结构化预测(Structered learning)
1 Introduction
- f:X→Y——输入输出不一定是向量
- 应用:语音辨识、翻译、文法剖析、位置监测、总结
Unified Framework:
- Training:F:X∗Y→R
- Testing:y~=argmaxy∈YF(x,y)
Statistics Unified Framework:
- Training:P:X∗Y→[0,1]
- Testing:Given x, y~=argmaxy∈YP(x,y)
2 three problems
- Evaluation:F(x,y)长什么样?
- Inference:怎么解argmax问题?
- Training:给定训练数据,如何训练F(x,y)?
3 结构化预测线性模型
- F(x,y)=w1ϕ1(x,y)+w2ϕ2(x,y)+…=wϕ(x,y)
- ∀y∈Y−{y^r},wϕ(xr,y^r)>wϕ(xr,y)
- 更新:w←w+ϕ(xr,y^r)−ϕ(xr,y~r)