A Fast Unified Model for Parsing and Sentence Understanding Compilation theory A Fast Unified Model for Parsing and Sentence Understanding Antoine SOUSTELLE - Pierre RAINERO 02/12/2018

Summary Problematic SPINN overview Shift-reduce parser SPINN and TreeLSTM Advantages of the SPINN Conclusion Antoine SOUSTELLE - Pierre RAINERO 02/12/2018

Problematic Antoine SOUSTELLE - Pierre RAINERO 02/12/2018

SPINN overview Stack-augmented Parser-Interpreter Neural Network Combines parsing and interpretation Shift-reduce parser Antoine SOUSTELLE - Pierre RAINERO 02/12/2018

Shift-reduce parser A = B + 3 Reduce Shift Step Parse Stack Look Ahead Unscanned Parser Action   id = B + 3 Shift 1 = B + 3 2 id = + 3 3 id = id + 4 id = value Reduce by value 5 id = sums Reduce by sums 6 id = sums + int 7 id = sums + int eof 8 id = sums + value 9 10 assign Done Antoine SOUSTELLE - Pierre RAINERO 02/12/2018

SPINN Designed to produce a vector representation of a sentence as its output SPINN = Shift-reduce + TreeLong Short-Term Memory based on neural network Neural network : improve encoding of sentences by using their structure Sentence are not linear sequence Antoine SOUSTELLE - Pierre RAINERO 02/12/2018

Advantages of the SPINN Most words have multiple senses or meanings : Use the context (hybrid). Supports batched computation Speedup of up to 25x over other tree-structured models Antoine SOUSTELLE - Pierre RAINERO 02/12/2018

Conclusion Recent (2016) Part of deep learning expansion Antoine SOUSTELLE - Pierre RAINERO 02/12/2018

Bibliography https://arxiv.org/abs/1603.06021 https://fr.slideshare.net/tuvistavie/tree-lstm https://www.wikipedia.org/ Antoine SOUSTELLE - Pierre RAINERO 02/12/2018