Statistical Language Models based on Neural Networks - Faculty of ...

More documents

Recommendations

Info

WER [%] 12.5 12 11.5 11 10.5 10 9.5 9 10 0 10 1 Figure 5.2: WER on Eval 92 after rescoring with increasing size of N-best list, the baseline is obtained with 5-gram model. probably more useful to either produce wider lattices to allow more diverse paths to be encoded in the lattice, or to use neural net language models directly during decoding. 5.2.1 Approximation of RNNME using n-gram models As was described in Section 3.4.3, it is possible to approximate complex generative language models by sampling huge amount of data and building usual n-gram models <strong>based</strong> on the generated data. This can be seen as an attempt to precompute large lookup table for likely n-gram entries. To demonstrate potential of this technique which allows trivial integration of RNN and RNNME models directly into decoders, 15 billion words were generated from RNNME-480 model. A Good-Turing smoothed 5-gram model was built on top of these data, and various results are reported in Table 5.5. It should be noted that while using modified Kneser-Ney smoothing provides slightly better results for standalone models <strong>based</strong> on the generated data, the results after interpolation with the baseline 5-gram model are worse than if Good-Turing smoothing is used. N Based on the results reported in Table 5.5, it is possible to obtain around 0.6% WER reduction by rescoring lattices using models that were trained on additional data that were generated from the RNNME-480 model. However, the n-gram model <strong>based</strong> on the 68 10 2 10 3
Table 5.5: Results for models <strong>based</strong> on data sampled from RNNME-480 model (15B words). Model Perplexity WER [%] heldout Eval 92 Eval 93 GT3 105 13.0 17.6 KN5 (no count cutoffs) 80 12.0 16.6 Approximated RNNME-480 80 11.7 16.2 Approximated RNNME-480 + KN5 75 11.4 16.0 Full RNNME-480 59 10.2 13.7 Table 5.6: Results for pruned models <strong>based</strong> on data sampled from RNNME-480 model (15B words). Model WER [%] Number of n-grams Eval 92 Eval 93 GT3 13.0 17.6 11.1M KN5 (no count cutoffs) 12.0 16.6 68M Approximated RNNME-480 + KN5 11.4 16.0 846M Approximated RNNME-480 + KN5, pruning 1e-9 11.6 16.0 33M Approximated RNNME-480 + KN5, pruning 1e-8 12.2 16.7 9.5M Approximated RNNME-480 + KN5, pruning 1e-7 12.9 17.5 1.8M generated data is huge, and cannot be used directly in the decoder. Thus, additional experiments were performed with models that were pruned down in size, as reported in Table 5.6. Pruning was performed using SRILM toolkit and entropy pruning technique described in [73]. It can be seen that even after pruning, the approximated models remain competitive with the baseline 5-gram model. Conclusion of the data sampling experiments is that it is possible to approximate computationally complex language models by precomputing results for frequent n-grams. In theory, by sampling infinite amount of data and by building n-gram models with infinite order, this technique can be used for converting RNN models into n-gram models without any loss of precision. However in practice, it seems difficult to obtain more that 20% - 30% of improvement that the original model provides. Still, even this can be interesting in some situations, as the approximated models can be used directly in decoders with no additional effort - the only thing that changes is the training data. 69
Page 1 and 2:
VYSOKÉ UČENÍ TECHNICKÉ V BRNĚ
Page 3 and 4:
Abstrakt Statistické jazykové mod
Page 5 and 6:
Contents 1 Introduction 4 1.1 Motiv
Page 7 and 8:
6.2.3 Reduction of Vocabulary Size
Page 9 and 10:
Maybe the most popular vision of fu
Page 11 and 12:
Chapter 6 presents further extensio
Page 13 and 14:
Chapter 2 Overview of Stati
Page 15 and 16:
2.1 Evaluation 2.1.1 Perplexity Eva
Page 17 and 18:
ALP can be used to obtain prior pro
Page 19 and 20:
• Good theoretical motivation •
Page 21 and 22: abilities of n-grams are stored in
Page 23 and 24: main (static) n-gram model. As the
Page 25 and 26: There are many popular examples sho
Page 27 and 28: y Chen et al., who proposed a so-ca
Page 29 and 30: confusion among researchers, and ma
Page 31 and 32: language model took almost a week u
Page 33 and 34: w(t) s(t-1) s(t) U V W y(t) Figure
Page 35 and 36: ate is halved at start of every new
Page 37 and 38: or using matrix-vector notation as
Page 39 and 40: information for more than 5 time st
Page 41 and 42: A simple solution to the exploding
Page 43 and 44: output layer changes to computation
Page 45 and 46: While RNN models can overcome this
Page 47 and 48: complex or random architectures (su
Page 49 and 50: While for any of the previous point
Page 51 and 52: where λ is the interpolation weigh
Page 53 and 54: model with default SRILM cutoffs pr
Page 55 and 56: experiments, we have used the one i
Page 57 and 58: Perplexity (Penn corpus) 145 140 13
Page 59 and 60: with syntactical NNLMs would be pre
Page 61 and 62: Table 4.3: Combination of individua
Page 63 and 64: Table 4.6: Results on Penn Treebank
Page 65 and 66: 4.6 Conclusion of the Model Combina
Page 67 and 68: were: 400 classes, hidden layer siz
Page 69 and 70: Entropy per word on the WSJ test da
Page 71: Table 5.3: Results on the WSJ setup
Page 75 and 76: trained together with a maximum ent
Page 77 and 78: wt-3 wt-2 wt-1 D D D P(wt|context)
Page 79 and 80: n-gram probabilities. However, it w
Page 81 and 82: Table 6.1: Training corpora for NIS
Page 83 and 84: Perplexity 360 340 320 300 280 260
Page 85 and 86: Entropy per word 9 8.5 8 7.5 7 6.5
Page 87 and 88: 1 a a a 1 2 3 P(w(t)|*) ONE TWO THR
Page 89 and 90: Table 6.4: Perplexity on the evalua
Page 91 and 92: Entropy reduction per word over KN4
Page 93 and 94: Table 6.6: Perplexity with the new
Page 95 and 96: Entropy reduction over KN5 -0.04 -0
Page 97 and 98: as a baseline, and 12.3% after resc
Page 99 and 100: Table 7.1: BLEU on IWSLT 2005 Machi
Page 101 and 102: Table 7.3: Size of compressed text
Page 103 and 104: Table 7.4: Accuracy of different la
Page 105 and 106: Table 7.6: Entropy on PTB with n-gr
Page 107 and 108: 8.1 Machine Learning One possible d
Page 109 and 110: that almost every non-trivial compu
Page 111 and 112: supervision such as one digit at a
Page 113 and 114: Chapter 9 Conclusion and Future Wor
Page 115 and 116: from the expensive part of the mode
Page 117 and 118: Bibliography [1] A. Alexandrescu, K
Page 119 and 120: [23] D. Filimonov, M. Harper. A joi
Page 121 and 122: [50] T. Mikolov, S. Kombrink, L. Bu
Page 123 and 124:
[77] W. Wang, M. Harper. The SuperA
Page 125 and 126:
Test Phase After the model is train
Page 127 and 128:
• compute sentence-level scores g
Page 129 and 130:
Appendix B: Data generated from mod
Page 131 and 132:
Appendix C: Example of decoded utte
Page 133:
AND SAYS THIS SWING IS WITHIN A REA
show all

Statistical Language Models based on Neural Networks - Faculty of ...

Create successful ePaper yourself

Delete template?

Save as template?