Abstract
Recently, many deep learning models have archived high results in question answering task with overall F1 scores above 0.88 on SQuAD datasets. However, many of these models have quite low F1 scores on why-questions. These F1 scores range from 0.57 to 0.7 on SQuAD v1.1 development set. This means these models are more appropriate to the extraction of answers for factoid questions than for why-questions. Why-questions are asked when explanations are needed. These explanations are possibly arguments or simply subjective opinions. Therefore, we propose an approach to finding the answer for why-question using discourse analysis and natural language inference. In our approach, natural language inference is applied to identify implicit arguments at sentence level. It is also applied in sentence similarity calculation. Discourse analysis is applied to identify the explicit arguments and the opinions at sentence level in documents. The results from these two methods are the answer candidates to be selected as the final answer for each why-question. We also implement a system with our approach. Our system can provide an answer for a why-question and a document as in reading comprehension test. We test our system with a Vietnamese translated test set which contains all why-questions of SQuAD v1.1 development set. The test results show that our system cannot beat a deep learning model in F1 score; however, our system can answer more questions (answer rate of 77.0%) than the deep learning model (answer rate of 61.0%).
Highlights
Question answering is a branch of information retrieval
We will find the answer of question Q “Why C?” by identifying the most appropriate elementary discourse units (EDUs), named S, for the question Q. is means the relation of S and C is the entailment with the highest score. en, we find all vertices {spi} connected to S by breadth-first search
We test the experiment systems on VnYQA dataset with NVIDIA Tesla M40 12GB GPU. e execution time and the GPU memory size of these models are shown in Table 7. e results in Table 7 show that our system needs more resources and it consumes more time than other systems because it uses two Bidirectional Encoder Representation from Transformers (BERT) fine-tuned models for EDU segmentation and natural language inference, and two stages of Rhetorical structure theory (RST) parsing at inner-sentential and intersentential levels
Summary
Question answering is a branch of information retrieval. Many early question answering systems used named entity extraction models to extract answer candidates from the retrieved documents; they selected the best five answer candidates for each question. ese systems were designed for answering factoid questions; their answers were usually nominal phrases of place, time, person’s name, etc.ese systems did not answer why-question well because the answers of why-questions are not always nominal phrases. Many early question answering systems used named entity extraction models to extract answer candidates from the retrieved documents; they selected the best five answer candidates for each question. Ese systems were designed for answering factoid questions; their answers were usually nominal phrases of place, time, person’s name, etc. Answering why-questions is a big question for many early systems and recent deep learning models. Is means those models were mostly trained for answering factoid questions. Rhetorical structure theory (RST) [13] views documents as sets of rhetorical relations between text units called elementary discourse units (EDUs) [22]. Erefore, RST-style parsing is very important to understand texts at document level. We can identify the premises and the conclusions of an argument or the reasons and the claims of an opinion if we have an efficient RST-style parser. Maple syrup was used to make sugar. is is why the tree is called a ‘sugar’ maple tree.” is text fragment presents an argument to explain the name “sugar maple.” We can recognize this argument and identify its premises and the conclusion by exploring its RSTstructure. is means we can find the answer of why-question in RST structures
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
