困惑度、BLEU、ROUGE详解:语言模型评估指标计算与解读
从熵到困惑度,从 n-gram 到嵌入相似度,掌握评估指标的本质。本节基于《AI工程》第3章「Evaluation Methodology」— Understanding Language Modeling Metrics(熵、交叉熵、困惑度)、Exact Evaluation(BLEU、ROUGE、嵌入相似度)。
一、指标全景图
Chip Huyen 在《AI工程》第3章系统阐述了语言模型评估指标。理解这些指标是设计评估体系、解读模型表现的基础。书中将指标分为:语言建模指标(熵、困惑度)、精确评估(功能正确性、参考相似度、嵌入)。
#mermaid-svg-VkwLEuY6FMGvn5tq{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-VkwLEuY6FMGvn5tq .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-VkwLEuY6FMGvn5tq .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-VkwLEuY6FMGvn5tq .error-icon{fill:#552222;}#mermaid-svg-VkwLEuY6FMGvn5tq .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-VkwLEuY6FMGvn5tq .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-VkwLEuY6FMGvn5tq .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-VkwLEuY6FMGvn5tq .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-VkwLEuY6FMGvn5tq .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-VkwLEuY6FMGvn5tq .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-VkwLEuY6FMGvn5tq .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-VkwLEuY6FMGvn5tq .marker{fill:#333333;stroke:#333333;}#mermaid-svg-VkwLEuY6FMGvn5tq .marker.cross{stroke:#333333;}#mermaid-svg-VkwLEuY6FMGvn5tq svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-VkwLEuY6FMGvn5tq p{margin:0;}#mermaid-svg-VkwLEuY6FMGvn5tq .edge{stroke-width:3;}#mermaid-svg-VkwLEuY6FMGvn5tq .section–1 rect,#mermaid-svg-VkwLEuY6FMGvn5tq .section–1 path,#mermaid-svg-VkwLEuY6FMGvn5tq .section–1 circle,#mermaid-svg-VkwLEuY6FMGvn5tq .section–1 polygon,#mermaid-svg-VkwLEuY6FMGvn5tq .section–1 path{fill:hsl(240, 100%, 76.2745098039%);}#mermaid-svg-VkwLEuY6FMGvn5tq .section–1 text{fill:#ffffff;}#mermaid-svg-VkwLEuY6FMGvn5tq .node-icon–1{font-size:40px;color:#ffffff;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-edge–1{stroke:hsl(240, 100%, 76.2745098039%);}#mermaid-svg-VkwLEuY6FMGvn5tq .edge-depth–1{stroke-width:17;}#mermaid-svg-VkwLEuY6FMGvn5tq .section–1 line{stroke:hsl(60, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-VkwLEuY6FMGvn5tq .disabled,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled circle,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled text{fill:lightgray;}#mermaid-svg-VkwLEuY6FMGvn5tq .disabled text{fill:#efefef;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-0 rect,#mermaid-svg-VkwLEuY6FMGvn5tq .section-0 path,#mermaid-svg-VkwLEuY6FMGvn5tq .section-0 circle,#mermaid-svg-VkwLEuY6FMGvn5tq .section-0 polygon,#mermaid-svg-VkwLEuY6FMGvn5tq .section-0 path{fill:hsl(60, 100%, 73.5294117647%);}#mermaid-svg-VkwLEuY6FMGvn5tq .section-0 text{fill:black;}#mermaid-svg-VkwLEuY6FMGvn5tq .node-icon-0{font-size:40px;color:black;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-edge-0{stroke:hsl(60, 100%, 73.5294117647%);}#mermaid-svg-VkwLEuY6FMGvn5tq .edge-depth-0{stroke-width:14;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-0 line{stroke:hsl(240, 100%, 83.5294117647%);stroke-width:3;}#mermaid-svg-VkwLEuY6FMGvn5tq .disabled,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled circle,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled text{fill:lightgray;}#mermaid-svg-VkwLEuY6FMGvn5tq .disabled text{fill:#efefef;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-1 rect,#mermaid-svg-VkwLEuY6FMGvn5tq .section-1 path,#mermaid-svg-VkwLEuY6FMGvn5tq .section-1 circle,#mermaid-svg-VkwLEuY6FMGvn5tq .section-1 polygon,#mermaid-svg-VkwLEuY6FMGvn5tq .section-1 path{fill:hsl(80, 100%, 76.2745098039%);}#mermaid-svg-VkwLEuY6FMGvn5tq .section-1 text{fill:black;}#mermaid-svg-VkwLEuY6FMGvn5tq .node-icon-1{font-size:40px;color:black;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-edge-1{stroke:hsl(80, 100%, 76.2745098039%);}#mermaid-svg-VkwLEuY6FMGvn5tq .edge-depth-1{stroke-width:11;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-1 line{stroke:hsl(260, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-VkwLEuY6FMGvn5tq .disabled,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled circle,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled text{fill:lightgray;}#mermaid-svg-VkwLEuY6FMGvn5tq .disabled