several of the systems showed statistically significant bias; that is, they consistently provide slightly higher sentiment intensity predictions for one race or one gender.
    Svetlana Kiritchenko and Saif Mohammad. (2018). Examining Gender and Race Bias in Two Hundred Sentiment Analysis Systems. In Proceedings of the 7th Joint Conference on Lexical and Computational Semantics (*SEM), New Orleans, USA, 2018 [pdf] [data]
  • Biases in Vision-Language Systems: We investigate bias and diversity in outputs of state-of-the-art text-to-image and large vision-language systems.
    Kathleen C. Fraser and Svetlana Kiritchenko. (2024) Examining Gender and Racial Bias in Large Vision--Language Models Using a Novel Dataset of Parallel Images. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (EACL), Malta, March 2024. [paper]
    Kathleen C. Fraser, Svetlana Kiritchenko, and Isar Nejadgholi. (2023) Diversity is Not a One-Way Street: Pilot Study on Ethical Interventions for Racial Bias in Text-to-Image Systems. In Proceedings of the 14th International Conference on Computational Creativity (ICCC), Waterloo, ON, Canada, June 2023. Best Short Paper Award [pdf]
    Kathleen C. Fraser, Isar Nejadgholi, and Svetlana Kiritchenko. (2023) A Friendly Face: Do Text-to-Image Systems Rely on Stereotypes when the Input is Under-Specified? In Proceedings of the Creative AI Across Modalities Workshop (CreativeAI @ AAAI), Washington, DC, USA, Feb. 2023. [pdf]
• Explainable NLP
  
  • Explainability for Fairness: Motivations for methods in explainable artificial intelligence (XAI) often include detecting, quantifying and mitigating bias, and contributing to making machine learning models fairer. However, exactly how an XAI method can help in combating biases is often left unspecified. We investigate the barriers preventing XAI methods from being used more widely in tackling fairness issues.
    Esma Balkir, Svetlana Kiritchenko, Isar Nejadgholi, and Kathleen C. Fraser. (2022) Challenges in Applying Explainability Methods to Improve the Fairness of NLP Models. In Proceedings of the Second Workshop on Trustworthy Natural Language Processing (TrustNLP @ NAACL), Seattle, WA, USA, July 2022. [pdf]
  • Necessity and Sufficiency for Explaining Text Classifiers: We present a novel feature attribution method for explaining text classifiers, and analyze it in the context of hate speech detection. Unlike previous feature attribution models, which usually provide a single importance score for each token, we instead provide two complementary and theoretically-grounded scores -- necessity and sufficiency -- resulting in more informative explanations.
    Esma Balkir, Isar Nejadgholi, Kathleen C. Fraser, and Svetlana Kiritchenko. (2022). Necessity and Sufficiency for Explaining Text Classifiers: A Case Study in Hate Speech Detection. In Proceedings of the Annual Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL), Seattle, WA, USA, July 2022. [pdf][code]
  • TCAV for Text Classifiers: We propose an interpretability technique, based on the Testing Concept Activation Vector (TCAV) method, to quantify the sensitivity of a trained model to human-defined concepts, and use that to explain the generalizability and fairness of abusive language detection models:
    Isar Nejadgholi, Svetlana Kiritchenko, Kathleen C. Fraser, and Esma Balkir. (2023) Concept-Based Explanations to Test for False Causal Relationships Learned by Abusive Language Classifiers. In Proceedings of the 7th Workshop on Online Abuse and Harms (WOAH), Toronto, ON, Canada, July 2023. [pdf]
    Isar Nejadgholi, Esma Balkir, Kathleen C. Fraser, and Svetlana Kiritchenko. (2022) Towards Procedural Fairness: Uncovering Biases in How a Toxic Language Classifier Uses Sentiment Information. In Proceedings of the Workshop on Analyzing and Interpreting Neural Networks for NLP (BlackboxNLP), Abu Dhabi, United Arab Emirates, Dec. 2022. [pdf][code]
    Isar Nejadgholi, Kathleen C. Fraser, and Svetlana Kiritchenko. (2022). Improving Generalizability in Implicitly Abusive Language Detection with Concept Activation Vectors. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, Dublin, Ireland, May 2022. [pdf][code]
• Ethical Challenges in Abusive Language Detection
  
