Seminars

View all Seminars  |  Download ICal for this event

Fragile Interpretations and Interpretable Models for NLP

Series: M.Tech (Research) Thesis Defense

Speaker: Ms. Ayushi Kohli

Date/Time: Oct 06 11:30:00

Location: CSA Seminar Hall (Room No. 254, First Floor)

Faculty Advisor: Dr. Susheela Devi

Abstract:
Deployment of deep learning models in critical areas is still an issue of concern,
as the cost of making a wrong decision is very high in these areas. As a result,
the final decision in these settings is human-centric. Also, these models act as
black boxes, and we are unaware of their internal workings. Therefore, the models
must be explainable to know their internal workings. So, if the model is explainable,
is it safe to deploy it in real-world settings involving huge risks?

Our work centres around the concept of fragile interpretations considering the models
robustness and the robustness of interpretations. We have proposed an algorithm that
perturbs the input text and generates adversarial examples with the same prediction
as the input but with different interpretations. Through our experiments, we have
provided a detailed analysis of whether these interpretations are reliable and whether
to trust the model or the interpretations. We have provided the reason for the fragility
in the case of NLP. Taking this into account, we have proposed two interpretable models,
one for a multi-task offensive language detection task and the other for a sentence-pair
similarity detection task.