Clever Geek Handbook

Language Modeling with Guided Convolutional Networks

The dominant approach to language modeling today is based on recurrent neural networks. Their success in modeling is often related to the ability of such networks to handle unbounded context. In this article, we develop a finite context approach using stacked (composite) convolutions, which can be more efficient as they allow sequential chunks of data to be parallelized. We propose a new simplified neuro-guided mechanism that is superior to that proposed by Oord et al. (2016b) [26]and investigate the impact of key architectural decisions for it. The proposed approach achieves the most significant results on the WikiText103 benchmark, even though it is characterized by long-term dependencies, as well as comparable results on the Google Billion Words benchmark. Our model reduces the latency in evaluating a proposal by an order of magnitude, compared to recurrent baseline values. As far as we know, this is the first time that the non-periodic approach is competitive with strong recurrent models in such large-scale language problems.





1. Introduction

Statistical language models estimate the probability distribution of a sequence of words by modeling the probability of the next word given the previous words, i.e.





wi - . (Yu & Deng, 2014)[34] (Koehn, 2010) [17].





, (Bengio et al., 2003 [1]; Mikolov et al., 2010 [2]; Jozefowicz et al., 2016 [14]) n- (Kneser & Ney, 1995 [16]; Chen & Goodman, 1996 [3]). , , , . , . (LSTM; Hochreiter et al., 1997[12]), .





. (LeCun & Bengio, 1995 [19]). ,





\mathcal{O} (N / k )

N k. ,





\mathcal {O} (N)

, , , , (Manning & Schutze, 1999 [20]; Steedman, 2002 [31]). , , (Glorot & Bengio, 2010 [6]).





. ( ) , . , , ( 2).





, (Jozefowicz et al., 2016 [14]). (GLU) , ( 5.2).





, , , , LSTM, Google Billion Word Benchmark (Chelba et al., 2013 [2]). WikiText-103, , , (Merity et al., 2016 [21]). , , (GLU) , LSTM- Oord et al. (2016 [26]; 4, 5).





2.

, , , . (Bengio et al., 2003 [1])





H = [h_0,. . . , h_N]





w_0,. . . , w_N,





P (w_i | h_i).

f H





h_i = f (h_{i βˆ’ 1}, w_{i βˆ’ 1}).

, i ( ).









f { } H = f βˆ— w

, , , . . , , , , ( 5).





1 . ,





D^{ | V | Γ— e}, | V | - ( ), e - .





w_0,. . . , w_N,





E = [D_{w0},. . . , D_{wN}].





h_0,. . . , h_L





m, n – , , k - ,





X \in \mathbb {R} ^{N \times m} - h_l

( , ),





W \in \mathbb {R}^{k \times m \times n}, b \in \mathbb {R}^n, V \in \mathbb {R}^{k\times m\times n}, c \in \mathbb {R}^n - , \sigma \otimes .

,





h_i

. , (Oord et al., 2016a [25]). ,





k-1,

, - , ,





k - .
1. .
1. .





X βˆ— W + b, Οƒ (X βˆ— V + c).

LSTM,





X βˆ— W + b

, . (GLU). E





H = h_Lβ—¦. . .β—¦h_0 (E).

(GLU) , (He et al., 2015a [10]). , 5 ().





- softmax, , , (Gutmann & Hyvarinen [9]) softmax (Morin & Bengio, 2005 [24]). , softmax. ( , ) – (Grave et al., 2016a [7]). – , .





3.

, , (Hochreiter & Schmidhuber, 1997 [12]). LSTMs , () . . . , , , () .





, , , . , , , . Oord et al. (2016b [26]) LSTM





tanh (X βˆ— W + b) βŠ—Οƒ (X βˆ— V + c)

. Kalchbrenner et al. (2016 [15]) .





(GLU) - , Dauphin & Grangier (2015) [35] , . , . LSTM, gated tanh unit (GTU),





, ( , ) -





tanh ' (X) Οƒ ' (X).

, (GLU)









βˆ‡X βŠ— Οƒ(X)

Οƒ (X). , . Β§5.2 , (GLU) .