  • We review a large body of NLP research on automatic abuse detection with a focus on ethical challenges, organized around eight established ethical principles: privacy, accountability, safety and security, transparency and explainability, fairness and non-discrimination, human control of technology, professional responsibility, and promotion of human values. In many cases, these principles relate not only to situational ethical codes, which may be context-dependent, but are in fact connected to universal human rights, such as the right to privacy, freedom from discrimination, and freedom of expression. We highlight the need to examine the broad social impacts of this technology, and to bring ethical and human rights considerations to every stage of the application life-cycle, from task formulation and dataset design, to model training and evaluation, to application deployment. Guided by these principles, we identify several opportunities for rights-respecting, socio-technical solutions to detect and confront online abuse, including 'nudging', 'quarantining', value sensitive design, counter-narratives, style transfer, and AI-driven public education applications.
    Svetlana Kiritchenko, Isar Nejadgholi, and Kathleen C. Fraser. Confronting Abusive Language Online: A Survey from the Ethical and Human Rights Perspective. Journal of Artificial Intelligence Research, 71: 431-478, July 2021. [pdf]
  • We provide an explanatory view of how LLMs can apply moral reasoning to both criticize and defend sexist language:
    Rongchen Guo, Isar Nejadgholi, Hillary Dawkins, Kathleen C. Fraser, and Svetlana Kiritchenko. (2024) Adaptable Moral Stances of Large Language Models on Sexist Content: Implications for Society and Gender Discourse. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP 2024), Miami, Florida, USA, November 2024. [pdf]
  • We explore the topic bias and the task formulation bias in cross-dataset generalization on the task of online abuse detection:
    Isar Nejadgholi and Svetlana Kiritchenko. On Cross-Dataset Generalization in Automatic Detection of Online Abuse. In Proceedings of the 4th Workshop on Online Abuse and Harms at EMNLP-2020, November 2020. [pdf]
  • We propose a unified framework for online abuse detection as a two-step process. First, online content is categorized around personal and identity-related subject matters. Second, severity of abuse is identified through comparative annotation within each category. The novel framework is guided by the Ethics by Design principle and is a step towards building more accurate and trusted models.
    Svetlana Kiritchenko and Isar Nejadgholi. Towards Ethics by Design in Online Abusive Content Detection. NRC Technical Report, October 2020. [pdf]

