How State-Of-The-Art Models Can Deal With Long-Form Question Answering
18846Authors
nan
Abstract
Question answering is an essential task in nat- ural language processing. According to our review, the datasets for this task often contain short answers. We raise a research question of whether state-of-the-art models can perform well when a longer answer is needed. We pro- pose a dataset that contains much longer an- swers called FitQA1 and conduct a brief per- formance analysis among current state-of-the- art models on this dataset. The robust trans- former architecture like ALBERT achieved 90.9% F1 on SQuAD 2.0 but only got 47.3% F1 score on FitQA. Our hypothesis is that for longer context, the model needs to be guided to focus on longer dependent words. We con- duct a curriculum-learning-based framework. Experimental results show that our approach could improve the performance with the ap- propriate answer length up to 55.3% on F1. Acknowledgments: This work was supported by JST CREST Grant Number JPMJCR1513 and the Asian Office of Aerospace R&D (AOARD), Air Force Office of Scientific Re- search (Grant no. FA2386-19-1-4041) 1
Keywords
Full Content
How State-Of-The-Art Models Can Deal With Long-Form Question Answering Minh-Quan Bui, Vu Tran, Ha-Thanh Nguyen, Le-Minh Nguyen Japan Advanced Institute of Science and Technology Nomi, Ishikawa, Japan {quanbui,vu.tran, nguyenhathanh, nguyenml}@jaist.ac.jp Abstract Question answering is an essential task in nat- ural language processing. According to our review, the datasets for this task often contain short answers. We raise a research question of whether state-of-the-art models can perform well when a longer answer is needed. We pro- pose a dataset that contains much longer an- swers called FitQA1 and conduct a brief per- formance analysis among current state-of-the- art models on this dataset. The robust trans- former architecture like ALBERT achieved 90.9% F1 on SQuAD 2.0 but only got 47.3% F1 score on FitQA. Our hypothesis is that for longer context, the model needs to be guided to focus on longer dependent words. We con- duct a curriculum-learning-based framework. Experimental results show that our approach could improve the performance with the ap- propriate answer length up to 55.3% on F1. Acknowledgments: This work was supported by JST CREST Grant Number JPMJCR1513 and the Asian Office of Aerospace R&D (AOARD), Air Force Office of Scientific Re- search (Grant no. FA2386-19-1-4041) 1 Introduction Machine reading comprehension (MRC), or the abil- ity to read and understand the unstructured text and answer questions about it remains a challenging task in natural language understanding. This chal- lenge spurs the development of large datasets and 1The dataset will be published along with the paper Article: 5 Fat Loss Myths You Still Believe Question: How to Limit Metabolic Compensation? Context: How to Limit Metabolic Compensation The good news is there are some ways to reduce metabolic compensation. Here are some things to do: Do your best to maintain as much muscle as you can. The metabolic rate will not slow as much and be more resistance to fat regain. This means to make weight lifting the dominant part of your fitness regime during fat loss. Cardio becomes a little more important after weight loss, when the metabolic rate has lessened. You may want to save your cardio for after, rather than during the competition diet. Eat more protein, see the first point above about maintaining muscle mass. And probably increase the amount of protein as a percent of total calories. Do this during, but perhaps more importantly, after fat loss. Cycle the calorie gap, having times where you’re in a strong deficit and other times where you’re in no deficit at all. The recent MATADOR study (minimizing adaptive thermogenesis and deactivating obesity rebound) showed this strategy got better results, had less metabolic adaptation, and much longer lasting results. Don’t eat like an asshole when it all ends. Focus on blander foods and less variety of them. Doing the traditional burger, pizza, and cheesecake binges will trigger the brain’s hedonistic response and cause you to want more of that same dopamine hit all this when the metabolism is at its most vulnerable in terms of fat storage. And finally you …
Content truncated to 500 words
Labeled Text
Question answering is an essential task in natural language processing. According to our review, the datasets for this task often contain short answers. We raise a research question of whether state-of-the-art models can perform well when a longer answer is needed. We propose a dataset that contains much longer answers called FitQA1 and conduct a brief performance analysis among current state-of-theart models on this dataset. The robust transformer architecture like ALBERT achieved 90.9% F1 on SQuAD 2.0 but only got 47.3% F1 score on FitQA. Our hypothesis is that for longer context, the model needs to be guided to focus on longer dependent words. We conduct a curriculum-learning-based framework. Experimental results show that our approach could improve the performance with the appropriate answer length up to 55.3% on F1. Acknowledgments: This work was supported by JST CREST Grant Number JPMJCR1513 and the Asian Office of Aerospace R&D (AOARD), Air Force Office of Scientific Research (Grant no. FA2386-19-1-4041) 1
Publication Details
- Paper ID:
- 18846
- Year:
- 2020
- Publication:
- TIPSTER
- Publication Type:
- conf
- DOI:
- nan