AIP will host two talks at 12:30-14:00 on Friday, September 14 at AIP (COREDO Nihonbashi 15F). Everyone who received this email are welcome to attend.
1) Pontus Stenetrop (Senior Research Associate at University College London)
"Recent Trends in Natural Language Processing: Reading Comprehension and Fact Verification – with additional advice for young researchers”
2) Takuma Yoneda (Master Student at Toyota Institute - intern at Pontus’s group at UCL)
“Work of FEVER(Fact Extraction and VERification) at UCL and Internship experience)
1) Talk by Pontus Stenetrop
Natural Language Processing, along with all of Artificial Intelligence, has been receiving considerable attention over the last few years. In this talk, we want to highlight two very active areas of research: Reading Comprehension and Fact Verification. Lastly, we hope to give concrete advice based on our own experiences on how to navigate and make good research progress as a young researcher – we will also be available after the talk for personal questions.
Reading Comprehension is the task of answering questions for a given text. We will give a historical background of the task, the recently introduced datasets, and modelling approaches for solving the task. We will then relate this to recent research from ourselves and others on weaknesses of existing datasets and models, such as “easy samples”, inability to handle “multi-step” reasoning – predicting future lines of research.
“Fake News” has been the focus of a lot attention over the last few years, motivating research on automated fact verification. In this task, a system is given a claim and is tasked to not only answer if the claim is true or not, but also to provide the relevant evidence. Our group at University College London recently participated in the FEVER shared task, where we finished second among 24 teams. We will give an outline of the task and our own system for solving it, with a focus on the process of how one approaches successfully participating in a shared task where time is very limited compared to “standard” academic research.
2) Talk by Takuma Yoneda
During my intern at UCLMR, I worked on the shared task called FEVER(Fact Extraction and VERification), which is a new shared task showing up in the context of increasing need for automated fact-checking.
Given a claim, this task requires a model to answer whether the claim is true/false as well as providing the evidences based on large corpus such as Wikipedia.
I work on it with two researchers in UCLMR and build up a model that is roughly composed of Information Retrieval (IR) part, which retrieves evidences for a claim, and Natural Language Inference (NLI) part, which predicts whether retrieved evidences can support/refute the claim, and an aggregation part, which aggregates predictions for each claim-evidence pair.
As a result, our model got the 2nd place out of 24 groups/individuals in the shared task.
I’ll talk briefly about the shared-task, architecture of the model and how I worked collaboratively with other (foreign) researchers.
While staying at one of the top-level laboratories in a foreign country, I found that the atmosphere in the laboratory and students’ attitude toward their research are totally different from those in Japan.
Students are so active and they are enjoying discussion on research topic almost at every moment. They also find interesting papers/blogs every day and share them on slack.
The intriguing things here is that even professors are also involved in these activities, which makes these interactions more valuable and reasonable.
In addition to these, we could have lots of opportunities to listen to and talk with famous researchers coming from other research institutes such as DeepMind, Google and Microsoft since they gather in London.
I’ll talk about my experience in the internship especially focusing on the difference in atmosphere/belief toward research compared with those in Japan, and how much young researchers could get through the intern in a foreign country.