4.

4.1





. -, Google Billion Word (Chelba et al., 2013 [2]) , , 800 . . , 3 , . 30 301 028 , . -, WikiText-103 - , 100 . , 200 . (Merity et al., 2016 [21]). GBW, , , . <S> </S> . Google Billion Word , WikiText-103 . <S> </ S> , </S>. ,





e^{ {\frac1N}\sum_{i}^N βˆ’ \log p(w_i|...,w_{iβˆ’1})}

.









4.2





Torch (Collobert et al., 2011 [5]) Tesla M40. , . 8 , , 1/8 . Nvidia NCCL. .





, (Sutskever et al., 2013 [32]). , . (Pascanu et al., 2013 [27]) (Salimans & Kingma, 2016 [28]).





1. . [k, n]. Β«BΒ» .
1. . [k, n]. Β«BΒ» .

. (2013) [27] , , RNN. RNN, .





, . , , 1.





4.3





. {1,. . . , 10}, {128,. . . , 256}, {128,. . . , 2048}, - {3,. . . , 5}. , , , , . , ( ., 2015b [11]), [1., 2.], 0,99 0,1. , .





5.

LSTM . LSTM RNN .





, GCNN LSTM Google . , softmax (Grave et al., 2016a [7]), . GCNN 38,1 , LSTM 39,8 ( 2).





2. Google Billion Word. GCNN LSTM .
2. Google Billion Word. GCNN LSTM .

, GCNN . 2 , , softmax softmax. softmax, GCNN . GCNN , LSTM Jozefowicz et al. (2016 [14]), , softmax. , , , 31,9 30,6 , 2 8 3 32 LSTM. , , (Shazeer et al., 2017 [30]), .





2. (Jozefowicz et al., 2016 [14]), softmax, softmax , .
2. (Jozefowicz et al., 2016 [14]), softmax, softmax , .

, GCNN . Google Billion Word - 20 . WikiText-103, , , . WikiText-103 , , 4000 . GCNN LSTM ( 3). GCNN-8 8 800 , LSTM - 1024 . , GCNN .





3. WikiText-103.
3. WikiText-103.

Gigaword Chen et al. (2016 [4]) . , , , 55,6 29,4. Penn tree. , GCNN LSTM : 108,7 109,3 . , , . LSTM, , GCNN , , .





5.1





. . , . , () . , () - , . , , , , . , . , , .





, 43,9 Google Billion Word. LSTM 2048 2, GCNN-8Bottleneck 7 Resnet, , (He et al., 2015a [10]), GCNN-8 . () k > 1 k = 1. k = 1 , . , .





4. . LSTM 2048 GCNN 43,9 Google Billion Word. GCNN 20 .
4. . LSTM 2048 GCNN 43,9 Google Billion Word. GCNN 20 .

LSTM 750 20, 15 000 . - 15 000 . 4 , LSTM GCNN . LSTM , 750 . , LSTM cuDNN, cuDNN , . , 1- cuDNN. LSTM, GCNN , , GCNN 20 .





3. WikiText-103 () Google Billion Word () . (GLU) .
3. WikiText-103 () Google Billion Word () . (GLU) .

5.2 ()





, . (GTU) LSTM





tanh (X βˆ— W + b) βŠ— Οƒ (X βˆ— V + c)

(Oord et al., 2016b [17]) , ReLU Tanh. , . 3 () , GLU , WikiText-103. , ReLU , . ReLU, GLU. , Tanh, GTU , , . GTU , , .





GTU Tanh , Tanh GTU . ( 3, ) , , GTU Tanh. , ReLU GLU,





ReLU (X) = X βŠ— (X> 0)

, . GLU .





3 () Google Billion Words. 100 - , . WikiText-103, . 5 GLU ReLU, LSTM RNN, (Jozefowicz et al., 2016 [14]) .





5.3





, , , . GLU , , GLU. (Manning & Schutze, 1999 [20]). , GLU





h_l(X) = X βˆ— W + b

- , softmax, -. GLU - (Mnih & Hinton, 2007 [23]),





h_l(X) = (X βˆ— W + b) βŠ— (X βˆ— V + c).
4. Google Billion Word () Wiki-103 (). , , 20.
4. Google Billion Word () Wiki-103 (). , , 20.