Sentiment and Emotion Analysis and Applications

• Abusive Language Detection
  
  • Understanding and Countering Stereotypes: We present a computational approach to interpreting stereotypes in text through the Stereotype Content Model (SCM), a comprehensive causal theory from social psychology. Further, we explore various strategies to counter stereotypical beliefs.
    Kathleen C. Fraser, Svetlana Kiritchenko, and Isar Nejadgholi. (2024) How Does Stereotype Content Differ across Data Sources? In Proceedings of the 13th Joint Conference on Lexical and Computational Semantics (*SEM 2024), Mexico City, Mexico, June 2024. [pdf]
    Isar Nejadgholi, Kathleen C. Fraser, Anna Kerkhof, and Svetlana Kiritchenko. (2024) Challenging Negative Gender Stereotypes: A Study on the Effectiveness of Automated Counter-Stereotypes. In Proceedings of the Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), Torino, Italy, May 2024. [pdf] [data]
    Kathleen C. Fraser, Svetlana Kiritchenko, and Isar Nejadgholi. (2022). Computational Modelling of Stereotype Content in Text. Frontiers in Artificial Intelligence, April, 2022. [paper]
    Kathleen C. Fraser, Isar Nejadgholi, and Svetlana Kiritchenko (2021). Understanding and Countering Stereotypes: A Computational Approach to the Stereotype Content Model. In Proceedings of the Joint Conference of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (ACL-IJCNLP 2021), August 2021. [pdf]
    Kathleen C. Fraser, Svetlana Kiritchenko, Isar Nejadgholi, and Anna Kerkhof. (2023) What Makes a Good Counter-Stereotype? Evaluating Strategies for Automated Responses to Stereotypical Text. In Proceedings of the First Workshop on Social Influence in Conversations (SICon), Toronto, ON, Canada, July 2023. [pdf]
    Kathleen C. Fraser, Svetlana Kiritchenko, and Isar Nejadgholi. (2022). Extracting Age-Related Stereotypes from Social Media Texts. In Proceedings of the Language Resources and Evaluation Conference (LREC-2022), Marseille, France, June 2022. [pdf][project webpage]
  • Detecting and Countering Aporophobia: Aporophobia, a social bias against the poor, is a common phenomenon online, yet so far has been overlooked in NLP research on toxic language. We demonstrate that aporophobic attitudes are indeed present in social media and argue that the existing NLP datasets and models are inadequate to effectively address this problem.
    Georgina Curto, Svetlana Kiritchenko, Kathleen C. Fraser, and Isar Nejadgholi. (2024) The Crime of Being Poor: Associations between Crime and Poverty on Social Media in Eight Countries. In Proceedings of the Sixth Workshop on NLP and Computational Social Science (NLP+CSS), Mexico City, Mexico, June 2024. [pdf]
    Svetlana Kiritchenko, Georgina Curto, Isar Nejadgholi, and Kathleen C. Fraser. (2023) Aporophobia: An Overlooked Type of Toxic Language Targeting the Poor. In Proceedings of the 7th Workshop on Online Abuse and Harms (WOAH), Toronto, ON, Canada, July 2023. Outstanding Paper Award [pdf]
• Examining the State of Being Alone (SOLO): we created a corpus of over 4 million tweets collected with query terms solitude, lonely, and loneliness, and used it to analyze the language and emotions associated with the state of being alone.
• WikiArt Emotions Dataset: a dataset of 4,105 pieces of art (mostly paintings) that has annotations for emotions evoked in the observer. The pieces of art were selected from WikiArt.org's collection for twenty-two categories from four western styles (Renaissance Art, Post-Renaissance Art, Modern Art, and Contemporary Art). In addition to emotions, the art was also annotated for whether it includes the depiction of a face and how much the observers like the art.
• Affect in Tweets
  
