ROCLING-2021 Shared Task: Dimensional Sentiment Analysis for Educational Texts

Abstract This paper presents a framework to answer the questions that require various kinds of inference mechanisms (such as Extraction, Entailment-Judgement, and Summariza- tion). Most of the previous approaches adopt a rigid framework which handles only one inference mechanism. Only a few of them adopt several answer generation modules for providing different mecha- nisms; however, they either lack an aggre- gation mechanism to merge the answers from various modules, or are too compli- cated to be implemented with neural net- works. To alleviate the problems men- tioned above, we propose a divide-and- conquer framework, which consists of a set of various answer generation modules, a dispatch module, and an aggregation mod- ule. The answer generation modules are de- signed to provide different inference mech- anisms, the dispatch module is used to se- lect a few appropriate answer generation modules to generate answer candidates, and the aggregation module is employed to select the final answer. We test our frame- work on the 2020 Formosa Grand Chal- lenge Contest dataset. Experiments show 1 https://fgc.stpi.narl.org.tw/activity/techai2018 that the proposed framework outperforms the state-of-the-art Roberta-large model by about 11.4%. Keywords: QA, Framework, Divide-and-Conquer strategy, Answer Aggregation, Inference mechanism 1 Introduction Natural Language Inference (NLI) is an important topic in the Artificial Intelligence (AI) field, and any NLI related issue can be checked by asking an appropriate corresponding question (Chen, 2018). Therefore, the Question Answering (QA) task has become a very suitable testbed for evaluating NLI models and checking the progress of current tech- niques. Accordingly, the Ministry of Science and Technology of Taiwan has organized the Formosa Grand Challenge Open Contest series1 (FGC) in 2018, which mainly evaluates the reasoning/infer- ence capability on natural texts, to promote the AI progress in Taiwan. Specifically, this open contest covers a variety of answer modes; that is, it needs different inference mechanisms (such as Extraction, Entailment-Judgement, Aggregative-Operation, A Flexible and Extensible Framework for Multiple Answer Modes Question Answering Cheng-Chung Fan Shang-Bao Luo Kuang-Yu Chang Meng-Tse Wu Tzu-Man Wu Chao-Chun Liang Kuan-Yu Chen � Keh-Yih Su Chia-Chih Kuo � Pei-Jun Liao Chiao-Wei Hsu Shih-Hong Tsai Aleksandra Smolka Hsin-Min Wang Yu Tsao � Institute of Information Science, Academia Sinica � Research Center for Information Technology Innovation, Academia Sinica � Department of Computer Science and Information Engineering, National Taiwan University of Science and Technology {jjfan, newsboy3423, simonc, cwhsu, moju, doublebite, tzum.wu, alsm, ccliang, whm, kysu}@iis.sinica.edu.tw, {jerrykuo7727, s2w81234}@gmail.com kychen@mail.ntust.edu.tw yu.tsao@citi.sinica.edu.tw The 33rd Conference on Computational Linguistics and Speech Processing (ROCLING 2021) Taoyuan, Taiwan, October 15-16, 2021. The Association for Computational Linguistics and Chinese Language Processing 33 etc.) to get the desired answer. As a result, the sys- tem/framework must be able to handle various an- swer modes at the same time. The previous frameworks for the QA task could be classified into two main categories according to the number of answer modules adopted: (1) Single answer generation module (Trischler et al., 2017; Chen, 2018; Shoeybi et al., 2020; Zhang et al., 2020), which involves only one answer mode, and allows merely one …

Content truncated to 500 words