Skip to main content

RNN language model with word clustering and class-based output layer


The recurrent neural network language model (RNNLM) has shown significant promise for statistical language modeling. In this work, a new class-based output layer method is introduced to further improve the RNNLM. In this method, word class information is incorporated into the output layer by utilizing the Brown clustering algorithm to estimate a class-based language model. Experimental results show that the new output layer with word clustering not only improves the convergence obviously but also reduces the perplexity and word error rate in large vocabulary continuous speech recognition.

1 Introduction

Statistical language models estimate the probability of a word occurring in a given context, which plays an important role in many natural language processing applications such as speech recognition, machine translation, and information retrieval. Standard n-gram back-off language models (LMs) are widely used for their simplicity and efficiency. However, in this approach, words are modeled as discrete symbols with richer linguistic information, such as syntax and semantic, ignored completely. Additionally, large numbers of parameters need to be estimated, and due to the sparsity characteristics of natural language, the probability of low- and zero-frequency events is estimated crudely and inaccurately using various smoothing algorithms.

The distributional hypothesis in linguistics states that words occurring in the same context tend to have similar meanings. It is a reasonable assumption that similar words occur in the same context with similar probability, for example, ‘America,’ ‘China,’ and ‘Japan’ which usually come after the same preposition or as the subject of a sentence. Based on this assumption, neural network language models (NNLMs)[1, 2] project the discrete word indices into a continuous space where similar words occur close together. The predicted probability of the next word is returned by a smooth function of the context representation, which alleviates the sparsity issue to some extent and leads to better generalization for unseen n-grams. In 2010, Elman’s recurrent neural network was first used for language modeling by Mikolov[3] and then an extension of this model was proposed in 2011[4, 5]. The recurrent neural network language model (RNNLM) has a longer memory and has recently performed better than other modeling methods[3, 4, 6]. Accordingly, we select the RNNLM as our baseline approach in this paper.

One key issue is the heavy computational cost for the RNNLM. As the output layer contains one unit for each word in the vocabulary, it is infeasible to train the model for large vocabulary with hundreds of thousands of words. Therefore, reducing the complexity of neural network language models has been an important topic. Perhaps one method is to estimate the several thousand most frequent words (the shortlist) via NNLMs, while other words are estimated via n-gram back-off models. Unfortunately, it has been shown that this technique causes severe degradation of performance for a small shortlist[1]. Other tree-structured output layer methods have also been proposed to speed up the NNLMs[7, 8]. In these methods, the tree structure of the output layer needs to be constructed carefully using linguistic knowledge such as WordNet[9] or word continuous representation. In general, speed and performance need to be balanced so that training and testing process is accelerated as much as possible, without deteriorating the performance of the model.

In this paper, we introduce a new method for constructing a class-based output layer using the Brown clustering algorithm. The closest previous work to this is a simple frequency-based word factorizing algorithm used to construct the output layer[4]. Words are roughly clustered according to their frequencies in this method, with training speed increasing but performance degraded. We extend this work to improve the performance of RNNLM and speed up the training. Words are clustered off-line and then the word classes are embedded into the output layer to estimate the class-based language model, where the RNN is used to estimate the conditional probability of classes and words.

This paper is organized as follows: In Section 2, we introduce our baseline RNNLM and the proposed Brown clustering method for constructing the output layer. Perplexity evaluation on a public corpus is performed in Section 3. Our proposed model is further evaluated on the Wall Street Journal (WSJ) and Switchboard speech-to-text tasks in Sections 4 and 5. Finally, Section 6 concludes this paper and gives the future work.

2 Model description

2.1 RNN language model

An Elman recurrent neural network (RNN)[10] is shown in Figure1. The hidden state is a function of the entire input history. RNNs are well known for their long memory and are widely used for dynamic system modeling and sequence prediction. Let V denotes the vocabulary with ‘1-of- |V|’ coding used in the input layer, so that the i th word of the vocabulary is encoded as a binary |V|-dim vector, where the i th element is set as 1 and all others are 0. Let h t  = sigmoid(W ih x t  + W hh ht−1), where x t and h t denote the input and the hidden activation at the current time step, respectively. The hidden state h t is activated by the current input x t and the previous hidden activation ht−1. DefineP( w t | w 1 t 1 )= o t =softmax( W ho h t ), where w t and w 1 t 1 denote the next word and the context. The output layer o t corresponds to the predicted probability of all words in the vocabulary. To speed up the training of RNNLM, a frequency-based extension of RNNLM is introduced in[4], which is a class-based model as shown in Figure2. LetP( w t | w 1 t 1 )=P(C( w t )| w 1 t 1 )P( w t |C( w t ), w 1 t 1 ), where C(w t ) denotes the class of the word w t . The predicted probabilities of classes and specific words are estimated to decrease the computational complexity.

