Log In Sign Up

A Comparison of Neural Models for Word Ordering

We compare several language models for the word-ordering task and propose a new bag-to-sequence neural model based on attention-based sequence-to-sequence models. We evaluate the model on a large German WMT data set where it significantly outperforms existing models. We also describe a novel search strategy for LM-based word ordering and report results on the English Penn Treebank. Our best model setup outperforms prior work both in terms of speed and quality.


page 1

page 2

page 3

page 4


On the Role of Pre-trained Language Models in Word Ordering: A Case Study with BART

Word ordering is a constrained language generation task taking unordered...

Sequence-to-Sequence Learning as Beam-Search Optimization

Sequence-to-Sequence (seq2seq) modeling has rapidly become an important ...

Online Versus Offline NMT Quality: An In-depth Analysis on English-German and German-English

We conduct in this work an evaluation study comparing offline and online...

Learning the Ordering of Coordinate Compounds and Elaborate Expressions in Hmong, Lahu, and Chinese

Coordinate compounds (CCs) and elaborate expressions (EEs) are coordinat...

Classical Structured Prediction Losses for Sequence to Sequence Learning

There has been much recent work on training neural attention models at t...

Humor@IITK at SemEval-2021 Task 7: Large Language Models for Quantifying Humor and Offensiveness

Humor and Offense are highly subjective due to multiple word senses, cul...

Studying word order through iterative shuffling

As neural language models approach human performance on NLP benchmark ta...

1 Introduction

Finding the best permutation of a multi-set of words is a non-trivial task due to linguistic aspects such as “syntactic structure, selective restrictions, subcategorization, and discourse considerations” [Elman1990]

. This makes the word-ordering task useful for studying and comparing different kinds of models that produce text in tasks such as general natural language generation 

[Reiter and Dale1997], image caption generation [Xu et al.2015], or machine translation [Bahdanau et al.2015]. Since plausible word order is an essential criterion of output fluency for all of these tasks, progress on the word-ordering problem is likely to have a positive impact on these tasks as well. Word ordering has often been addressed as syntactic linearization which is a strategy that involves using syntactic structures or part-of-speech and dependency labels  [Zhang and Clark2011, Zhang et al.2012, Zhang and Clark2015, Liu et al.2015, Puduppully et al.2016]. It has also been addressed as LM-based linearization which relies solely on language models and obtains better scores  [de Gispert et al.2014, Schmaltz et al.2016]

. Recently, Schmaltz et al. no-syntax showed that recurrent neural network language models 

[Mikolov et al.2010, Rnnlms]

with long short-term memory 

[Hochreiter and Schmidhuber1997, Lstm] cells are very effective for word ordering even without any explicit syntactic information.

We continue this line of work and make the following contributions. We compare several language models on the word-ordering task and propose a bag-to-sequence neural architecture that equips an LSTM decoder with explicit context of the bag-of-words (Bow) to be ordered. This model performs particularly strongly on WMT data and is complementary to an Rnnlm

: combining both yields large BLEU gains even for small beam sizes. We also propose a novel search strategy which outperforms a previous heuristic. Both techniques together surpass prior work on the Penn Treebank at

4x the speed.

2 Bag-to-Sequence Modeling with Attentional Neural Networks

Given the Bow {at, bottom, heap, now, of, the, the, we, ’re, .}, a word-ordering model may generate an output string now we ’re at the bottom of the heap .“. We can use an Rnnlm [Mikolov et al.2010]

to assign it a probability

by decomposing into conditionals:

Figure 1: (a) Attention-based seq2seq model and (b) bag2seq model used in this work.

Since we have access to the input Bows, we extend the model representation by providing the network additionally with the Bow to be ordered, thereby allowing it to focus explicitly on all tokens it generates in the output during decoding. Thus, instead of modeling the a priori distribution of sentences as in Eq. 1, we condition the distribution on :


This dependency is realized by the neural attention mechanism recently proposed by Bahdanau et al. bahdanau. The resulting bag-to-sequence model (bag2seq) is inspired by the attentional sequence-to-sequence model RNNsearch (seq2seq

) proposed by Bahdanau et al. bahdanau for neural machine translation between a source sentence

and a target sentence . Fig. 0(a) illustrates how seq2seq generates the -th target token using the decoder state

and the context vector

. The context vector is the weighted sum of source side annotations which encode sequence information.

