Subjective NLP tasks and complex human values

The flexibility of input structure of text-to-text pretrained models have enabled few-shots adaptions to new tasks using prompts. I have recently started exploring how the same flexibility can be used to improve the model’s performance in subjective NLP tasks like emotion classification, hate speech detection, stance detection, or elaboration of concepts in a domain-specific text. In my latest AAAI 2023 submission, I used this flexibility to enable users to selectively simplify the contents of medical texts.

Quality labeling has always been a challenge for subjective NLP tasks due to the lack of a well-defined standard. Our recent project on mining human values from online reviews, submitted to CHI 2023 exposed us to some of those challenges. We found that low inter-annotator agreement aligns directly with the low performance of large language models like RoBERTa. One of the interesting reasons for subjectivity in data labels turned out to be the ranking of labels (words) in English vocabulary, especially when one label was a hyperonym of another label or belonged to the synset of another label in the Wordnet sense or when one label had strong correlation (co-occurrence in multi-label setting) with another label.

Controllable language generation

Current language models are giants and have learned indiscriminately from the internet. Much research has been devoted to what these models are learning and how to make them unlearn concepts or connections that we do not want them to learn. There is also a great scope of collaboration between humans with our natural communication skills and language models with their vast extra-human knowledge. I am excited to test out how we can use the recent developments in controllable text generation to align natural language models to human values, values that our ideal selves care about.

Links

I have been meaning to share the links to some of the pages that I follow for learning Neural Networks, Deep Learning, Motion planning algorithms, Q-Learning using function approximation and Deep Reinforcement Learning. Most of these pages are good if you want to code from scratch. 

This is an online book on Neural Networks and Deep Learning that I am following right now. The codes are all in Python. The exercises are thoughtful and the explanation is quite easy to follow. Here's the link: Michael Nielsen's online book on Neural Network and Deep Learning

A great site for graph search algorithms like BFS, BFS, Dijkstra's and A* and its variants (gaming applications): Amit's A* Pages

David Silver's course on Reinforcement Learning

News

October 19 - 21, 2016 : I just attended Grace Hopper Celebration in Houston. I have heard from many former attendees that this conference boosts your spirits and resets your goals if you are distracted. I agree. It is a "celebration" in its true sense. You do feel celebrated as a woman, starting from the keynotes by some highly successful women in tech to meeting representatives from different companies. But what touched me the most is to meet a "galaxy" of women in computing who come from a plethora of backgrounds, like some of the most accomplished graduate students to mothers who decided to pursue computer science education late in their life. There were 16000 of us. No matter how much of an outlier that you think you are, you are sure to find someone who had very similar life experiences as you. I was thrilled to meet an engineer from Google who started out her career in arts and architecture, very much like I did. As for the technical talks, I loved the workshop on design of self-driving cars, offered by four Google engineers. They demonstrated how a mechanical engineer, a systems engineer, a user experience product manager and a software engineer would collaboratively solve a technical challenge in a corporate environment. Overall GHC exceeded my expectations.