text{fill:#efefef;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-2 rect,#mermaid-svg-VkwLEuY6FMGvn5tq .section-2 path,#mermaid-svg-VkwLEuY6FMGvn5tq .section-2 circle,#mermaid-svg-VkwLEuY6FMGvn5tq .section-2 polygon,#mermaid-svg-VkwLEuY6FMGvn5tq .section-2 path{fill:hsl(270, 100%, 76.2745098039%);}#mermaid-svg-VkwLEuY6FMGvn5tq .section-2 text{fill:#ffffff;}#mermaid-svg-VkwLEuY6FMGvn5tq .node-icon-2{font-size:40px;color:#ffffff;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-edge-2{stroke:hsl(270, 100%, 76.2745098039%);}#mermaid-svg-VkwLEuY6FMGvn5tq .edge-depth-2{stroke-width:8;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-2 line{stroke:hsl(90, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-VkwLEuY6FMGvn5tq .disabled,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled circle,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled text{fill:lightgray;}#mermaid-svg-VkwLEuY6FMGvn5tq .disabled text{fill:#efefef;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-3 rect,#mermaid-svg-VkwLEuY6FMGvn5tq .section-3 path,#mermaid-svg-VkwLEuY6FMGvn5tq .section-3 circle,#mermaid-svg-VkwLEuY6FMGvn5tq .section-3 polygon,#mermaid-svg-VkwLEuY6FMGvn5tq .section-3 path{fill:hsl(300, 100%, 76.2745098039%);}#mermaid-svg-VkwLEuY6FMGvn5tq .section-3 text{fill:black;}#mermaid-svg-VkwLEuY6FMGvn5tq .node-icon-3{font-size:40px;color:black;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-edge-3{stroke:hsl(300, 100%, 76.2745098039%);}#mermaid-svg-VkwLEuY6FMGvn5tq .edge-depth-3{stroke-width:5;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-3 line{stroke:hsl(120, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-VkwLEuY6FMGvn5tq .disabled,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled circle,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled text{fill:lightgray;}#mermaid-svg-VkwLEuY6FMGvn5tq .disabled text{fill:#efefef;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-4 rect,#mermaid-svg-VkwLEuY6FMGvn5tq .section-4 path,#mermaid-svg-VkwLEuY6FMGvn5tq .section-4 circle,#mermaid-svg-VkwLEuY6FMGvn5tq .section-4 polygon,#mermaid-svg-VkwLEuY6FMGvn5tq .section-4 path{fill:hsl(330, 100%, 76.2745098039%);}#mermaid-svg-VkwLEuY6FMGvn5tq .section-4 text{fill:black;}#mermaid-svg-VkwLEuY6FMGvn5tq .node-icon-4{font-size:40px;color:black;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-edge-4{stroke:hsl(330, 100%, 76.2745098039%);}#mermaid-svg-VkwLEuY6FMGvn5tq .edge-depth-4{stroke-width:2;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-4 line{stroke:hsl(150, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-VkwLEuY6FMGvn5tq .disabled,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled circle,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled text{fill:lightgray;}#mermaid-svg-VkwLEuY6FMGvn5tq .disabled text{fill:#efefef;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-5 rect,#mermaid-svg-VkwLEuY6FMGvn5tq .section-5 path,#mermaid-svg-VkwLEuY6FMGvn5tq .section-5 circle,#mermaid-svg-VkwLEuY6FMGvn5tq .section-5 polygon,#mermaid-svg-VkwLEuY6FMGvn5tq .section-5 path{fill:hsl(0, 100%, 76.2745098039%);}#mermaid-svg-VkwLEuY6FMGvn5tq .section-5 text{fill:black;}#mermaid-svg-VkwLEuY6FMGvn5tq .node-icon-5{font-size:40px;color:black;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-edge-5{stroke:hsl(0, 100%, 76.2745098039%);}#mermaid-svg-VkwLEuY6FMGvn5tq .edge-depth-5{stroke-width:-1;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-5 line{stroke:hsl(180, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-VkwLEuY6FMGvn5tq .disabled,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled circle,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled text{fill:lightgray;}#mermaid-svg-VkwLEuY6FMGvn5tq .disabled text{fill:#efefef;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-6 rect,#mermaid-svg-VkwLEuY6FMGvn5tq .section-6 path,#mermaid-svg-VkwLEuY6FMGvn5tq .section-6 circle,#mermaid-svg-VkwLEuY6FMGvn5tq .section-6 polygon,#mermaid-svg-VkwLEuY6FMGvn5tq .section-6 path{fill:hsl(30, 100%, 76.2745098039%);}#mermaid-svg-VkwLEuY6FMGvn5tq .section-6 text{fill:black;}#mermaid-svg-VkwLEuY6FMGvn5tq .node-icon-6{font-size:40px;color:black;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-edge-6{stroke:hsl(30, 100%, 76.2745098039%);}#mermaid-svg-VkwLEuY6FMGvn5tq .edge-depth-6{stroke-width:-4;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-6 line{stroke:hsl(210, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-VkwLEuY6FMGvn5tq .disabled,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled circle,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled text{fill:lightgray;}#mermaid-svg-VkwLEuY6FMGvn5tq .disabled text{fill:#efefef;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-7 rect,#mermaid-svg-VkwLEuY6FMGvn5tq .section-7 path,#mermaid-svg-VkwLEuY6FMGvn5tq .section-7 circle,#mermaid-svg-VkwLEuY6FMGvn5tq .section-7 polygon,#mermaid-svg-VkwLEuY6FMGvn5tq .section-7 path{fill:hsl(90, 100%, 76.2745098039%);}#mermaid-svg-VkwLEuY6FMGvn5tq .section-7 text{fill:black;}#mermaid-svg-VkwLEuY6FMGvn5tq .node-icon-7{font-size:40px;color:black;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-edge-7{stroke:hsl(90, 100%, 76.2745098039%);}#mermaid-svg-VkwLEuY6FMGvn5tq .edge-depth-7{stroke-width:-7;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-7 line{stroke:hsl(270, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-VkwLEuY6FMGvn5tq .disabled,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled circle,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled text{fill:lightgray;}#mermaid-svg-VkwLEuY6FMGvn5tq .disabled text{fill:#efefef;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-8 rect,#mermaid-svg-VkwLEuY6FMGvn5tq .section-8 path,#mermaid-svg-VkwLEuY6FMGvn5tq .section-8 circle,#mermaid-svg-VkwLEuY6FMGvn5tq .section-8 polygon,#mermaid-svg-VkwLEuY6FMGvn5tq .section-8 path{fill:hsl(150, 100%, 76.2745098039%);}#mermaid-svg-VkwLEuY6FMGvn5tq .section-8 text{fill:black;}#mermaid-svg-VkwLEuY6FMGvn5tq .node-icon-8{font-size:40px;color:black;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-edge-8{stroke:hsl(150, 100%, 76.2745098039%);}#mermaid-svg-VkwLEuY6FMGvn5tq .edge-depth-8{stroke-width:-10;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-8 line{stroke:hsl(330, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-VkwLEuY6FMGvn5tq .disabled,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled circle,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled text{fill:lightgray;}#mermaid-svg-VkwLEuY6FMGvn5tq .disabled text{fill:#efefef;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-9 rect,#mermaid-svg-VkwLEuY6FMGvn5tq .section-9 path,#mermaid-svg-VkwLEuY6FMGvn5tq .section-9 circle,#mermaid-svg-VkwLEuY6FMGvn5tq .section-9 polygon,#mermaid-svg-VkwLEuY6FMGvn5tq .section-9 path{fill:hsl(180, 100%, 76.2745098039%);}#mermaid-svg-VkwLEuY6FMGvn5tq .section-9 text{fill:black;}#mermaid-svg-VkwLEuY6FMGvn5tq .node-icon-9{font-size:40px;color:black;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-edge-9{stroke:hsl(180, 100%, 76.2745098039%);}#mermaid-svg-VkwLEuY6FMGvn5tq .edge-depth-9{stroke-width:-13;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-9 line{stroke:hsl(0, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-VkwLEuY6FMGvn5tq .disabled,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled circle,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled text{fill:lightgray;}#mermaid-svg-VkwLEuY6FMGvn5tq .disabled text{fill:#efefef;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-10 rect,#mermaid-svg-VkwLEuY6FMGvn5tq .section-10 path,#mermaid-svg-VkwLEuY6FMGvn5tq .section-10 circle,#mermaid-svg-VkwLEuY6FMGvn5tq .section-10 polygon,#mermaid-svg-VkwLEuY6FMGvn5tq .section-10 path{fill:hsl(210, 100%, 76.2745098039%);}#mermaid-svg-VkwLEuY6FMGvn5tq .section-10 text{fill:black;}#mermaid-svg-VkwLEuY6FMGvn5tq .node-icon-10{font-size:40px;color:black;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-edge-10{stroke:hsl(210, 100%, 76.2745098039%);}#mermaid-svg-VkwLEuY6FMGvn5tq .edge-depth-10{stroke-width:-16;}#mermaid-svg-VkwLEuY6FMGvn5tq .section-10 line{stroke:hsl(30, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-VkwLEuY6FMGvn5tq .disabled,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled circle,#mermaid-svg-VkwLEuY6FMGvn5tq .disabled text{fill:lightg