To modify seq2seq for problems with unordered input, we make the encoder architecture order-invariant by replacing the recurrent layer with non-recurrent transformations of the word embeddings, as indicated by the missing arrows between source positions in Fig. 0(b). For convenience, we formalize as sequence in which words are sorted, e.g. alphabetically, so that we can refer to the -th word in the Bow. The model can be trained to recover word order in a sentence by using as input and the original sequence as target. This network architecture does not prevent words outside the Bow to appear in the output. Therefore, we explicitly constrain our beam decoder by limiting its available output vocabulary to the remaining tokens in the input bag at each time step, thereby ensuring that all model outputs are valid permutations of the input.

3 Search

Beam search is a popular decoding algorithm for neural sequence models [Sutskever et al.2014, Bahdanau et al.2015]. However, standard beam search suffers from search errors when applied to word ordering and Schmaltz et al. no-syntax reported that gains often do not saturate even with a large beam of 512. They suggested adding external unigram probabilities of the remaining words in the Bow

as future cost estimates to the beam-search scoring function and reported large gains for an

-gram LM and Rnnlm. We re-implement this future cost heuristic, , and further propose a new search heuristic, , which collects internal unigram statistics during decoding. We keep hypotheses in the beam if their score is close to a theoretical upper bound, the product of the best word probabilities given any history within the explored search space. For each word we maintain a heuristic score estimate which we initialize to 0. Each time the search algorithm visits a new context, we update the estimates such that is the current best score for :


where is the set of contexts (i.e. ordered prefixes in the form of ) explored by beam search so far. Thus, instead of computing a future cost, we compare the actual score of a partial hypothesis with the product of heuristic estimates of its words. This is especially useful for model combinations since all models are taken into account. We also implement hypothesis recombination to further reduce the number of search errors. More formally, at each time step our beam search keeps the best hypotheses according to scoring function using partial model score and estimates :


4 Experimental Setup

We evaluate using data from the English-German news translation task [Bojar et al.2015, WMT] and using the English Penn Treebank data [Marcus et al.1993, PTB]. Since additional knowledge sources are often available in practice, such as access to the source sentence in a translation scenario, we also report on bilingual experiments for the WMT task.

4.1 Data and evaluation

The WMT parallel training data includes Europarl v7, Common Crawl, and News Commentary v10. We use news-test2013 for tuning model combinations and news-test2015 for testing. All monolingual models for the WMT task were trained on the German news2015 corpus (51.3M sentences). For PTB, we use preprocessed data by Schmaltz et al. no-syntax for a fair comparison (40k sentences for training). We evaluate using the multi-bleu.perl script for WMT and for PTB.

4.2 Model settings

For WMT, the bag2seq parameter settings follow the recent NMT systems trained on WMT data. We use a 50k vocabulary, 620 dimensional word embeddings and 1000 hidden units in the decoder Lstm cells. On the encoder side, the input tokens are embedded to form annotations of the same size as the hidden units in the decoder. The Rnnlm is based on the “large” setup of Zaremba et al. rnnlm-google which uses an Lstm. Nplm

, a 5-gram neural feedforward language model, was trained for 10 epochs with a vocabulary size of 100k, 150 input and output units, 750 hidden units and 100 noise samples

[Vaswani et al.2013]. The -gram language model is a 5-gram model estimated with Srilm [Kneser and Ney1995]. For the bilingual setting, we implemented a seq2seq NMT system following Bahdanau et al. bahdanau using a beam size of 12 in line with recent NMT systems for WMT [Sennrich et al.2016]. Rnnlm, bag2seq and seq2seq

were implemented using TensorFlow 

[Abadi et al.2015] 111 and we used sgnmt for beam decoding222

Following Schmaltz et al. no-syntax, our neural models for PTB have a vocabulary of 16,161 incl. two different tokens and the Rnnlm is based on the “medium” setup of Zaremba et al. rnnlm-google. bag2seq uses 300 dimensional word embeddings and 500 hidden units in the decoder Lstm. We also compare to Gyro [de Gispert et al.2014] which explicitly targets the word-ordering problem. We extracted 1-gram to 5-gram phrase rules from the PTB training data and used an -gram LM for decoding. For model combinations, we combine the predictive distributions in a log-linear model and tune the weights by optimizing BLEU on the validation set with the BOBYQA algorithm [Powell2009].