Figure 1
figure 1

Recurrent neural network language model.

Figure 2
figure 2

Class-based output layer for RNNLM.

To speed this up, a simple frequency-based factorizing method is used to construct the equivalence class of words in the class-based model. Compared with RNNLM without a class-based layer, the perplexity (PPL) of the model shown in Table1 is 10% higher for the Penn Treebank Corpus (one million words). Details can be found in Section 3.

Table 1 Perplexities on test set of Penn Treebank Corpus

2.2 Word clustering for output layer

2.2.1 Frequency-based clustering

Frequency-based word clustering is referred to as the frequency binning factorization method[4], where words are assigned to classes proportionally. Figure3 gives the unigram cumulative probability distribution for the Penn Treebank Corpus, which describes the well-known phenomenon of Zipf’s law in natural language. This method divides the cumulative probability into K partitions to form K frequency binnings which correspond to K clusters, a very rough word partition which only considers frequency. Zipf’s law states that given some corpus of natural language utterances, the frequency of any word is inversely proportional to its rank in the frequency table. This means that the very few high-frequency words occupy most of the text corpus. It can be seen in Figure3 that the most frequent 150 words and 2,800 words occupy more than 50% and 80% of the text corpus, respectively. In order to have 100 equal clusters, the top 150 words make up the first 50 clusters and the top 2,800 words make up the first 80 clusters, which means the last 7,200 words form 20 clusters. Therefore, clusters containing high-frequency words are very small, possibly even containing one word. In contrast, most words are in the remaining clusters, each containing hundreds or even thousands of words. Thus, the clustering results of this method depend severely on frequency distribution of the training corpus, leading to unsatisfactory clustering results.

Figure 3
figure 3

Cumulative unigram probability distribution for Penn Treebank Corpus with about one million words (Zipf’s law).

2.2.2 Brown clustering

Brown clustering is a data-driven hierarchical word clustering algorithm[11, 12], which is widely used in natural language processing. The input to this algorithm is text, which is a sequence of words w1,w2,...,w n , and the output of the clustering algorithm is a binary tree, the leaves of which are words. In this paper, we interpret all leaves with the same parent node as a cluster in the tree. The Brown clustering algorithm was first proposed to estimate a class-based language model[11]. Let V, C, and T denote the vocabulary, the word clusters and the text corpus, respectively. The optimization object is the cross entropy of the text corpus:

loss ( C ) = 1 | T | log P ( w 1 ... w | T | ) = 1 | T | log i = 1 | T | P ( C ( w i ) | C ( w i 1 ) P ( w i | C ( w i ) ) )

where |T| denotes the length of text and C(·) maps the words to the specific clusters. P(C(w i )|C(wi−1)) and P(w i |C(w i )) can be estimated by frequency. Therefore, the object loss function can be rewritten as follows:

loss ( C ) = 1 | T | i = 1 | T | log n c i , c i 1 n c i 1 n w i n c i = c , c C P ( c , c ) log P ( c , c ) P ( c ) P ( c ) w V P ( w ) log P ( w ) ,

where n w i denotes the occurrence count of pattern w i that occurs in the corpus.

Initially, the algorithm starts with each word in its own cluster. As long as there are more than one cluster left, the algorithm merges the two clusters that minimizes the loss of the clustering result as shown in Equation 2. The naive algorithm has time complexity O(|V|3)[11] and is impractical for hundreds of thousands of words. Fortunately, a variant algorithm with time complexity O(|V|K2 + |T|) was proposed in[13], where K denotes the number of clusters. The algorithm is described as follows with details available in[13].

Input: text corpus T, the number of clusters K

Output: K word clusters

  •  Take the K most frequent words, put each into its own cluster c1,c2,...,c K

  •  For i = K + 1 : |V|

  1. 1.

    Create a new cluster c K+1 for the i th most frequent word, giving K + 1 clusters.

  2. 2.

    Choose two clusters from c 1,c 2,...,c K+1 to merge, selecting these that minimize Equation 2.

  •  Implement |V| − K merges to create a full hierarchical word clusters.

In this paper, we partition the words into clusters using this algorithm and demonstrate its effectiveness for RNN language modeling in terms of both perplexity and word error rate.

