






摘要: 基于圖編碼器的路徑推理方法, 將知識(shí)圖譜多輪對(duì)話的實(shí)體間關(guān)系作為節(jié)點(diǎn)圖, 編碼器根據(jù)每輪對(duì)話對(duì)節(jié)點(diǎn)逐次編碼從而模擬語(yǔ)義推理過(guò)程, 最終預(yù)測(cè)當(dāng)前對(duì)話的答案
實(shí)體, 解決了對(duì)話中存在缺省詞和指代詞的問(wèn)題以及復(fù)雜語(yǔ)境下的特征提取問(wèn)題. 實(shí)驗(yàn)結(jié)果表明, 該方法更關(guān)注實(shí)體間的關(guān)系, 有助于保持推理的完整性和準(zhǔn)確性, 在一定程度上證
明了將上下文建模為關(guān)系節(jié)點(diǎn)圖的實(shí)用性和有效性.
關(guān)鍵詞: 知識(shí)圖譜; 自然語(yǔ)言處理; 多輪問(wèn)答; 卷積神經(jīng)網(wǎng)絡(luò)
中圖分類號(hào): TP391" 文獻(xiàn)標(biāo)志碼: A" 文章編號(hào): 1671-5489(2025)01-0076-07
Multi Round Conversational Model Based on Path Reasoning in Knowledge Graph
HUA Qingyuan1, PENG Tao1,2, CUI Hai2, BI Haijia1
(1. College of Computer Science and Technology, Jilin University, Changchun 130012, China;2. Key Laboratory of Symbolic Computation and Knowled
ge Engineering of Ministry of Education,Jilin University, Changchun 130012, China)
Abstract: Based on a path reasoning method of graph encoder, we used the entity relationships between multi rounds of dialogue in the k
nowledge graph as a node graph. The encoder sequentially encoded the nodes according to each round of dialogue to simulate the semantic reasoning process,
and utimately predicted the answer entity for the current dialogue. This approach solved the problems of missing words and pronouns in dialogues, as well as feature extraction problems i
n complex contexts. The experimental results show that the method focused more on the relationships between entities, which helped to maintain the integrity and accuracy of reasoning. To a cer
tain extent, it proved the practicality and effectiveness of modeling context as a relational node graph.
Keywords: knowledge graph; natural language process; multi round of question answering; convolutional neural network
收稿日期: 2024-02-07.
第一作者簡(jiǎn)介: 化青遠(yuǎn)(1999—), 男, 漢族, 碩士研究生, 從事知識(shí)圖譜問(wèn)答的研究, E-mail: huaqy21@mails.jlu.edu.cn. 通信作
者簡(jiǎn)介: 彭 濤(1977—), 男, 漢族, 博士, 教授, 博士生導(dǎo)師, 從事數(shù)據(jù)挖掘及Web挖掘、 信息檢索、 機(jī)器學(xué)習(xí)和自然語(yǔ)言處理的研究, E-mail: tpeng@jlu.edu.cn.
基金項(xiàng)目: 吉林省科技廳重點(diǎn)科技研發(fā)項(xiàng)目(批準(zhǔn)號(hào): 20210201131GX).
隨著大規(guī)模知識(shí)圖譜的發(fā)展, 基于知識(shí)圖譜的智能推理模型, 特別是基于知識(shí)圖譜的問(wèn)答領(lǐng)域(knowledge-based question answering, KBQA)受到廣泛關(guān)注. KBQA作為新一代搜索引擎[1], 不僅需要對(duì)
自然語(yǔ)言進(jìn)行語(yǔ)義分析和理解, 而且還依賴于知識(shí)圖譜中實(shí)體和關(guān)系的推理, 通常是從一組實(shí)體或關(guān)系開(kāi)始, 根據(jù)自然語(yǔ)言中給定的推理和約束檢索一組答案. 其中, 對(duì)話形式的知識(shí)
圖譜問(wèn)答(conversational KBQA, ConvKBQA)旨在基于知識(shí)圖譜推斷和檢索對(duì)話中的答案, 模型需要理解上下文并完善缺失的或參考的問(wèn)題.
主流的KBQA方法包括語(yǔ)義檢索(子圖檢索)[2]和基……