Master's Thesis
Refine
Document Type
- Master's Thesis (19) (remove)
Keywords
Institute
- Institute for Web Science and Technologies (19) (remove)
The purpose of this thesis is to explore the sentiment distributions of Wikipedia concepts.
We analyse the sentiment of the entire English Wikipedia corpus, which includes 5,669,867 articles and 1,906,375 talks, by using a lexicon-based method with four different lexicons.
Also, we explore the sentiment distributions from a time perspective using the sentiment scores obtained from our selected corpus. The results obtained have been compared not only between articles and talks but also among four lexicons: OL, MPQA, LIWC, and ANEW.
Our findings show that among the four lexicons, MPQA has the highest sensitivity and ANEW has the lowest sensitivity to emotional expressions. Wikipedia articles show more sentiments than talks according to OL, MPQA, and LIWC, whereas Wikipedia talks show more sentiments than articles according to ANEW. Besides, the sentiment has a trend regarding time series, and each lexicon has its own bias regarding text describing different things.
Moreover, our research provides three interactive widgets for visualising sentiment distributions for Wikipedia concepts regarding the time and geolocation attributes of concepts.
This thesis focuses on approximate inference in assumption-based argumentation frameworks. Argumentation provides a significant idea in the computerization of theoretical and practical reasoning in AI. And it has a close connection with AI, engaging in arguments to perform scientific reasoning. The fundamental approach in this field is abstract argumentation frameworks developed by Dung. Assumption-based argumentation can be regarded as an instance of abstract argumentation with structured arguments. When facing a large scale of data, a challenge of reasoning in assumption-based argumentation is how to construct arguments and resolve attacks over a given claim with minimal cost of computation and acceptable accuracy at the same time. This thesis proposes and investigates approximate methods that randomly select and construct samples of frameworks based on graphical dispute derivations to solve this problem. The presented approach aims to improve reasoning performance and get an acceptable trade-off between computational time and accuracy. The evaluation shows that for reasoning in assumption-based argumentation, in general, the running time is reduced with the cost of slightly low accuracy by randomly sampling and constructing inference rules for potential arguments over a query.
Commonsense reasoning can be seen as a process of identifying dependencies amongst events and actions. Understanding the circumstances surrounding these events requires background knowledge with sufficient breadth to cover a wide variety of domains. In the recent decades, there has been a lot of work in extracting commonsense knowledge, a number of these projects provide their collected data as semantic networks such as ConceptNet and CausalNet. In this thesis, we attempt to undertake the Choice Of Plausible Alternatives (COPA) challenge, a problem set with 1000 questions written in multiple-choice format with a premise and two alternative choices for each question. Our approach differs from previous work by using shortest paths between concepts in a causal graph with the edge weight as causality metric. We use CausalNet as primary network and implement a few design choices to explore the strengths and drawbacks of this approach, and propose an extension using ConceptNet by leveraging its commonsense knowledge base.
This thesis explores and examines the effectiveness and efficacy of traditional machine learning (ML), advanced neural networks (NN) and state-of-the-art deep learning (DL) models for identifying mental distress indicators from the social media discourses based on Reddit and Twitter as they are immensely used by teenagers. Different NLP vectorization techniques like TF-IDF, Word2Vec, GloVe, and BERT embeddings are employed with ML models such as Decision Tree (DT), Random Forest (RF), Logistic Regression (LR) and Support Vector Machine (SVM) followed by NN models such as Convolutional Neural Network (CNN), Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) to methodically analyse their impact as feature representation of models. DL models such as BERT, DistilBERT, MentalRoBERTa and MentalBERT are end-to-end fine tuned for classification task. This thesis also compares different text preprocessing techniques such as tokenization, stopword removal and lemmatization to assess their impact on model performance. Systematic experiments with different configuration of vectorization and preprocessing techniques in accordance with different model types and categories have been implemented to find the most effective configurations and to gauge the strengths, limitations, and capability to detect and interpret the mental distress indicators from the text. The results analysis reveals that MentalBERT DL model significantly outperformed all other model types and categories due to its specific pretraining on mental data as well as rigorous end-to-end fine tuning gave it an edge for detecting nuanced linguistic mental distress indicators from the complex contextual textual corpus. This insights from the results acknowledges the ML and NLP technologies high potential for developing complex AI systems for its intervention in the domain of mental health analysis. This thesis lays the foundation and directs the future work demonstrating the need for collaborative approach of different domain experts as well as to explore next generational large language models to develop robust and clinically approved mental health AI systems.
We examine the systematic underrecognition of female scientists (Matilda effect) by exploring the citation network of papers published in the American Physical Society (APS) journals. Our analysis shows that articles written by men (first author, last author and dominant gender of authors) receive more citations than similar articles written by women (first author, last author and dominant gender of authors) after controlling for the journal of publication, year of publication and content of the publication. Statistical significance of the overlap between the lists of references was considered as the measure of similarity between articles in our analysis. In addition, we found that men are less likely to cite articles written by women and women are less likely to cite articles written by men. This pattern leads to receiving more citations by articles written by men than similar articles written by women because the majority of authors who published in APS journals are male (85%). We also observed Matilda effect reduces when articles are published in journals with the highest impact factors. In other words, people’s evaluation of articles published in these journals is not affected by the gender of authors significantly. Finally, we suggested a method that can be applied by editors in academic journals to reduce the evaluation bias to some extent. Editors can identify missing citations using our proposed method to complete bibliographies. This policy can reduce the evaluation bias because we observed papers written by female scholars (first author, last author, the dominant gender of authors) miss more citations than articles written by male scholars (first author, last author, the dominant gender of authors).