  
  • SemEval-2018 Shared Task on Affect in Tweets: ~22,000 tweets in English, Arabic, and Spanish annotated for emotion and sentiment. The data is annotated for coarse classes (i.e., no anger, low anger, moderate anger, high anger) as well as for fine-grained real-valued scores indicating the intensity of emotion. The task included five subtasks: 1. emotion intensity regression, 2. emotion intensity ordinal classification, 3. valence (sentiment) regression, 4. valence ordinal classification, and 5. emotion classification. Seventy-five teams (about 200 team members) participated in the shared task.
    Mohammad, S., Bravo-Marquez, F., Salameh, M., and Kiritchenko, S. (2018). Semeval-2018 Task 1: Affect in tweets. In Proceedings of the International Workshop on Semantic Evaluation (SemEval-2018), New Orleans, LA, USA, June 2018 [pdf]
  • Relationships between Affect Categories: the tweets were annotated for many emotion (or affect) dimensions, from both the basic emotion model and the VAD model. We used these data to analyze the relationships between affect categories. We calculated the extent to which pairs of emotions co-occur in tweets. We showed the extent to which the intensities of affect dimensions correlate. We also calculated affect-affect intensity ratios which help identify the tweets for which the two affect scores correlate and the tweets for which they do not.
    Mohammad, S. and Kiritchenko, S. (2018). Understanding Emotions: A Dataset of Tweets to Study Interactions between Affect Categories. In Proceedings of the 11th edition of the Language Resources and Evaluation Conference, May 2018, Miyazaki, Japan [pdf]
• Stance
  • Quantifying Qualitative Data for Understanding Controversial Issues
  • SemEval-2016 Shared Task on Stance Detection in Tweets: 4870 English tweets annotated for stance towards six commonly known targets in the United States ('Atheism', 'Climate Change is a Real Concern', 'Feminist Movement', 'Hillary Clinton', 'Legalization of Abortion', and 'Donald Trump'). The annotations were done manually via crowdsourcing.
    Mohammad, S., Kiritchenko, S., Sobhani, P., Zhu, X., and Cherry, C. (2016) SemEval-2016 Task 6: Detecting Stance in Tweets. Proceedings of the International Workshop on Semantic Evaluation (SemEval-2016), San Diego, California, 2016 [pdf]
    Mohammad, S., Kiritchenko, S., Sobhani, P., Zhu, X., and Cherry, C. (2016) A Dataset for Detecting Stance in Tweets. Proceedings of the 10th edition of the Language Resources and Evaluation Conference (LREC), Portoroz, Slovenia, 2016 [pdf] [interactive visualization]
  • Automatic Stance Detection System: we designed and implemented a simple, but effective stance detection system that obtained an F-score higher than the one obtained by the more complex, best-performing system in the SemEval-2016 competition. We used a linear-kernel SVM classifier that leveraged word and character n-grams as well as word-embedding features drawn from additional unlabeled data.
    Mohammad, S., Sobhani, P., Kiritchenko, S. (2017). Stance and Sentiment in Tweets. ACM Transactions on Internet Technology, 17(3), 2017 [pdf]
  • Stance vs. Sentiment: the SemEval-2016 stance data were also annotated for 'target of opinion' and sentiment. We conducted a detailed analysis of the dataset and performed several experiments to tease out the interactions between stance and sentiment. We showed that sentiment features are not as effective for stance detection as they are for sentiment prediction. Moreover, an oracle system that had access to gold sentiment and target of opinion annotations was able to predict stance with an F-score of only 59.6%, 10% lower than the F-score obtained by our automatic stance detection system. We also showed that even though humans are capable of detecting stance towards a given target from texts that express opinion towards a different target, automatic systems perform poorly on such data.
    Mohammad, S., Sobhani, P., Kiritchenko, S. (2017). Stance and Sentiment in Tweets. ACM Transactions on Internet Technology, 17(3), 2017 [pdf] [interactive visualization]
• Best-Worst Scaling
  
  • Sentiment composition lexicons created with Best-Worst Scaling
• Sentiment Analysis in Arabic
• Personality Detection from Text
  • Relationship between Emotions and Personality: we performed experiments to show that fine-grained emotion categories such as that of excitement, guilt, yearning, and admiration are useful in automatically detecting personality from text. We generated a large lexicon of word-emotion associations from tweets with emotion-word hashtags pertaining to 585 fine-grained emotions. The fine-grained emotion category features extracted form this lexicon significantly improved performance of all classifiers over the majority baseline, and matched or outperformed a known set of baseline features -- Mairesse et al. (2007). The improvements were in large majority of cases above and beyond those obtained using features from coarse affect categories and word information content. Stream-of-consciousness essays and collections of Facebook posts marked with personality traits of the author were used as the test sets.
    Mohammad, S., Kiritchenko S. (2015) Using Hashtags to Capture Fine Emotion Categories from Tweets. Computational Intelligence, 31(2): 301-326, 2015 [pdf]
• Automatic Sentiment Analysis Systems
  • Aspect Based Sentiment Analysis
    
