The Fifth Elephant 2020 edition

Natural Language Processing models are susceptible to learning spurious and shallow patterns in the dataset which does not generalize well to real world data. Given that dataset serves as the proxy for the actual task on hand, often deep learning NLP models learn from the spurious shallow patterns in the dataset instead of solving the actual task on hand. The presence of such non-robust brittle features has been shown to lead to poor real world generalization performance in various tasks such as sentiment analysis, question answering, natural language inference etc. Hence performance on a handful of test datasets does not often translate to real world. It is essential to analyze the data and identify whether the model is depending on such shallow and spurious patterns. As a second step it is essential to mitigate such impact on the learnt model to improve the real world performance. In this talk, we discuss the current state of art methods in identifying such shallow surface cues in NLP datasets and cover a range of techniques to mitigate such cues and to build models which don’t depend on them.

Outline

Part I - Is your NLP model solving the dataset instead of the actual task?
We cover existing research literature on identification of shallow surface cues/patterns in datasets and thus motivate the need for building models which don’t depend on such surface cues. (We also briefly cover the work on how adversarial examples can be shown to arise from the model dependence on such shallow cues).
Part II – Identifying, Analyzing and Mitigating spurious dataset cues from NLP models.
We then cover techniques which can identify model dependence on such spurious cues, and discuss mitigating and eliminating techniques. We consider two real world NLP tasks namely natural language inference and question answering and show how these techniques are applicable in eliminating such shallow surface cues from the model learning.
This talk is focussed on intermediate and advanced NLP developers/practitioners who are interested in building robust NLP models. Prior knowledge of basic NLP is assumed.

References
(1) https://arxiv.org/abs/2002.04108
(2) https://arxiv.org/abs/1908.10763
(3) https://arxiv.org/abs/1909.03683
(4) https://arxiv.org/abs/1911.03861
(5) https://arxiv.org/abs/2001.01565
(6) https://arxiv.org/abs/1905.02175

Requirements

None

Speaker bio

Sandya Mannarswamy is an independent researcher in Natural Language Processing. She was a senior research scientist at Conduent Labs India in the Natural Language Processing research group. She holds a Ph.D. in computer science from Indian Institute of Science, Bangalore. She has 19 years of software industry experience, at various R&D labs, including Hewlett Packard Ltd, IBM and Xerox Research. Her work spans a number of areas natural language processing, machine learning, compiler optimizations, developer tools and file systems, with a number of publications and patents.

Links

• https://scholar.google.co.in/citations?user=i27nd3oAAAAJ&hl=en
• https://www.linkedin.com/in/sandya/
• https://hasgeek.com/anthillinside/2019/schedule/rigorous-evaluation-of-nlp-models-for-real-world-deployment-eZZTRPfVp6dUM4q32ZPAKL