Bachelor/Master Thesis

In research on human language processing, we are often interested in visual attention i.e. what objects or words study participants fixated. Traditionally, such research has used expensive eye-tracking equipment. However, today, many laptops have good cameras, which opens new perspectives for running experiments online using crowd-sourcing. In this thesis, you will compare data quality and experimental results using the traditional vs. crowd-sourced setup for a study on pragmatic processing. 

While researchers working on pragmatic communication games have typically focused only on the outcomes of these communication games, we’re now interested in the moment-to-moment behavior of participants in these games, specifically, how and when they examine various parts of the game’s visual display.

You will help us implement and analyze web-hosted webcam eye-tracking experiments to collect this data from a crowd-sourcing platform, and you will help explore the data to extract useful generalizations and decide on useful metrics for classifying trials and participants. Depending on your interests, the experiment can also be run in the eye-tracking lab for comparative results.

Natural language generation (NLG) holds immense potential to revolutionize information access and communication. However, a critical challenge lies in ensuring the reliability of generated text. Fabricating information, commonly referred to as "hallucination," undermines trust and limits the practical applications of NLG.

This research focuses on establishing robust mechanisms for reliable generation, particularly within data-to-text tasks. We utilize the drone technology domain as a compelling example, where accurate communication with human pilots is vital. Here, reliable NLG can generate informative messages about the drone's environment and mission details, enhancing situational awareness and decision-making.

You can build on our existing research, and contribute to (but not limited to): developing tailored evaluation metrics that assess both factual accuracy and domain-specific suitability, incorporating controlled generation techniques to explore different modeling techniques while improving reliability.

A popular paradigm to study language processing in the psycholinguistic field is self-paced reading: a task for which participants read a passage word-by-word or phrase-by-phrase, pressing a button to get the next word or phrase displayed. The time taken to press the button gives an indication of the processing difficulty at each stage.

To gain a deeper understanding of language processing, we need to know not only how people process written text, but also how they process spoken text. In this project, we will therefore see if results of previous reading research can be replicated in a spoken setting — using a novel self-paced listening paradigm: instead of reading the texts, people have to listen to the texts, and press a button to hear the next word or phrase.

You will build on existing research that has used self-paced reading, and create spoken versions of the materials used in these studies. The data will be collected via a crowdsourcing platform. You will analyze the data to gain insight into whether the effects found in the earlier research can be replicated in this new paradigm.

Discourse relations are logical relations between segments of texts that make the text coherent. They are often marked by discourse connectives. For example, the connective "because" marks a "reason" relation. Each language has its own collection of connectives and they often do not have mutual cross-lingual correspondences. For instance, the French connective "en effet" can be translated to "indeed" or "in fact" in English. When it is used to mark a "cause" relation, it is often omitted (implicitated) in the English translation (Zufferey 2016). Previous work used automatic word alignment to induce French connective lexicons and discourse annotation by projecting annotations from English (Laali and Kosseim 2014, Laali 2017). In this project, we would like to use this technique to study how discourse relations are marked in English-French translations.

References:
Laali, Majid, and Leila Kosseim. "Inducing discourse connectives from parallel texts." Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers. 2014.
Zufferey, Sandrine. "Discourse connectives across languages: Factors influencing their explicit or implicit translation." Languages in Contrast. International Journal for Contrastive Linguistics 16.2 (2016): 264-279.
Laali, Majid. Inducing discourse resources using annotation projection. Diss. Concordia University, 2017.