NLI Introduction
Challenges
- 词语匹配的多元性(同义词、一词多义)lexical gap、lexical variant;
- 短语匹配的结构性、多粒度计算语义;
- 文本匹配的层次性;
- 利用不相似部分的信息;
- 自然语言的基础现象:Variability of Semantic expression
- 是语言歧义的对偶问题;
- 语义表达的多变性、语言结构的多样性;
Approachs
models on NLI(Natural Language Infernce), concluded as: sentence-encoding-based vs. interaction-based
- 基于encoding的基本框架:
- input encoding —>concat vector —>vector & element-wise difference/product—> prediction
- Pros:
- 1、可预处理;2、易于建立索引;
- Cons:
- 1、失去语义焦点,易语义偏移;2、词上下文重要性难以衡量;
- 基于交互的基本框架:
- input encoding —> inference(interaction) —> composition —> prediction
- Pros:
- 1、把握语义焦点;2、可以对上下文重要性进行合理的建模;
- Cons:
- 1、信息损失(主要由交互带来);
Measure of Similarity
Models of Sentence Encoding Based
- BiLSTM-Max
- NSE
- Deep Gated Attn. BiLSTM
- Residual Stacked Encoder
- Reinforced Self-Attention Network
- Distance-based Self-Attention Network
- Hierarchical BiLSTM with Max Pooling
- Dynamic Self-Attention Model
Models of Interaction Based
- Decomposable Attention
- ESIM
- KIM
- Densely Interactive Inference Network (DIIN)
- BIMPM
- Multi-Way Attention
- DR-BiLSTM
- CAFE
- Densely-Connected Recurrent and Co-Attentive Network
- DMAN
- SLRC
- AFN