Rnnlm Nplm -gram bag2seq seq2seq Bleu
Table 1: German word ordering on news-test2015 with beam=12, single models/combinations. Monolingual models use heuristic , bag2seq as a single model and bilingual models use no heuristic.

5 Results

5.1 Word Ordering on WMT data

The top of Tab. 1 shows that bag2seq outperforms all other language models by up to 4.2 BLEU on ordering German (bold numbers highlight its improvements). This suggests that explicitly presenting all available tokens to the decoder during search enables it to make better word order choices. A combination of Rnnlm, Nplm and -gram LM yields a higher score than the individual models, but further adding bag2seq yields a large gain of 4.5 BLEU confirming its suitability for the word-ordering task.

In the bilingual setting in the bottom of Tab. 1, the seq2seq model is given English input text and the beam decoder is constrained to generate permutations of German Bows. This is effectively a translation task with knowledge of the target Bows and seq2seq provides a strong baseline since it uses source sequence information. Still, adding bag2seq yields a 2.9 BLEU gain and adding it to the combination of all other models still improves by 1.8 BLEU. This suggests that it could also help for machine translation rescoring by selecting hypotheses that constitute good word orderings.

5.2 Word Ordering on the Penn Treebank

Tab. 2 shows the performance of different models and search heuristics on the Penn Treebank: using no heuristic (none) vs. and described in Section 3. Numbers in bold mark the best result for a given model. We compare against the LM-based method of de Gispert et al. gyro and the -gram and Rnnlm (Lstm) models of Schmaltz et al. no-syntax, of which the latter achieves the best BLEU score of 42.7. We can reproduce or surpass prior work for -gram and Rnnlm and show that outperforms for these models. This also holds when adding a 900k sample from the English Gigaword corpus as proposed by Schmaltz et al. no-syntax.333Results omitted from Tab. 2 to save space. However, bag2seq underperforms Rnnlm at this large beam size.

Model none
Previous work beam=512
Gyro444Note that this model has an advantage because longer sentences are processed in chunks of maximum length 20. 42.2
Ngram-512 38.6
Lstm-512 42.7
This work beam=512
-gram 35.7 38.6 38.9
Rnnlm 38.6 43.2 44.2
bag2seq 37.1 33.6 37.1
Table 2: BLEU scores for PTB word-ordering task (test). Ngram-512 and Lstm-512 are quoted from Schmaltz et al. (2016).

Since decoding is slow for large beam sizes, we compare bag2seq to the -gram and Rnnlm using a small beam of size 5 in Tab. 3. The first three rows show that decoding without heuristics is much easier with bag2seq and outperforms -gram and Rnnlm by a large margin with 33.4 BLEU. The Rnnlm needs heuristic to match this performance. For bag2seq, using heuristic estimates is worse than just using its partial scores for search. We suspect that its partial model scores are obfuscated by the heuristic estimates and the amount of their contribution should probably be tuned on a heldout set. Using the same beam size, ensembles yield better results but the best results are achieved by combining Rnnlm and bag2seq (37.9 BLEU). This confirms our findings on WMT data that these models are highly complementary for word ordering. The results for beam=64 follow this pattern and identify an interaction between heuristics and beam size. While we get the best results for beam=5 using , heuristic seems to perform better for larger beams, perhaps because the internal unigram statistics become more reliable. Finally, Rnnlm+bag2seq with and beam=64 outperforms Lstm-512 by 0.8 BLEU. This is significant because decoding in this configuration is also 4x faster than decoding with a single Rnnlm and beam=512 as shown in Fig. 2.

Model none
-gram 23.3 30.1 26.5
Rnnlm 24.5 33.6 29.7
bag2seq 33.4 27.0 31.7
Rnnlm-ensemble 25.5 34.2 30.6
bag2seq-ensemble 34.8 35.1 32.8
Rnnlm+bag2seq 35.7 37.9 34.4
Rnnlm 34.6 40.9 42.5
bag2seq 36.2 31.4 36.5
Rnnlm-ensemble 35.4 42.4 43.2
Rnnlm+bag2seq 40.5 43.1 43.5
Table 3: BLEU scores for PTB word-ordering task for different search heuristics and beam sizes (test).
Figure 2: Decoding time in relation to beam size for PTB word ordering task (test).

6 Conclusion

We have compared various models for the word-ordering task and proposed a new model architecture inspired by attention-based sequence-to-sequence models that helps performance for both German and English tasks. We have also proposed a novel search heuristic and found that using a model combination together with this heuristic and a modest beam size provides a good trade-off between speed and quality and outperforms prior work on the PTB task.


