Insertion-based Decoding with Automatically Inferred Generation Order

[kweonwooj]

Abstract

• propose a novel decoding algorithm - INDIGO, which generates text in an arbitrary order via insertion operation
• achieve competitive or even better performance in machine translation than conventional left-to-right generation.
  • Dataset : WMT16 RoEn, WMT18 EnTr, KFTT EnJa

Details

INDIGO

• INsertion based Decoding with Inferred Generation Order
• assumes generation orders as latent variables
• use relative position representation to capture generation order
• use Transformer model with relative position
• maximize evidence lower-bound (ELBO) of the original objective function and study four approximate posterior distribution of generation orders
  

Neural Autoregressive Decoding

• neural autoregressive model commonly learns the probability of a Y given X via product of probability of each token given X and previously generated tokens Y_t
• a common way to decode such sequence model was from left-to-right, as it is a natural for most human-beings to read sequences (strong inductive bias)
• however, L2R may not be the optimal option for generating all sequences
  • Japanese tend to produce better result in R2L
  • code generation is beneficial when generated on abstract syntax tree etc

Ordering as latent variable

• add order function pi to the conditional probability
• L2R can be recovered if z_t = t
  

Relative Representation of Positions

• it is essential to use relative representation to model position because we do not know how many tokens will be generated at the end
• relational vector is used to model relative positions of all tokens at each timestep. accumulating relational vector across all timestep leads to relational matrix

Insertion based Decoding

• INDIGO predicts next token and its relative position at each timestep, shown in Alg 1
  
  

Learning

• maximizing marginalized likelihood is intractable because we need to consider all T! permutations of tokens, given that tokens are now order-free
• instead, we maximize the evidence lower bound of original objective by introducing an approximate posterior distribution of generation orders which we can flexibly control
  
  

Experiment - Machine Translation

• Datasets
  • WMT16 RoEn 620k / 2k / 2k
  • WMT18 EnTr 207k / 3k / 3k
  • KFTT EnJa 405k / 1k / 1k
• Result
  • except for Random order, all pre-defined orders perform relatively similar, but L2R / R2L is best
  • Adaptive Order with beam=8 performs better than L2R, R2L in all language pairs
    

Experiment - Word Order Recovery / Code Generation

• improvement via INDIGO is more vivid in word order recovery and code generation tasks
  
  

Personal Thoughts

• paper was bit difficult to read
• predicting tokens and its positions autoregressively is an interesting idea
• wish more ablation was on what kind of tokens model predicts first in terms of POS, frequency etc
• interesting to see common-first approach is worse than L2R/R2L. surprised to find out how strong L2R inductive bias is

Link : https://arxiv.org/pdf/1902.01370.pdf
Authors : Gu et al. 2019

[lx385095967]

Do you understand the meaning of matrix B in the SAO method?