5 , GLU, , . 40 , GLU 20 . (115) 67,6 5- -, . , , 61 Google Billion Word, 5- -, (Ji et al., 2015 [13]).





5.4





Figure 5. Learning curves in Google Billion Word for models with varying degrees of nonlinearity.
5. Google Billion Word .

4 CNN. . , , , 40 , WikiText-103, . , , , . , , 40 . 4 , WikiText-103 , Google Billion Word, . WikiText-103 , Google Billion Word, 20. , 4000 , , 30 .





5.5





. , . - . 6 , . , . (1 0,01), . , ,   , .





Figure 6. Impact of weight normalization and gradient clipping on Google Billion Word.
6. Google Billion Word.

6.

. , , , () . , , , . , WikiText-103. Google Billion Word , .





Spoiler

  1. Bengio, Yoshua, Ducharme, Rejean, Vincent, Pascal, and Jauvin, Christian. A neural probabilistic language model. journal of machine learning research, 3(Feb):1137–1155, 2003.





  2. Chelba, Ciprian, Mikolov, Tomas, Schuster, Mike, Ge, Qi, Brants, Thorsten, Koehn, Phillipp, and Robinson, Tony. One billion word benchmark for measuring progress in statistical language modeling. arXiv preprint arXiv:1312.3005, 2013.





  3. Chen, Stanley F and Goodman, Joshua. An empirical study of smoothing techniques for language modeling. In Proceedings of the 34th annual meeting on Association for Computational Linguistics, pp. 310–318. Association for Computational Linguistics, 1996.





  4. Chen, Wenlin, Grangier, David, and Auli, Michael. Strategies for training large vocabulary neural language models. CoRR, abs/1512.04906, 2016.





  5. Collobert, Ronan, Kavukcuoglu, Koray, and Farabet, Clement. Torch7: A Matlab-like Environment for Machine Learning. In BigLearn, NIPS Workshop, 2011. URL http://torch.ch.





  6. Glorot, Xavier and Bengio, Yoshua. Understanding the difficulty of training deep feedforward neural networks. The handbook of brain theory and neural networks, 2010.





  7. Grave, E., Joulin, A., Cisse, M., Grangier, D., and Jegou, H. Efficient softmax approximation for GPUs. ArXiv e-prints, September 2016a.





  8. Grave, E., Joulin, A., and Usunier, N. Improving Neural Language Models with a Continuous Cache. ArXiv e-prints, December 2016b.





  9. Gutmann, Michael and Hyvarinen, Aapo. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models.





  10. He, Kaiming, Zhang, Xiangyu, Ren, Shaoqing, and Sun, Jian. Deep residual learning for image recognition. arXiv preprint arXiv:1512.03385, 2015a.





  11. He, Kaiming, Zhang, Xiangyu, Ren, Shaoqing, and Sun, Jian. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE International Conference on Computer Vision, pp. 1026–1034, 2015b.





  12. Hochreiter, Sepp and Schmidhuber, Jurgen. Long short-term memory. Neural computation, 9(8):1735–1780, 1997.





  13. Ji, Shihao, Vishwanathan, SVN, Satish, Nadathur, Anderson, Michael J, and Dubey, Pradeep. Blackout: Speeding up recurrent neural network language models with very large vocabularies. arXiv preprint arXiv:1511.06909, 2015.





  14. Jozefowicz, Rafal, Vinyals, Oriol, Schuster, Mike, Shazeer, Noam, and Wu, Yonghui. Exploring the limits of language modeling. arXiv preprint arXiv:1602.02410, 2016.





  15. Kalchbrenner, Nal, Espeholt, Lasse, Simonyan, Karen, van den Oord, Aaron, Graves, Alex, and Kavukcuoglu, Koray. Neural Machine Translation in Linear Time. arXiv, 2016.





