How I Used RoBERTa to Learn Common Sense Knowledge for Spatial Reasoning

Ellen Schellekens

Sep 13

-

5 min read

Humans have a natural profound understanding about the physical world we are born in. We know how objects move, and where they are located in relation to us. For example, if we know that the cloud is above the house, it's obvious that the house is below the cloud. This kind of common sense knowledge is learned implicitly and is difficult to capture by machine systems. This difficulty, combined with the importance of common sense knowledge to human intelligence, makes this a very fascinating research topic.Language models like BERT have enabled breakthroughs in many applications, such as dialogue systems and generative language modelling. These models are pre-trained on huge datasets, in an unsupervised manner. Would these models be able to reason about relative locations with only this initial pre-training? And are they able to learn this reasoning through further specific training? This is the topic of my Master's Thesis.To study this, I constructed a new dataset for relative locations. This new dataset is the first textual dataset that takes into consideration three dimensions, and of which the examples are based on real life situations. With this dataset, I evaluated the RoBERTa language model both after pre-training and with additional finetuning. I found that the model cannot reason about relative locations after only pre-training, but it is able to learn this reasoning during specific finetuning. Read all the details about my research in our Medium blog

Previous

Related projects

Related events

No events found.