Neural Lattice-to-Sequence Models for Uncertain Inputs

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1380–1389 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Neural Lattice-to-Sequence Models for Uncertain Inputs Matthias Sperber1, Graham Neubig2, Jan Niehues1, Alex Waibel1 1Karlsruhe Institute of Technology, Germany 2Carnegie Mellon University, USA matthias.sperber@kit.edu Abstract The input to a neural sequence-to- sequence model is often determined by an up-stream system, e.g. a word seg- menter, part of speech tagger, or speech recognizer. These up-stream models are potentially error-prone. Representing in- puts through word lattices allows mak- ing this uncertainty explicit by captur- ing alternative sequences and their poste- rior probabilities in a compact form. In this work, we extend the TreeLSTM (Tai et al., 2015) into a LatticeLSTM that is able to consume word lattices, and can be used as encoder in an attentional encoder- decoder model. We integrate lattice poste- rior scores into this architecture by extend- ing the TreeLSTM’s child-sum and forget gates and introducing a bias term into the attention mechanism. We experiment with speech translation lattices and report con- sistent improvements over baselines that translate either the 1-best hypothesis or the lattice without posterior scores. 1 Introduction In many natural language processing tasks, we will require a down-stream system to consume the input of an up-stream system, such as word segmenters, part of speech taggers, or automatic speech recognizers. Among these, one of the most prototypical and widely used examples is speech translation, where a down-stream translation sys- tem must consume the output of an up-stream au- tomatic speech recognition (ASR) system. Previous research on traditional phrase-based or tree-based statistical machine translation have used word lattices (e.g. Figure 1) as an effective tool to pass on uncertainties from a previous step 0: <s> 1: ah 5: qué 8: </s> 2: hay 4: qué 3: que 6: bueno 7: bueno 1 .4 .6 1 1 .2 .8 1 0: <s> 1: ah 5: qué 8: </s> 2: hay 4: qué 3: que 6: bueno 7: bueno 1 Figure 1: A lattice with 3 possible paths and pos- terior scores. Translating the whole lattice poten- tially allows for recovering from errors in its 1- best hypothesis. (Ney, 1999; Casacuberta et al., 2004). Several works have shown quality improvements by trans- lating lattices, compared to translating only the single best upstream output. Examples include translating lattice representations of ASR output (Saleem et al., 2004; Zhang et al., 2005; Matusov et al., 2008), multiple word segmentations, and morphological alternatives (Dyer et al., 2008). Recently, neural sequence-to-sequence (seq2seq) models (Kalchbrenner and Blun- som, 2013; Sutskever et al., 2014; Bahdanau et al., 2015) have often been preferred over the traditional methods for their strong empirical results and appealing end-to-end modeling. These models force us to rethink approaches to handling lattices, because their recurrent design no longer allows for efficient lattice decoding using dynamic programming as was used in the earlier works. As a remedy, Su et al. (2017) proposed replac- ing the sequential encoder by a lattice encoder …

Content truncated to 500 words