This work was partially supported by the U.K. Engineering and Physical Sciences Research Council (EPSRC grant EP/L027623/1). We thank the authors of Schmaltz et al. no-syntax for sharing their preprocessed data and helping to reproduce their results.


  • [Abadi et al.2015] Martın Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, et al. 2015.

    TensorFlow: Large-scale machine learning on heterogeneous systems, 2015.

    1. Software available from
  • [Bahdanau et al.2015] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In ICLR.
  • [Bojar et al.2015] Ondřej Bojar, Rajen Chatterjee, Christian Federmann, Barry Haddow, Matthias Huck, Chris Hokamp, Philipp Koehn, Varvara Logacheva, Christof Monz, Matteo Negri, Matt Post, Carolina Scarton, Lucia Specia, and Marco Turchi. 2015. Findings of the 2015 workshop on statistical machine translation. In Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 1–46. Association for Computational Linguistics.
  • [de Gispert et al.2014] Adrià de Gispert, Marcus Tomalin, and W Byrne. 2014. Word ordering with phrase-based grammars. In EACL, pages 259–268.
  • [Elman1990] Jeffrey L Elman. 1990. Finding structure in time. Cognitive science, 14(2):179–211.
  • [Hochreiter and Schmidhuber1997] Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural computation, 9(8):1735–1780.
  • [Kneser and Ney1995] Reinhard Kneser and Hermann Ney. 1995. Improved backing-off for m-gram language modeling. In ICASSP, volume 1, pages 181–184.
  • [Liu et al.2015] Yijia Liu, Yue Zhang, Wanxiang Che, and Bing Qin. 2015. Transition-based syntactic linearization. In NAACL.
  • [Marcus et al.1993] Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. 1993. Building a large annotated corpus of English: The Penn Treebank. Computational linguistics, 19(2):313–330.
  • [Mikolov et al.2010] Tomas Mikolov, Martin Karafiát, Lukas Burget, Jan Cernockỳ, and Sanjeev Khudanpur. 2010. Recurrent neural network based language model. In Interspeech, volume 2, page 3.
  • [Powell2009] Michael JD Powell. 2009. The BOBYQA algorithm for bound constrained optimization without derivatives. Cambridge NA Report NA2009/06, University of Cambridge, Cambridge.
  • [Puduppully et al.2016] Ratish Puduppully, Yue Zhang, and Manish Shrivastava. 2016. Transition-based syntactic linearization with lookahead features. In Proceedings of NAACL-HLT, pages 488–493.
  • [Reiter and Dale1997] Ehud Reiter and Robert Dale. 1997. Building applied natural language generation systems. Natural Language Engineering, 3(01):57–87.
  • [Schmaltz et al.2016] Allen Schmaltz, Alexander M Rush, and Stuart M Shieber. 2016. Word ordering without syntax. In EMNLP.
  • [Sennrich et al.2016] Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Edinburgh neural machine translation systems for wmt 16. In Proceedings of the First Conference on Machine Translation, pages 371–376.
  • [Sutskever et al.2014] Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014. Sequence to sequence learning with neural networks. In Proceedings of NIPS.
  • [Vaswani et al.2013] Ashish Vaswani, Yinggong Zhao, Victoria Fossum, and David Chiang. 2013. Decoding with large-scale neural language models improves translation. In EMNLP.
  • [Xu et al.2015] Kelvin Xu, Jimmy Lei Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhutdinov, Richard S. Zemel, and Yoshua Bengio. 2015. Show, attend and tell: Neural image caption generation with visual attention. arXiv preprint arXiv:1502.03044.
  • [Zaremba et al.2014] Wojciech Zaremba, Ilya Sutskever, and Oriol Vinyals. 2014. Recurrent neural network regularization. arXiv preprint arXiv:1409.2329.
  • [Zhang and Clark2011] Yue Zhang and Stephen Clark. 2011. Syntax-based grammaticality improvement using CCG and guided search. In EMNLP, pages 1147–1157.
  • [Zhang and Clark2015] Yue Zhang and Stephen Clark. 2015. Discriminative syntax-based word ordering for text generation. Computational Linguistics, 41(3):503–538.
  • [Zhang et al.2012] Yue Zhang, Graeme Blackwood, and Stephen Clark. 2012. Syntax-based word ordering incorporating a large-scale language model. In EACL, pages 736–746.