Recently, Huawei Cloud has made great progress in the field of knowledge computing. Based on the accumulation of leading technologies in the field of natural language processing, Huawei Cloud Speech semantic Innovation Lab and Huawei Noah’s Ark Laboratory won the first place in the international authoritative HotpotQA evaluation Fullwiki Setting track, and reached the first place in the answer fuzzy accuracy and Joint fuzzy accuracy (Joint F1) and other indicators.

HotpotQA Review (Fullwiki Setting track) Latest list (as of January 20, 2021)

As an important information carrier of enterprises, digital documents record a large amount of enterprise knowledge. How to quickly help employees get answers from digital documents is the pain point of enterprise document maintenance personnel. To solve this problem, on the one hand, strong reading comprehension techniques are needed, and on the other hand, a complex open domain question answering process is needed.

HotpotQA evaluation data is a new question-answering data set jointly launched by Stanford university, CMU and The University of Montreal. It is led by Professor Manning of Stanford University and Bengio, one of the three big deep learning companies. The data set consists of multi-hop complex questions and corresponding answers, and also contains supporting evidence to explain the source of the answers. In Stanford’s previously announced SQUAD task, which machines have repeatedly outperformed humans, it is relatively simple to find answers to questions from a single chapter. However, HotpotQA reviews are much more challenging when they need to find answers to questions from wikipedia or multiple articles and return a chain of reasoning from question to answer. The review also attracted well-known research institutions such as Google, Microsoft and Facebook, as well as well-known universities such as CMU, Stanford, University of Washington, Tsinghua University and Peking University.

Challenge the technical high point of multi-hop knowledge reasoning

The difficulty with HotpotQA’s review is that the machine has to jump through multiple articles to come up with an answer and return supporting evidence. The test was divided into two categories: Distractor Setting and Fullwiki Setting. Fullwiki Setting is more complex and closer to practical value than Distractor Setting. It needs to extract documents from the whole Wikipedia document, then extract paragraphs from the document, and finally extract answers from the paragraphs. The Distractor Setting track provides 10 alternate chapters. How to search for candidates from a large number of document paragraphs, and then understand the content of candidate paragraphs and extract supporting evidence has become the key of the competition.

Huawei Cloud and Huawei Noah’s Ark Laboratory put forward a new search target hop, which is used to collect reasoning evidence hidden in Wikipedia, solve complex multi-hop problems, and sort and merge the answers with Beam Search.

Search Hops from Wikipedia Text Map

Nowadays, enterprises are facing the transformation of digitalization and knowledge, and the key technology of knowledge transformation is knowledge understanding and reasoning, which is of great significance and promotion to enterprise knowledge transformation. In 2020, Huawei Cloud will launch the knowledge computing solution. Enterprises can build their own knowledge computing platform based on huawei cloud knowledge computing solution, which can be used for research and development, production, operation, sales, after-sales service and other enterprise core processes. At present, the scheme has been pioneered in petroleum, automobile, medical, chemical fiber, coal coking, steel, transportation and other industries.

Click follow to learn about the fresh technologies of Huawei Cloud