An abductive inference approach to cross-focus loss based on pre-trained models

A reasoning method and pre-training technology, applied in the field of artificial intelligence, can solve problems such as assignment, low probability of occurrence, and no way to judge, and achieve the effect of improving robustness and accuracy.

Active Publication Date: 2021-11-09
INST OF AUTOMATION CHINESE ACAD OF SCI
View PDF6 Cites 0 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

2) L2R 2 The method sorts these hypotheses according to the number of times they appear in the data set, but it is actually difficult to assign and sort the probability of occurrence. For these three hypotheses, we have no way to directly assign their probability of occurrence, and we There is also no way to judge which of the three hypotheses has the highest probability of occurrence and which has the lowest probability of occurrence

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
  • An abductive inference approach to cross-focus loss based on pre-trained models
  • An abductive inference approach to cross-focus loss based on pre-trained models
  • An abductive inference approach to cross-focus loss based on pre-trained models

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0054] Such as figure 1 Shown, O 1 and O 2 For any observation pair, there are four corresponding hypotheses in this observation pair, where H 1 and H 2 For the correct assumption, H 3 and H 4 wrong assumption. The L2R2 ranking algorithm will rank these four probabilities and select the part with a higher probability as the correct answer. But there are still some flaws in it. First, for some hypotheses, we have no way to judge which one is more likely to be the answer, for example figure 1 H in 1 with H 2 , the meanings of these two sentences are not much different in essence, so it is difficult to compare their probabilities as answers. And for two unrelated answers, it is even more difficult to compare their probabilities as answers. Second, the L2R2 method sums the probabilities of all correct answers to a fixed value, when assuming H 1 When the probability of being the answer increases, the sum of the probabilities of other hypotheses will also decrease, which ...

Embodiment 2

[0093] According to the abductive inference method based on the cross focus loss of the pre-trained model described in embodiment 1, a certain event relationship classification based on clue mining is applied, and the specific facts are as follows:

[0094] Event relationship detection is a natural language processing technology that deeply determines the correlation between two events and what kind of logical relationship they have. Its core task is to use the event as the basic semantic unit, and realize the identification and judgment of the logical relationship between events by analyzing the semantic correlation features between relationship type) two research tasks. Event relationship recognition is mainly to judge whether there is a logical or semantic relationship between two events, and pre-collecting samples for deep relationship detection between events is an important prerequisite for in-depth analysis of event logic relationships; event relationship judgment is in...

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 present invention provides an abductive reasoning method based on the cross-focus loss of the pre-training model, including: the observation pair O 1 and O 2 Combining with all assumptions, the input sequence is obtained; input the single input variable in the input sequence into the pre-training model to obtain the feature matrix corresponding to the sentence level, and then sum the word dimensions of the feature matrix to obtain the feature vector; traverse all the input sequences in the input sequence A single input variable to obtain the feature vector sequence; input the feature vector sequence into the bidirectional long-term short-term memory network to obtain the distributed feature representation, and then use the fully connected layer to map and sum to obtain the score of each input; the N labels in the input sequence Form N groups with true values ​​and all false values, and perform softmax within the group to obtain cross-prediction values; introduce clustering factors and weight factors, improve FocalLoss, and obtain training loss functions; optimize the training loss function , to get the optimal abductive reasoning model.

Description

technical field [0001] The present invention relates to the field of artificial intelligence, especially intelligent classification, for automatic classification and source tracing of language and events, and in particular to an abductive reasoning method based on cross-focus loss of pre-trained models. Background technique [0002] Human beings can understand natural language texts about everyday situations through past experience or known common sense. Given two observations O 1 and O 2 and two hypotheses H 1 and H 2 . Observing O 1 Under the condition, according to O 2 This result to guess leads to O 1 change to O 2 What are the causes of , and then from the candidate hypothesis H 1 , H 2 choose a more reasonable hypothesis. aNLI can also be said to look for the most plausible explanation. [0003] The purpose of abductive reasoning is mainly to help people understand texts and to capture whether there is a causal relationship between texts. The existing tech...

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
Patent Type & Authority Patents(China)
IPC IPC(8): G06N5/04G06N3/04
CPCG06N5/04G06N3/044
Inventor 陶建华徐铭杨国花张大伟刘通
Owner INST OF AUTOMATION CHINESE ACAD OF SCI
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