A Sentence Interaction Network for Modeling Dependence between Sentences Biao Liu, Minlie Huang Tsinghua University
Motivation In answer selection: The semantic relations between sentences is crucial for some NLP tasks.
Motivation We want to model the semantic relation between two sentences. What do cats look like? Cats have large eyes and furry bodies.
Motivation Convolutional Neural Network Architectures for Matching Natural Language Sentences
Motivation Dynamic Pooling and Unfolding Recursive Autoencoders for Paraphrase Detection
Background RNN
Background LSTM
Method Step 1 Use a LSTM to model the two sentences
Method Step 2: Introduce interactions
Method Step 2
Method Much complex semantic relations can be modeled with a vector. Through gates, different words can have different weights for classification.
Method SIN It’s powerful to model interactions between words, but not strong enough for phrase interactions. SIN-CONV Add a convolution layer to model phrases.
Experiments Answer Selection To select correct answers from a set of candidates for a given question.
Experiments Answer Selection Results
Experiments Dialogue Act Analysis To identify the dialogue act of a sentence in a dialogue.
Experiments Dialogue Act Analysis
Interaction Analysis Interaction Mechanism Analysis
Thanks