3 Penn Treebank Corpus evaluation

The proposed language model is evaluated on the Wall Street Journal portion of the Penn Treebank which is preprocessed by lowercasing words, removing punctuation and replacing numbers with the ‘N’ symbol. Sections 00-20 (930K words) are used as training sets, sections 21-22 as validation sets (74K words), and sections 23-24 as test sets (82K words). The vocabulary size is 10K, including a special token for unknown words.

We compare the proposed model with the baseline RNNLM model[5, 6]. We denote our proposed model as RNNLM-Brown and the baseline as RNNLM-Freq for convenience, where the ‘Brown’ and ‘Freq’ mean the Brown clustering and the frequency-based clustering, respectively. Both models have the same basic configuration (200 hidden units) for comparisons in the following experiments. The truncated backpropagation through time algorithm (BPTT) is used for training the RNNLMs using ten time steps. When the perplexity decreases very slowly or increases, the learning rate is halved. The basic 5-gram back-off language model (LM-KN5) is trained with the modified Kneser-Ney smoothing algorithm. Figure4 shows the convergence process of the validation set’s perplexity for RNNLM-Freq and RNNLM-Brown, where RNNLM-Brown with 13 epochs obtains the same perplexity as RNNLM-Freq with 24 epochs. We can see that the proposed RNNLM-Brown converges twice faster and obtains lower perplexity on the validation set. Accordingly, appropriate word clustering for the output layer can speed up the convergence, which is especially important for a large training corpus.

Figure 4
figure 4

PPL convergence of RNNLM-Freq/Brown on validation sets.

In the following experiments, perplexity and training speed are evaluated with different sizes of class layers, as shown in Table2. The first two columns refer to the baseline (RNNLM-Freq) and the interpolated model with LM-KN5 (RNNLM-Freq + KN5), respectively, which is consistent with the results reported in[4]. The last two columns correspond to the proposed language model (RNNLM-Brown) and the interpolated version with LM-KN5 (RNNLM-Brown + KN5). The full model uses the entire vocabulary for the class layer, in which each word corresponds to a separate cluster. The interpolated coefficients are determined according to the validation set. To make the observation easier, we plot the test perplexity and the speed of training in Figure5, where the blue and red lines correspond to the left y-axis for perplexity and the green lines correspond to the right y-axis for the training speed. The training speed is evaluated by the number of words processed per second on a machine with an Intel® Core™2 Quad CPU Q9400 at 2.66 GHz, 8-GB RAM. In practice, the training speed first increases and then decreases with the increasing number of clusters. There is a trade-off between the perplexity and the training speed, especially for a larger corpus. From Figure5, 100 or 200 clusters are the best choices balancing this, and the training speed is 15 times faster than that of the full model. Empirically, the best number of clusters is around | V | , where |V| denotes the size of vocabulary. In the experiments, the perplexity is reduced by about 5% for both single and interpolated models without reducing the speed of training, compared with the baseline. Moreover, the performance of the proposed RNNLM-Brown is much closer to that of the full model without the class layer.

Table 2 Comparisons of perplexities on test set of Penn Treebank Corpus with different sizes of class layer
Figure 5
figure 5

Detailed comparisons of perplexity and speed for different RNNLMs on Penn Treeback Corpus.

4 WSJ speech recognition experiment

To evaluate the performance of the proposed language model for speech recognition, we use the WSJ task which is a standard task for language model evaluation. The acoustic model is a 6,000-tied-state continuous model with 32 Gaussians, trained on the WSJ1 Corpus; the details of which can be found in[14]. The LM training text contains 26 million words from the entire NYT-1994 section of the English Gigaword Corpus. The top 64,000 most frequent words are selected as the vocabulary and other words are mapped to a special token. Trigram (LM-KN3) and 5-gram (LM-KN5) models are trained on the text using the MITLM toolkit[15] with the modified Kneser-Ney smoothing algorithm for decoding and rescoring.

In the experiment, we use the NIST 1993 CSR Hub and Spoke Benchmark Test Corpora[16] as our test bed. We select the hub 1 and spoke 1, 2 and 4 sections for evaluations, which contain 1,251 utterances (about 25K words, 2.5 h of voice data) All the utterances are divided equally into two parts as the development set and the evaluation set.

Due to the complexity of training the neural network language model, this requires several days or an even longer time period to converge for several million training words. Thus, we randomly select about five million words from the NYT-1994 section of English Gigaword Corpus as the training data, 500K words as the validation data for early stopping and 500K words as the test data for perplexity evaluation. The top 30K frequent words are selected as the vocabulary. The truncated BPTT is used for training the different RNNLMs with ten time steps. Two hundred hidden units are used. The learning rate is initially set to 0.1 and halved when the perplexity of the validation data is increased. The detailed results are given in Table3, where consistent improvements are observed with different class sizes for the larger training corpus. The perplexity is reduced by approximately 5%, compared with the baseline.

