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 Master Thesis is an exploratory research to determine whether it is feasible to construct a subjectivity lexicon using Wikipedia. The key hypothesis is that that all quotes in Wikipedia are subjective and all regular text are objective. The degree of subjectivity of a word, also known as ''Quote Score'' is determined based on the ratio of word frequency in quotations to its frequency outside quotations. The proportion of words in the English Wikipedia which are within quotations is found to be much smaller as compared to those which are not in quotes, resulting in a right-skewed distribution and low mean value of Quote Scores.
The methodology used to generate the subjectivity lexicon from text corpus in English Wikipedia is designed in such a way that it can be scaled and reused to produce similar subjectivity lexica of other languages. This is achieved by abstaining from domain and language-specific methods, apart from using only readily-available English dictionary packages to detect and exclude stopwords and non-English words in the Wikipedia text corpus.
The subjectivity lexicon generated from English Wikipedia is compared against other lexica; namely MPQA and SentiWordNet. It is found that words which are strongly subjective tend to have high Quote Scores in the subjectivity lexicon generated from English Wikipedia. There is a large observable difference between distribution of Quote Scores for words classified as strongly subjective versus distribution of Quote Scores for words classified as weakly subjective and objective. However, weakly subjective and objective words cannot be differentiated clearly based on Quote Score. In addition to that, a questionnaire is commissioned as an exploratory approach to investigate whether subjectivity lexicon generated from Wikipedia could be used to extend the coverage of words of existing lexica.
The content aggregator platform Reddit has established itself as one of the most popular websites in the world. However, scientific research on Reddit is hindered as Reddit allows (and even encourages) user anonymity, i.e., user profiles do not contain personal information such as the gender. Inferring the gender of users in large-scale could enable the analysis of gender-specific areas of interest, reactions to events, and behavioral patterns. In this direction, this thesis suggests a machine learning approach of estimating the gender of Reddit users. By exploiting specific conventions in parts of the website, we obtain a ground truth for more than 190 million comments of labeled users. This data is then used to train machine learning classifiers to use them to gain insights about the gender balance of particular subreddits and the platform in general. By comparing a variety of different approaches for classification algorithm, we find that character-level convolutional neural network achieves performance with an 82.3% F1 score on a task of predicting a gender of a user based on his/her comments. The score surpasses 85% mark for frequent users with more than 50 comments. Furthermore, we discover that female users are less active on Reddit platform, they write fewer comments and post in fewer subreddits on average, when compared to male users.
Current political issues are often reflected in social media discussions, gathering politicians and voters on common platforms. As these can affect the public perception of politics, the inner dynamics and backgrounds of such debates are of great scientific interest. This thesis takes user generated messages from an up-to-date dataset of considerable relevance as Time Series, and applies a topic-based analysis of inspiration and agenda setting to it. The Institute for Web Science and Technologies of the University Koblenz-Landau has collected Twitter data generated beforehand by candidates of the European Parliament Election 2019. This work processes and analyzes the dataset for various properties, while focusing on the influence of politicians and media on online debates. An algorithm to cluster tweets into topical threads is introduced. Subsequently, Sequential Association Rules are mined, yielding wide array of potential influence relations between both actors and topics. The elaborated methodology can be configured with different parameters and is extensible in functionality and scope of application.
The output of eye tracking Web usability studies can be visualized to the analysts as screenshots of the Web pages with their gaze data. However, the screenshot visualizations are found to be corrupted whenever there are recorded fixations on fixed Web page elements on different scroll positions. The gaze data are not gathered on their fixated fixed elements; rather they are scattered on their recorded scroll positions. This problem has raised our attention to find an approach to link gaze data to their intended fixed elements and gather them in one position on the screenshot. The approach builds upon the concept of creating the screenshot during the recording session, where images of the viewport are captured on visited scroll positions and lastly stitched into one Web page screenshot. Additionally, the fixed elements in the Web page are identified and linked to their fixations. For the evaluation, we compared the interpretation of our enhanced screenshot against the video visualization, which overcomes the problem. The results revealed that both visualizations equally deliver accurate interpretations. However, interpreting the visualizations of eye tracking Web usability studies using the enhanced screenshots outperforms the video visualizations in terms of speed and it requires less temporal demands from the interpreters.
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.
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.
“Did I say something wrong?” A word-level analysis of Wikipedia articles for deletion discussions
(2016)
This thesis focuses on gaining linguistic insights into textual discussions on a word level. It was of special interest to distinguish messages that constructively contribute to a discussion from those that are detrimental to them. Thereby, we wanted to determine whether “I”- and “You”-messages are indicators for either of the two discussion styles. These messages are nowadays often used in guidelines for successful communication. Although their effects have been successfully evaluated multiple times, a large-scale analysis has never been conducted. Thus, we used Wikipedia Articles for Deletion (short: AfD) discussions together with the records of blocked users and developed a fully automated creation of an annotated data set. In this data set, messages were labelled either constructive or disruptive. We applied binary classifiers to the data to determine characteristic words for both discussion styles. Thereby, we also investigated whether function words like pronouns and conjunctions play an important role in distinguishing the two. We found that “You”-messages were a strong indicator for disruptive messages which matches their attributed effects on communication. However, we found “I”-messages to be indicative for disruptive messages as well which is contrary to their attributed effects. The importance of function words could neither be confirmed nor refuted. Other characteristic words for either communication style were not found. Yet, the results suggest that a different model might represent disruptive and constructive messages in textual discussions better.
Topic models are a popular tool to extract concepts of large text corpora. These text corpora tend to contain hidden meta groups. The size relation of these groups is frequently imbalanced. Their presence is often ignored when applying a topic model. Therefore, this thesis explores the influence of such imbalanced corpora on topic models.
The influence is tested by training LDA on samples with varying size relations. The samples are generated from data sets containing a large group differences i.e language difference and small group differences i.e. political orientation. The predictive performance on those imbalanced corpora is judged using perplexity.
The experiments show that the presence of groups in training corpora can influence the prediction performance of LDA. The impact varies due to various factors, including language-specific perplexity scores. The group-related prediction performance changes for groups when varying the relative group sizes. The actual change varies between data sets.
LDA is able to distinguish between different latent groups in document corpora if differences between groups are large enough, e.g. for groups with different languages. The proportion of group-specific topics is under-proportional to the share of the group in the corpus and relatively smaller for minorities.
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.