    We participated in the SemEval-2014 Aspect Based Sentiment Analysis shared task (Task 4):
    - detecting aspect categories (
    1st place
    )
    - detecting sentiment towards aspect categories (
    1st place
    )
    - detecting aspect terms (
    3rd place
    )
    - detecting sentiment towards aspect terms (
    1st and 2nd places
    )
    Kiritchenko, S., Zhu, X., Cherry, C., and Mohammad, S. (2014) NRC-Canada-2014: Detecting Aspects and Sentiment in Customer Reviews. Proceedings of the 8th International Workshop on Semantic Evaluation Exercises (SemEval-2014), Dublin, Ireland, 2014 [pdf]
  • Sentiment Analysis of Tweets
    
    We participated in the SemEval-2014 Sentiment Analysis in Twitter shared task (Task 9):
    - term-level sentiment classification (
    overall 1st place
    )
    - message-level sentiment classification (
    overall 1st place
    )
    Kiritchenko, S., Zhu, X., Mohammad, S. (2014). Sentiment Analysis of Short Informal Texts. Journal of Artificial Intelligence Research, 50:723-762, 2014 [pdf]
    Zhu, X., Kiritchenko, S., and Mohammad, S. (2014) NRC-Canada-2014: Recent Improvements in the Sentiment Analysis of Tweets. Proceedings of the 8th International Workshop on Semantic Evaluation Exercises (SemEval-2014), Dublin, Ireland, 2014 [pdf]
    
    
    We participated in the SemEval-2013 Sentiment Analysis in Twitter shared task (Task 2):
    - term-level sentiment classification (
    1st place
    )
    - message-level sentiment classification (
    1st place
    )
    Mohammad, S., Kiritchenko, S., and Zhu, X. (2013) NRC-Canada: Building the State-of-the-Art in Sentiment Analysis of Tweets. Proceedings of the 7th International Workshop on Semantic Evaluation Exercises (SemEval-2013), 2013 [pdf]
• Large-Scale Automatically Generated Sentiment Lexicons

Medical Informatics

• Social Media Mining for Health-Related Applications
  • AMIA-2017 Shared Task on Social Media Mining for Health Applications
    
    We participated in the Second Social Media Mining for Health Applications (SMM4H) Shared Task at AMIA-2017:
    - classification of tweets mentioning adverse drug reactions (
    1st place
    )
    - classification of tweets describing personal medication intake (
    3rd place
    )
    Kiritchenko, S., Mohammad, S., Morin, J., and de Bruijn, B. (2017). NRC-Canada at SMM4H Shared Task: Classifying Tweets Mentioning Adverse Drug Reactions and Medication Intake. Proceedings of the Social Media Mining for Health Applications Workshop at AMIA-2017, Washington, DC, USA, 2017 [pdf]
• Clinical Records Mining
  • Clinical Document Coding: though it can be viewed as multi-label document classification, the coding problem has the interesting property that most code assignments can be supported by a single phrase found in the input document. We propose a Lexically-Triggered Hidden Markov Model (LT-HMM) that leverages these phrases to improve coding accuracy. The LT-HMM works in two stages: first, a lexical match is performed against a term dictionary to collect a set of candidate codes for a document. Next, a discriminative HMM selects the best subset of codes to assign to the document by tagging candidates as present or absent. By confirming codes proposed by a dictionary, the LT-HMM can share features across codes, enabling strong performance even on rare codes. In fact, we are able to recover codes that do not occur in the training set at all.
    Kiritchenko, S., and Cherry, C. (2011) Lexically-Triggered Hidden Markov Models for Clinical Document Coding. Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics (ACL), 2011 [pdf]
  • i2b2-2010 Challenge in Natural Language Processing for Clinical Data
    
