teaching assistant

I was a teaching assistant in the following courses and seminars at UniNe. iHover over a course to read a short summary. The course material is available either on our Ilias, Claroline, or Moodle collaborative learning platform. iStudents must first register to a course before being allowed to see its content. Furthermore, I (co-)advice(d) students for their master theses and supervised junior researchers.

courses

• Artificial Intelligence
  Spring 2018, Spring 2017, & Spring 2016
  Prof. Jacques Savoy
  Intelligent agent paradigm. Uninformed search (depth-first search, breadth-first search). Informed search (A*, heuristics). Local search (hill climbing, simulated annealing). Genetic programming. Constraint satisfaction problem. AND/OR tree. Games (minimax, alpha-beta algorithm). Knowledge representation (proposition, predicates, frames, semantic nets, rules). Expert-system (forward and backward chaining). Neural networks. Machine learning (Naive Bayes). Planning.
• Natural Language Processing
  Autumn 2017, Autumn 2016, & Spring 2015
  Prof. Jacques Savoy
  Python and XML. Linguistics (morphology, syntax, semantics). Simple statistical approaches (KWIC, concordances). Spelling detection and correction. Statistical models (counting words, bigrams, entropy). Parsing. Markov chains. Hidden Markov chains. Text categorization, sentiment analysis, authorship attribution. Cryptography. Question/Answering. Text summarization.
• Machine Learning and Data Mining
  Autumn 2016, Autumn 2015, & Autumn 2014
  Prof. Jacques Savoy
  Data Mining perspective. Association rules. Decision trees. Instance-based learning (nearest neighbors). Clustering. Support Vector machines (SVM). Evaluation (Train & Test, Cross-validation, Leaving-one-out). Data Streams.

seminars

• Seminar Natural Language Processing
  Autumn 2017, Autumn 2016, Autumn 2015, & Autumn 2014
  Prof. Jacques Savoy
  Conduct your own research (on a limited amount of data) based on one/few scientific papers. Implement correctly a given algorithm. Test it, understand its weakness, analyze its behavior, evaluate its performance (comparative basis, statistical tests), and discover the limit(s) of the proposed model.
• R&D Workshop
  Spring 2016 & Spring 2015
  Dr. Hugues Mercier, Prof. Peter Kropf, & Prof. Jacques Savoy
  Even though there is no unique correct way to conduct scientific research, the main objective is to present general methods and skills that will help the students do research efficiently. Work on a reading/writing research project. This may include reading scientific articles on a given topic, extending an existing algorithm, do a comprehensive survey of the state of the art techniques in a specific field. The students will write a report in the form of a scientific paper.

supervision of junior researchers

• Psychological Profiling of Twitter Users
  Master Thesis at the University of Neuchâtel, 2017.
  This thesis discusses the link between the psychological makeup of people and written text. More precisely, we examine the usefulness of specific word categories and other linguistic style elements commonly found in Twitter tweets to detect the five basic psychological traits known as the Big Five Personality Traits. The analysis is done using English and Spanish tweets. We also examine the usefulness of three different classification methods for this task; Naive Bayes, k-Nearest Neighbor as well as a classifier based on the Winnow algorithm.
• Author Verification over Topically Diverse Text Streams
  Master Thesis at the University of Neuchâtel, 2017.
  This thesis deals with the problem of author verification, which is one of the most challenging tasks in style-based text categorization. Author verification is the task of determining whether documents of unknown authorship are written by a certain author. The set of possible candidates is not always limited to a given finite closed set. In particular, we focus on two datasets, one related to political speeches and one related to newspaper articles. The main goal was to determine whether two unknown texts (possibly cross-genre) are written by the same author. Experiments with a variety of style-based features were tested in a supervised classification system using two distances, namely Delta and SPATIUM-L1.
• Limits of Deep Learning
  Research project in the Master program, 2016, two students.
  Current natural language models try to capture semantic and syntactic relationships between words. This allows them to answer questions in the form of analogies: "X is to Y as U is to ?". The performance of the word2vec tool was studied on this task. To this end, the sensitivity to different hyper-parameter choices was analyzed. The models (in)ability to learn and answer on domain-specific knowledge was demonstrated. To increase the models accuracy on solving analogies, a mean-shift procedure was proposed as a preliminary step to compute the query vector. The effectiveness of various combination of analogy interpretations for solving for the most similar word-vector are tested on a generalised question set.
• Automatic Gender Detection in Written Texts
  Research project in the Master program, 2015.
  Nowadays a lot of research is made in gender recognition for written texts. However, the algorithms are mostly complex and often exhaustive to the English language. The purpose of this research was to explain a simplified version of a gender-recognition algorithm and test it on two languages (English and Spanish) to determine if there are accuracy changes or if the algorithm performs equally on different languages. Three distinct variants of the algorithm were discussed in this work and compared in terms of accuracy and speed.