  16. Kneser, Reinhard and Ney, Hermann. Improved backing-off for m-gram language modeling. In Acoustics, Speech, and Signal Processing, 1995. ICASSP-95., 1995 International Conference on, volume 1, pp. 181–184. IEEE, 1995.





  17. Koehn, Philipp. Statistical Machine Translation. Cambridge University Press, New York, NY, USA, 1st edition, 2010. ISBN 0521874157, 9780521874151.





  18. Kuchaiev, Oleksii and Ginsburg, Boris. Factorization tricks for LSTM networks. CoRR, abs/1703.10722, 2017. URL http://arxiv.org/abs/1703.10722.





  19. LeCun, Yann and Bengio, Yoshua. Convolutional networks for images, speech, and time series. The handbook of brain theory and neural networks, 3361(10):1995, 1995.





  20. Manning, Christopher D and Schutze, Hinrich. Foundations of statistical natural language processing, 1999.





  21. Merity, S., Xiong, C., Bradbury, J., and Socher, R. Pointer Sentinel Mixture Models. ArXiv e-prints, September 2016.





  22. Mikolov, TomaΒ΄s, Martin, Karafiat, Burget, Lukas, Cernocky, Jan, and Khudanpur, Sanjeev. Recurrent Neural Network based Language Model. In Proc. of INTERSPEECH, pp. 1045–1048





  23. Mnih, Andriy and Hinton, Geoffrey. Three new graphical models for statistical language modelling. In Proceedings of the 24th international conference on Machine learning, pp. 641–648. ACM, 2007.





  24. Morin, Frederic and Bengio, Yoshua. Hierarchical probabilistic neural network language model. In Aistats, volume 5, pp. 246–252. Citeseer, 2005.





  25. Oord, Aaron van den, Kalchbrenner, Nal, and Kavukcuoglu, Koray. Pixel recurrent neural networks. arXiv preprint arXiv:1601.06759, 2016a.





  26. Oord, Aaron van den, Kalchbrenner, Nal, Vinyals, Oriol, Espeholt, Lasse, Graves, Alex, and Kavukcuoglu, Koray. Conditional image generation with pixelcnn decoders. arXiv preprint arXiv:1606.05328, 2016b.





  27. Pascanu, Razvan, Mikolov, Tomas, and Bengio, Yoshua. On the difficulty of training recurrent neural networks. In Proceedings of The 30th International Conference on Machine Learning, pp. 1310–1318, 2013.





  28. Salimans, Tim and Kingma, Diederik P. Weight normalization: A simple reparameterization to accelerate training of deep neural networks. arXiv preprint arXiv:1602.07868, 2016.





  29. Shazeer, Noam, Pelemans, Joris, and Chelba, Ciprian. Skip-gram language modeling using sparse non-negative matrix probability estimation. arXiv preprint arXiv:1412.1454, 2014.





  30. Shazeer, Noam, Mirhoseini, Azalia, Maziarz, Krzysztof, Davis, Andy, Le, Quoc V., Hinton, Geoffrey E., and Dean, Jeff. Outrageously large neural networks: The sparsely-gated mixtureof-experts layer. CoRR, abs/1701.06538, 2017. URL http://arxiv.org/abs/1701.06538.





  31. Steedman, Mark. The syntactic process. 2002.





  32. Sutskever, Ilya, Martens, James, Dahl, George E, and Hinton, Geoffrey E. On the importance of initialization and momentum in deep learning. 2013.





  33. Wang, Mingxuan, Lu, Zhengdong, Li, Hang, Jiang, Wenbin, and Liu, Qun. gencnn: A convolutional architecture for word sequence prediction. CoRR, abs / 1503.05034, 2015. URL http://arxiv.org/abs/1503.05034 .





  34. Yu, Dong and Deng, Li. Automatic Speech Recognition: A Deep Learning Approach. Springer Publishing Company, Incorporated, 2014. ISBN 1447157788, 9781447157786.





  35. Dauphin, Yann N and Grangier, David. Predicting distributions with linearizing belief networks. arXiv preprint arXiv: 1511.05622, 2015.















More articles:


All Articles