Most social media platforms allow users to freely express their opinions, feelings, and beliefs. However, in recent years the growing propagation of hate speech, offensive language, racism and sexism on the social media outlets have drawn attention from individuals, companies, and researchers. Today, sexism both online and offline with different forms, including blatant, covert, and subtle lan- guage, is a common phenomenon in society. A notable amount of work has been done over identifying sexist content and computationally detecting sexism which exists online. Although previous efforts have mostly used peoples’ activities on social media platforms such as Twitter as a public and helpful source for collecting data, they neglect the fact that the method of gathering sexist tweets could be biased towards the initial search terms. Moreover, some forms of sexism could be missed since some tweets which contain offensive language could be misclassified as hate speech. Further, in existing hate speech corpora, sexist tweets mostly express hostile sexism, and to some degree, the other forms of sexism which also appear online was disregarded. Besides, the creation of labeled datasets with manual exertion, relying on users to report offensive comments with a tremendous effort by human annotators is not only a costly and time-consuming process, but it also raises the risk of involving discrimination under biased judgment.
This thesis generates a novel sexist and non-sexist dataset which is constructed via "UnSexistifyIt", an online web-based game that incentivizes the players to make minimal modifications to a sexist statement with the goal of turning it into a non-sexist statement and convincing other players that the modified statement is non-sexist. The game applies the methodology of "Game With A Purpose" to generate data as a side-effect of playing the game and also employs the gamification and crowdsourcing techniques to enhance non-game contexts. When voluntary participants play the game, they help to produce non-sexist statements which can reduce the cost of generating new corpus. This work explores how diverse individual beliefs concerning sexism are. Further, the result of this work highlights the impact of various linguistic features and content attributes regarding sexist language detection. Finally, this thesis could help to expand our understanding regarding the syntactic and semantic structure of sexist and non-sexist content and also provides insights to build a probabilistic classifier for single sentences into sexist or non-sexist classes and lastly find a potential ground truth for such a classifier.
Ontologies are valuable tools for knowledge representation and important building blocks of the Semantic Web. They are not static and can change over time. Changing an ontology can be necessary for various reasons: the domain that is represented by an ontology can change or an ontology is reused and must be adapted to the new context. In addition, modeling errors could have been introduced into the ontology which must be found and removed. The non-triviality of the change process has led to the emerge of ontology change as an own field of research. The removal of knowledge from ontologies is an important aspect of this change process, because even the addition of new knowledge to an ontology potentially requires the removal of older, conflicting knowledge. Such a removal must be performed in a thought-out way. A naïve change of concepts within the ontology can easily remove other, unrelated knowledge or alter the semantics of concepts in an unintended way [2]. For these reasons, this thesis introduces a formal operator for the fine-grained retraction of knowledge from EL concepts which is partially based on the postulates for belief set contraction and belief base contraction [3, 4, 5] and the work of Suchanek et al. [6]. For this, a short introduction to ontologies and OWL 2 is given and the problem of ontology change is explained. It is then argued why a formal operator can support this process and why the Description Logic EL provides a good starting point for the development of such an operator. After this, a general introduction to Description Logic is given. This includes its history, an overview of its applications and common reasoning tasks in this logic. Following this, the logic EL is defined. In a next step, related work is examined and it is shown why the recovery postulate and the relevance postulate cannot be naïvely employed in the development of an operator that removes knowledge from EL concepts. Following this, the requirements to the operator are formulated and properties are given which are mainly based on the postulates for belief set and belief base contraction. Additional properties are developed which make up for the non-applicability of the recovery and relevance postulates. After this, a formal definition of the operator is given and it is shown that the operator is applicable to the task of a fine-grained removal of knowledge from EL concepts. In a next step, it is proven that the operator fulfills all the previously defined properties. It is then demonstrated how the operator can be combined with laconic justifications [7] to assist a human ontology editor by automatically removing unwanted consequences from an ontology. Building on this, a plugin for the ontology editor Protégé is introduced that is based on algorithms that were derived from the formal definition of the operator. The content of this work is then summarized and a final conclusion is drawn. The thesis closes with an outlook into possible future work.
Our work finds the fine grained edits in context of neighbouring tokens in Wikipedia articles. We cluster those edits according to similar neighbouring context. We encode neighbouring context into vector space using word vectors. We evaluate clusters returned by our algorithm on extrinsic and intrinsic metric and compare it with previous work. We analyse the relation between extrinsic and intrinsic measurements of fine grained edit tokens.
Belief revision is the subarea of knowledge representation which studies the dynamics of epistemic states of an agent. In the classical AGM approach, contraction, as part of the belief revision, deals with the removal of beliefs in knowledge bases. This master's thesis presents the study and the implementation of concept contraction in the Description Logic EL. Concept contraction deals with the following situation. Given two concept C and D, assuming that C is subsumed by D, how can concept C be changed so that it is not subsumed by D anymore, but is as similar as possible to C? This approach of belief change is different from other related work because it deals with contraction in the level of concepts and not T-Boxes and A-Boxes in general. The main contribution of the thesis is the implementation of the concept contraction. The implementation provides insight into the complexity of contraction in EL, which is tractable since the main inference task in EL is also tractable. The implementation consists of the design of five algorithms that are necessary for concept contraction. The algorithms are described, illustrated with examples, and analyzed in terms of time complexity. Furthermore, we propose an new approach for a selection function, adapt for the concept contraction. The selection function uses metadata about the concepts in order to select the best from an input set. The metadata is modeled in a framework that we have designed, based on standard metadata frameworks. As an important part of the concept contraction, the selection function is responsible for selecting the best concepts that are as similar as possible to concept C. Lastly, we have successfully implemented the concept contraction in Python, and the results are promising.