Dynamic reasoning network and method for multi-hop questions and answers

A dynamic, networked technology, applied in inference methods, biological neural network models, electrical and digital data processing, etc., to solve the problems that text information cannot be fully utilized, texts are rarely accessed, and paragraph selection is not performed.

Active Publication Date: 2020-10-23
SICHUAN UNIV
View PDF10 Cites 9 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

[0004] So far, the above three models have achieved some success, but there are still some limitations; first, most existing models directly reason in all given passages without passage selection to remove interfering passages, which increases the the amount of data processed; second, after existing models encode each text into a representation vector, each text is always rarely accessed, whether it is a question or a paragraph
And the model may not be able to obtain enough information by accessing the text only once or twice, which leads to the underutilization of text information

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • Dynamic reasoning network and method for multi-hop questions and answers
  • Dynamic reasoning network and method for multi-hop questions and answers
  • Dynamic reasoning network and method for multi-hop questions and answers

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0066] refer to figure 1 , is a schematic structural diagram of a dynamic reasoning network oriented to multi-hop question answering in the present invention. Specifically, a dynamic reasoning network oriented to multi-hop question answering includes:

[0067] Paragraph selector 1, which receives paragraphs and questions, and selects subparagraphs related to the answers to questions in the paragraphs to eliminate interfering information;

[0068] In this embodiment, the paragraph selector 1 includes a pre-trained BERT model with a sentence classification layer. The pre-trained BERT model is used to predict the similarity between questions and paragraphs. The paragraph selector 1 connects "[CLS]"+question +"[SEP]" + document + "[SEP]" as input for questions and paragraphs, and outputs a matching score between 0 and 1. Paragraph selection follows the following rules: if the paragraph contains an answer, the label is 2; if the paragraph contains at least one supporting statement...

Embodiment 2

[0086] refer to image 3 It is a flow diagram of a dynamic reasoning method for multi-hop question answering in the present invention, specifically, a dynamic reasoning method for multi-hop question answering, comprising the following steps:

[0087] S600: Receive a paragraph and a question, select at least one sub-paragraph related to the answer to the question in the paragraph; then perform step S700;

[0088] In this embodiment, after receiving the paragraphs and questions that need to be inferred, the paragraph selector 1 in Embodiment 1 will select the paragraphs in the paragraphs related to the answers to the questions as sub-paragraphs, so as to eliminate the interference information, and Coding the question with the resulting subparagraph, specifically, the question is set to Q=[q 1 ,q 2 ,...,q m ]∈R m×h , the subparagraph is set to P=[p 1 ,p 2 ,...,p n ]∈R n×h , m and n are the lengths of the question and paragraph respectively, h is the size of the hidden sta...

Embodiment 3

[0139] In this embodiment, the validity of the system of embodiment 1 and the method of embodiment 2 is verified, specifically, this embodiment is in the HotpotQA data set (the latest benchmark data set for multi-hop reasoning across multiple paragraphs) Evaluate the inference network of the present invention on the TriviaQA data set (a benchmark data set based on information retrieval (IR) construction), and compare the results of other models with the same parameter data,

[0140] In this example, baseline (the model used when Yang, Zhilin, et al. proposed the hotpot qa dataset in 2018), GRN (a model with a good ranking but unpublished papers on the leaderboard of the Hotpot qa dataset in 2019), QFE (the model proposed by Nishida, Kosuke, et al. in 2019), DFGN (the model proposed by Xiao, Yunxuan, et al. in 2019) are compared with the system of the present invention, and EM and F1 are used as measurement indicators, and EM is the exact match value and F1 is the F1 score.

...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

PUM

No PUM Login to view more

Abstract

The invention provides a dynamic reasoning network and method for multi-hop questions and answers, and the network comprises: a paragraph selector which receives paragraphs and questions, and selectssub-paragraphs related to answers of the questions from the paragraphs; the encoding module that is used for enhancing interaction between the problem and the sub-paragraphs by using collaborative attention, and calculating to obtain final vector representation of the sub-paragraphs and vector representation of the problem; the entity graph construction module that is used for constructing an entity graph; the dynamic reasoning module that is used for reasoning the entity graph, repeatedly reading texts to simulate the process of analyzing information by people and constructing a problem remodeling mechanism so as to repeatedly read problems and related important parts; and the answer prediction module that is connected with the encoding module, is connected with the dynamic reasoning module and is used for receiving the final vector representation of the sub-paragraphs and outputting to obtain four types of prediction. The network establishes a question remodeling mechanism, and the mechanism can repeatedly read questions to simulate the reading habits of people so as to improve the understanding and reasoning ability of the multi-hop reasoning question and answer model.

Description

technical field [0001] The invention belongs to the field of machine reading comprehension, and in particular relates to a dynamic reasoning network and method for multi-hop question answering. Background technique [0002] Machine reading comprehension is a task that obtains the correct answer to a given question by reasoning over a set of texts, while multi-hop reasoning question answering is a subtask of machine reading comprehension that aims to find the answer to a given question across multiple passages; Most current multi-hop reasoning question answering models usually obtain answers by visiting the question only once, so the model may not be able to obtain enough textual information. [0003] The multi-hop reasoning question answering model mainly has three research directions. The first one is based on the memory network, which uses the memory unit to combine the question with the information obtained in each round, and through continuous iterative reasoning after s...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

Application Information

Patent Timeline
no application Login to view more
IPC IPC(8): G06N5/04G06N3/02G06F40/205G06F40/279
CPCG06N3/02G06N5/04G06F40/205G06F40/279
Inventor 琚生根李晓辉陈润
Owner SICHUAN UNIV
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products

Browse by: Latest US Patents, China's latest patents, Technical Efficacy Thesaurus, Application Domain, Technology Topic.

© 2024 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap