Multi-level retrieval with semantic Axiomatic Fuzzy Set clustering for question answering

This paper considers open-domain and multi-hop reading comprehension tasks that require complex multi-step reasoning processes. The study is particularly challenging because it requires a model to learn to explore “bridge” information to connect text snippets relevant to the answer. Unlike the usual neural-network-based retrieval models, which are difficult to interpret, this paper proposes a coarse-to-fine unsupervised evidence sentences retrieval model based on the Axiomatic Fuzzy Sets clustering with both reasoning ability and interpretability. According to the entities that appeared in a question, a chained inference retrieval was carried out to get the coarser candidate documents from knowledge bases. Then, sentence-level multi-feature scoring rules based on the part of speech and grammar are proposed. The Axiomatic Fuzzy Sets clustering algorithm based on the feature scores selects finer and sentence-level evidence by semantic descriptions. The retrieval process of the candidate sentences is unsupervised and straightforward, which does not require word embedding. Our model achieves state-of-the-art results in three open-domain QA datasets: HotpotQA, SQuAD Open and Natural Questions Open.