《電子技術(shù)應(yīng)用》
您所在的位置:首頁(yè) > 其他 > 設(shè)計(jì)應(yīng)用 > 基于句粒度提示的大語(yǔ)言模型時(shí)序知識(shí)問(wèn)答方法
基于句粒度提示的大語(yǔ)言模型時(shí)序知識(shí)問(wèn)答方法
網(wǎng)絡(luò)安全與數(shù)據(jù)治理
李志東,,羅琪彬,,喬思龍
華北計(jì)算技術(shù)研究所大數(shù)據(jù)研發(fā)中心,北京100083
摘要: 知識(shí)問(wèn)答是自然語(yǔ)言處理領(lǐng)域的研究熱點(diǎn)之一,,而時(shí)序知識(shí)問(wèn)答還需考慮知識(shí)的時(shí)序關(guān)系,,更是研究難點(diǎn)所在,。當(dāng)前時(shí)序知識(shí)問(wèn)答方法通常將知識(shí)和問(wèn)題的詞向量相似度作為回答的重要依據(jù),忽略了知識(shí)所蘊(yùn)含的句粒度語(yǔ)義信息,。對(duì)此,,提出了一種基于句粒度提示的大語(yǔ)言模型時(shí)序知識(shí)問(wèn)答方法,首先通過(guò)對(duì)句粒度提示的改進(jìn),,讓大語(yǔ)言模型高效學(xué)習(xí)句粒度語(yǔ)義信息,,同時(shí)驗(yàn)證大語(yǔ)言模型在Zeroshot、Fewshot及弱監(jiān)督微調(diào)下時(shí)序知識(shí)問(wèn)答能力,。在ICEWS0515數(shù)據(jù)集上進(jìn)行實(shí)驗(yàn),,所提方法回答正確準(zhǔn)確率得到可觀提升,體現(xiàn)了基于句粒度提示的大語(yǔ)言模型時(shí)序知識(shí)問(wèn)答方法的有效性,。
中圖分類號(hào):TP391.1
文獻(xiàn)標(biāo)識(shí)碼:ADOI:10.19358/j.issn.2097-1788.2023.12.002
引用格式:李志東,,羅琪彬,,喬思龍.基于句粒度提示的大語(yǔ)言模型時(shí)序知識(shí)問(wèn)答方法[J].網(wǎng)絡(luò)安全與數(shù)據(jù)治理,2023,,42(12):7-13.
Large language model based on sentence granularity prompts for temporal knowledge Q&A approach
Li Zhidong,,Luo Qibin,Qiao Silong
Big Data R&D Center, North China Institute of Computing Technology, Beijing 100083, China
Abstract: Knowledge Q&A is one of the hot research topics in the field of natural language processing, and temporal knowledge Q&A is a difficult area of Q&A reasoning because it also needs to consider the temporal relationship of knowledge. Today′s research usually focuses on the word vector similarity between knowledge and questions as an important basis for answering, while ignoring the sentence granularity semantic information embedded in the knowledge. In this paper, we propose a method of temporal knowledge Q&A for large language models based on sentence granularity prompts. Firstly, by improving the sentence granularity prompts, the large language models can learn the sentence granularity semantic information efficiently, and then the temporal knowledge Q&A ability of large language models under Zeroshot, Fewshot and weaklysupervised finetuning is verified. The experiments are conducted on the ICEWS0515 dataset , and the accuracy of answers is significantly improved, which demonstrates the effectiveness of the temporal knowledge Q&A method for large language models based on sentence granularity prompts.
Key words : temporal knowledge graph questionanswering; large language models; prompt learning,;natural language processing

引言

業(yè)務(wù)系統(tǒng)中具有多種不同時(shí)間序列的數(shù)據(jù)信息,,將這些數(shù)據(jù)通過(guò)相關(guān)性和因果關(guān)系相聯(lián)系形成知識(shí)圖譜有助于快速深入地掌握時(shí)序信息。此外,,數(shù)據(jù)信息在時(shí)間維度上的語(yǔ)義表達(dá)不同,,包括年、月,、日等不同粒度,,跨時(shí)間粒度的語(yǔ)義表達(dá)會(huì)對(duì)問(wèn)答結(jié)果產(chǎn)生影響。由此,,時(shí)序知識(shí)圖譜(Temporal Knowledge Graph,,TKG)的產(chǎn)生可以對(duì)不同的時(shí)間序列數(shù)據(jù)生成一個(gè)多層的、多粒度的知識(shí)圖譜,,使得時(shí)序之間的關(guān)系得以清晰描述,。基于知識(shí)圖譜的問(wèn)答系統(tǒng)(Question Answering System based on Knowledge Graphs, KGQA)最早被用于提高企業(yè)的核心競(jìng)爭(zhēng)力,,由于企業(yè)經(jīng)營(yíng)過(guò)程中沉淀了許多知識(shí)但并不能得到很好的利用,,KGQA的出現(xiàn)使得知識(shí)的完全利用成為了可能。而TKG是在傳統(tǒng)的知識(shí)圖譜上對(duì)時(shí)間進(jìn)行延伸,,在三元組中加入時(shí)間維度,,格式為“[頭實(shí)體 關(guān)系 尾實(shí)體 時(shí)間]”。


作者信息

李志東,,羅琪彬,,喬思龍

(華北計(jì)算技術(shù)研究所大數(shù)據(jù)研發(fā)中心,北京100083)


文章下載地址:http://forexkbc.com/resource/share/2000005870


weidian.jpg

此內(nèi)容為AET網(wǎng)站原創(chuàng),,未經(jīng)授權(quán)禁止轉(zhuǎn)載,。