A method and system for machine reading comprehension based on multi-step evidence reasoning
A technology of evidence reasoning and reading comprehension, applied in the field of multi-step evidence reasoning machine reading comprehension methods and systems, which can solve problems such as low accuracy and recall, failure to show modeling problem relationships, and error cascades, etc., to achieve simple implementation Effect
Active Publication Date: 2022-07-22
INST OF INFORMATION ENG CHINESE ACAD OF SCI
View PDF6 Cites 0 Cited by
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Although the method based on the binary classifier is easy to implement, it has poor interpretability, and the accuracy and recall of unanswerable questions are also low.
Based on relational networks or verifiable methods, although they have a certain increase in accuracy, they do not show the relationship between the modeling problem and the article, and may cause problems such as error cascades
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 moreImage
Smart Image Click on the blue labels to locate them in the text.
Smart ImageViewing Examples
Examples
Experimental program
Comparison scheme
Effect test
Embodiment Construction
[0029] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0030] The general framework of the present invention is as follows figure 1 shown, including a global decoder, a multi-step evidence inference engine, and an answer detector.
[0031] 1. Global Encoder
[0032] figure 2 is the schematic diagram of the global encoder. The global encoder contains a pretrained language model and a global encoding module for better semantic understanding.
[0033] Similar to the processing method of the current pre-trained language model, the present invention adds a classification embedded identifier [CLS], and separates the article X with a set tag [SEP] P and question X Q , where X P Used to represent all characters in the article, the same as X Q Represents all characters in the question. So the input of the present invention can be expressed as [CLS,X Q , SEP, X P , Sep]. The input character v...
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
Login to View More Abstract
The invention discloses a machine reading comprehension method and system based on multi-step evidence reasoning. The steps of the method include: 1) input the article P and the question Q into a global encoder to generate an input character vector representation U; 2) encode U to obtain an input representation g and input it to a multi-step evidence inference engine; 3) multi-step evidence inference The machine infers according to g to obtain the starting evidence vector and ending evidence vector related to question Q; 4) According to the final calculation of the multi-step evidence reasoning engine, the starting evidence vector s related to question Q is obtained T+1 , the end evidence vector e T+1 and the expression vector q of the problem cls Calculate the score of question Q. When the score is higher than the set threshold θ, it is determined that the question Q cannot be answered; otherwise, it is determined that there is an answer to the question Q, and the starting position start-position and the answer obtained from the article P are extracted. end-position.
Description
technical field [0001] The invention belongs to the field of natural language processing, and particularly relates to a multi-step evidence reasoning machine reading comprehension method and system. Background technique [0002] Machine reading comprehension (MRC) has seen a boom in recent research in the field of natural language processing, with various neural network models rapidly approaching human parity on some benchmarks, with the aim of encouraging machines to be able to understand a given passage content and answer the questions. Extractive machine reading comprehension is one of the branches. It mainly extracts a continuous segment from the article as the final answer of the text. However, there is a huge assumption in this task, that is, every question can be answered in the article. , in this case, the model only needs to match the paragraphs that are most similar to the question, and does not really understand whether the question is implied by the text. How t...
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
Login to View More Patent Type & Authority Patents(China)
IPC IPC(8): G06F16/33G06F16/332G06F16/35G06K9/62G06N3/04G06N5/04
CPCG06F16/3344G06F16/3346G06F16/3329G06F16/35G06N5/04G06N3/047G06N3/048G06N3/044G06F18/2415G06F18/241
Inventor 胡玥彭伟
Owner INST OF INFORMATION ENG CHINESE ACAD OF SCI