Multi-Context Reasoning for Question Answering
This paper deals with the problem of Answer Processing (AP) in a context-aware setting. In this context we refer to the context-aware semantic processing task which involves inferring the relevant information from a sentence. However, there are few clear criteria that can achieve the best scores for an appropriate task without the knowledge or ability of the human reader. To address this, this paper presents a new framework to model task-specific semantic information from a corpus using multi-scale attention mechanism. The framework is based on a novel method that we call Multi-Selection-Context Multiparameter Attention (M-CEAM). Our system generates sentences in a high dimensional context with multi-scale attention mechanism, but the task is different from typical human-authored text. We provide an efficient implementation of our framework by means of a supervised training and annotation pipeline for our system. In our experimental results, we show that M-CEAM outperforms state-of-the-art semantic and inference-based approaches on several tasks.