Table 3 Comparisons of perplexities on test set of five million training corpus with different sizes of class layer

In the following experiments, the 200-best hypotheses are generated using back-off trigram trained on the entire 26 million words and then rescored with different language models for comparisons. The interpolated coefficient of RNNLM and LM-KN5 is determined on the development set. Detailed results can be found in Tables4 and5. We can see that the word error rate (WER) for evaluation set is consistently reduced by 0.3% to 0.7% absolutely, compared with that of RNNLM-Freq, and 1.4% to 1.7% compared with the 1-best hypothesis (i.e. more than 10% relative reduction of WER).

Table 4 WER for development set rescored with different RNNLMs and LM-KN5
Table 5 WER for evaluation set rescored with different RNNLMs and LM-KN5

5 Switchboard speech recognition experiment

In this section, the effectiveness of our proposed model on the task of speech-to-text transcription is evaluated on the 309h Switchboard-I training set[17], a larger corpus than WSJ Corpus. The system uses 13-dimensional PLP features with rolling-window mean-variance normalization and up to third-order derivatives, reduced to 39 dimensions by HLDA. The speaker-independent three-state cross-word triphones share 9,308 CART-tied states. The GMM-HMM baseline system has 40-Gaussian mixtures per state, trained with maximum likelihood and refined discriminatively with the boosted maximum-mutual-information criterion.

The data for system development is the 1831-segment SWB part of the NIST 2000 Hub5 evaluation set (Hub5’00-SWB). The FSH half of the 6.3 h Spring 2003 NIST Rich Transcription set (RT03S-FSH) acts as the evaluation set. Based on Kneser-Ney smoothing, a back-off trigram language model (LM-KN3) was trained on the 2000h Fisher transcripts containing 20 million tokens for decoding, where the vocabulary is limited to 53K words and unknown words are mapped into a special token <unk>. Note that no other unkown text is used to train LMs for interpolations so that the following experimental results are easily repeatable. The pronouncing dictionary comes from the CMU pronouncing dictionary[18].

Two models (RNNLM-Freq/Brown) with 300 hidden units are trained on the entire training text for comparisons. The perplexity convergence of RNNLM-Freq and RNNLM-Brown on Hub5’00-SWB set is shown in Figure6. It can be seen that waiting long enough, the RNNLM with brown clustering converges better than that with frequency partitions. Moreover, the perplexity of RNNLM-Brown with 9 epochs competes with that of RNNLM-Freq with 16 epochs; it means that RNNLM-Brown converges twice faster than RNNLM-Freq.

Figure 6
figure 6

PPL convergence of RNNLM-Freq/Brown on Hub5’00-SWB set.

Subsequently, these models are further compared to rescore N-best hypotheses. For convenience, 100-best hypotheses are generated from Hub5’00-SWB and RT03S-FSH and rescored by different LMs. The interpolation weights are tuned on Hub5’00-SWB, and the performances of these LMs are evaluated on RT03S-FSH. These intermediate models during the training are used for rescoring, and the performance of these models in word error rate is plotted in Figure7 for comparisons, where the RNNLM-Brown converges much faster and better than RNNLM-Freq. We can see that RNNLM-Brown with 9 epochs obtains the same WER as RNNLM-Freq with 16 epochs.

Figure 7
figure 7

WER convergence of 100-best rescoring with RNNLM-Freq/Brown on Hub5’00-SWB set.

The performances of different LMs in perplexity and word error rate are shown in Table6, where Hub5’00-SWB and RT03S-FSH are used for validation and evaluation, respectively. We can see that the WER of our proposed model RNNLM-Brown interpolated with LM-KN5 obtains the lowest perplexity 51.24 and word error rate 25.3% on the evaluation set. In a word, our proposed RNNLM-Brown converges faster and better in the experiment.

Table 6 100-Best rescoring with different LMs on Hub5’00-SWB and RT03S-FSH

6 Conclusions

In this paper, the Brown word clustering algorithm is proposed to construct a class layer for the RNNLM. Experimental results show that our proposed RNNLM-Brown improves the perplexity and decreases the word error rate obviously. The performance of our proposed RNNLM-Brown is much closer to that of the full model without a class layer. Moreover, our proposed RNNLM-Brown converges twice faster than RNNLM-Freq, which is critical for a large-scale dataset. Additionally, we notice that the outputs of the brown clustering algorithm include a binary tree structure of clusters, which is not used in this paper. In future work, we will further investigate this tree-structured output layer according to the hierarchical word cluster results. Additionally, we will also investigate whether soft clustering of words[19] can be incorporated into the RNNLM to further improve its performance.



