Stanford CS224N NLP with Deep Learning | Winter 2021 | Lecture 12 - Question Answering | Summary and Q&A

TL;DR

Question Answering and Reading Comprehension models, such as LSTM-based models and BERT, have revolutionized the field, achieving high performance on standard datasets like SQuAD. However, adversarial examples and out-of-domain distributions still pose challenges.

Key Insights

• ⚾ QA/RC models have made significant advancements with the introduction of LSTM-based models and BERT.
• 🫠 Pre-training objectives play a crucial role in the performance of reading comprehension models, and improvements can be made to enhance their capabilities.

Transcript

okay hi everyone um welcome back to we're now in past the halfway point week six of cs 224 n um and so let me just give um a couple of quick announcements first um so today is the day that you have to have done the mid um quarter survey by hundreds of people have but if you haven't this is your last chance um to get the half point for that today is... Read More

Questions & Answers

Q: What are some challenges that QA/RC models face?

QA/RC models face challenges such as adversarial examples that can trick the models and out-of-domain distributions that are different from the training data, leading to decreased performance.

Q: What are some popular QA/RC datasets?

Some popular QA/RC datasets include SQuAD, Stanford Question Answering Dataset, which consists of annotated passages and question-answer pairs, and MS MARCO, which focuses on document ranking and passage ranking tasks.

Q: What are some key differences between LSTM-based models and BERT?

LSTM-based models rely on recurrent layers to capture sequential information, while BERT utilizes the transformer architecture, which is parallelizable and allows for more efficient training. BERT's pre-training on large amounts of text has been shown to be highly effective.

Q: How can pre-training objectives be improved for reading comprehension?

One approach is to consider contiguous spans as pre-training objectives, mimicking the target answer span in QA/RC tasks. Another approach is to focus on predicting the start and end positions of the answer span, compressing all the necessary information into these two points.

Summary & Key Takeaways

• Question Answering and Reading Comprehension (QA/RC) models have made significant progress in recent years, driven by deep learning techniques.

• Models like LSTM-based models and BERT have demonstrated high performance on standard datasets, such as SQuAD.

• However, these models still struggle with adversarial examples and out-of-domain distributions, which can lead to decreased performance.