    We participated in the 4th i2b2/VA Challenge in Natural Language Processing for Clinical Data:
    - extraction of medical problems, tests, and treatments (
    1st place
    )
    - classification of assertions made on medical problems (
    1st place
    )
    - classification of relations between medical problems, tests, and treatments (
    2nd place
    )
    de Bruijn, B., Cherry, C., Kiritchenko, S., Martin, J., Zhu, X. (2011) Machine-Learned Solutions for Three Stages of Clinical Information Extraction: the State of the Art at i2b2 2010. Journal of the American Medical Informatics Association (JAMIA), 18:557-562, 2011 [pdf]
    Zhu, X., Cherry, C., Kiritchenko, S., Martin, J., de Bruijn, B. (2013) Detecting Concept Relations in Clinical Text: Insights from a State-of-The-Art Model. Journal of Biomedical Informatics, 46(2):275-85, 2013 [pdf]
• Biomedical Information Extraction from Journal Articles
  • ExaCT: an automatic information extraction system that assists users with locating and extracting key trial characteristics (e.g., eligibility criteria, sample size, drug dosage, primary outcomes) from full-text journal articles reporting on randomized controlled trials (RCTs). ExaCT consists of two parts: an information extraction (IE) engine that searches the article for text fragments that best describe the trial characteristics, and a web browser-based user interface that allows human reviewers to assess and modify the suggested selections. A demo version of ExaCT is available here.
    Kiritchenko, S., de Bruijn, B., Carini, S., Martin, J., and Sim, I. (2010) ExaCT: Automatic Extraction of Clinical Trial Characteristics from Journal Publications. BMC Medical Informatics and Decision Making 2010, 10:56 [pdf][video]
  • Application of Hierarchical Classification to Bioinformatics:
    - We index biomedical articles with Medical Subject Headings (MeSH). MeSH is a manually built controlled vocabulary for the biomedical domain where terms are arranged into several hierarchical structures called MeSH Trees. We address this task as a hierarchical text categorization task with class hierarchies derived from the MeSH vocabulary. In other words, we classify biomedical articles from the Medline library into one or several keywords from a specified MeSH Tree.
    - We apply the learning and evaluation techniques designed for hierarchical text categorization to the task of functional annotation of genes from biomedical literature. The purpose of this task is to retrieve the known functionality of a group of genes from the literature and translate it into a controlled vocabulary of the Gene Ontology (GO). For this, we classify biomedical articles describing the functionality of a given gene into one or several functional classes from GO.
    Kiritchenko, S., Matwin, S. and Famili, F. (2005) Functional Annotation of Genes Using Hierarchical Text Categorization. Proceedings of the BioLINK SIG: Linking Literature, Information and Knowledge for Biology, a joint meeting of the ISMB BioLINK Special Interest Group on Text Data Mining and the ACL Workshop on Linking Biological Literature, Ontologies and Databases: Mining Biological Semantics (held at ISMB-05), 2005 [pdf]
    Kiritchenko, S., Matwin, S., and Famili, F. (2004) Hierarchical Text Categorization as a Tool of Associating Genes with Gene Ontology Codes. Proceedings of the Second European Workshop on Data Mining and Text Mining for Bioinformatics (held at ECML-04), pp. 26-30, 2004 [pdf]
    Kiritchenko, S. (2006) Hierarchical Text Categorization: Algorithms, Evaluation, and Applications. Ph.D. Thesis, University of Ottawa, 2006 [pdf]

Hierarchical Classification

Semi-Supervised Learning

• Email Classification with Co-Training: the main problems in text classification are lack of labeled data, as well as the cost of labeling the unlabeled data. We address these problems by exploring co-training - an algorithm that uses unlabeled data along with a few labeled examples to boost the performance of a classification system. We experimented with co-training on the email domain. Our results showed that the performance of co-training depends on the learning algorithm it uses. In particular, Support Vector Machines significantly outperformed Naive Bayes on email data.
  Kiritchenko, S., and Matwin, S. (2001) Email Classification with Co-Training. Proceedings of CASCON 2001, pages 192-201, Toronto, Canada, 2001. CASCON-2011 Award: Most Influential Paper from a decade ago [pdf]