Automatic speech recognition


Backpropagation through time algorithm


Standard back-off n-gram language model


Neural network language model


Recurrent neural network language model


Freq-based word clustering


Brown word clustering


Kneser-Ney smoothing algorithm


5-gram based on Kneser-Ney smoothing algorithm




Word error rate


Wall Street Journal


  1. Balado F, Hurley NJ, Mccarthy EP, Silvestre GCM: Continuous space language models. Comput. Speech. Lang 2007, 21(3):492-518. 10.1016/j.csl.2006.09.003

    Article  Google Scholar 

  2. Son LH, Allauzen R, Wisniewski G, Yvon F: Training continuous space language models: some practical issues. Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP) Massachusetts, 9–11 October 2010, pp. 778–788

    Google Scholar 

  3. Tomas M, Martin K, Lukas B, Jan HG, Sanjeev K: Recurrent neural network based language model. Proceedings of the Annual Conference of International Speech Communication Association (INTERSPEECH) Chiba, 26–30 September 2010, pp. 1045–1048

    Google Scholar 

  4. Tomas M, Stefan K, Lukas B, Jan HG, Sanjeev K: Extensions of recurrent neural network language model. Proceedings of IEEE International Conference on Acoustic, Speech and Signal Processing (ICASSP) Prague, 22–27 May 2011

    Google Scholar 

  5. Tomas M, Anoop D, Stefan K, Lukas B, Jan HG: RNNLM - recurrent neural network language modeling toolkit. Proceedings of Automatic Speech Recognition and Understanding Workshop (ASRU) Waikoloa, 11–15 December 2011

    Google Scholar 

  6. Tomas M, Anoop D, Stefan K, Lukas B, Jan HG: Empirical evaluation and combination of advanced language modeling techniques. Proceedings of the Annual Conference of International Speech Communication Association (INTERSPEECH 2011) Florence, 27–31 August 2011

    Google Scholar 

  7. Morin F, Bengio Y: Hierarchical probabilistic neural network language model. Proceedings of AISTATS Christchurch, 6–8 January 2005, pp. 246–252

    Google Scholar 

  8. Le HS, Oparin I, Allauzen A, Gauvain JL, Yvon F: Structured Output Layer neural network language model. Proceedings of IEEE International Conference on Acoustic, Speech and Signal Processing (ICASSP) Prague, 22–27 May 2011, pp. 5524–5527

    Google Scholar 

  9. Fellbaum C: WordNet: An Electronic Lexical Database. Cambridge: MIT Press; 1998.

    MATH  Google Scholar 

  10. Elman JL: Finding structure in time. Cogn. Sci 1990, 14(2):179-211. 10.1207/s15516709cog1402_1

    Article  Google Scholar 

  11. Brown PF, deSouza PV, Mercer RL, Pietra VJD, Lai JC: Class-based n-gram models for natural language. Comput. Linguist 1992, 18(4):467-479.

    Google Scholar 

  12. Martin S, Liermann J, Ney H: Algorithms for bigram and trigram word clustering. Speech Commun 1998, 24: 1253-1256.

    Article  Google Scholar 

  13. Liang P: Semi-supervised learning for natural language processing. Master’s thesis, MIT, 2005

    Google Scholar 

  14. Vertanen K: Baseline WSJ acoustic models for HTK and Sphinx: training recipes and recognition experiments. 2006.

    Google Scholar 

  15. Hsu BJP: MIT language modeling toolkit. 2009.

    Google Scholar 

  16. 1993 ARPA CSR Hub and Spoke Benchmark Tests Corpora 1993.

  17. Godfrey J, Holliman E: Switchboard-1 Release. Philadelphia: Linguistic Data Consortium; 1997.

    Google Scholar 

  18. The CMU Pronouncing Dictionary Release 0.7a 2007.

  19. Su Y: Bayesian class-based language models. Proceedings of IEEE International Conference on Acoustic, Speech and Signal Processing (ICASSP) Prague, 22–27 May 2011

    Google Scholar 

Download references


This work was supported by the National Natural Science Foundation of China under grant nos. 61273268, 61005019 and 90920302, and in part by Beijing Natural Science Foundation Program under grant no. KZ201110005005.

Author information

Authors and Affiliations


Corresponding author

Correspondence to Yongzhe Shi.

Additional information

Competing interests

The authors declare that they have no competing interests.

Authors’ original submitted files for images

Rights and permissions

Open Access This article is distributed under the terms of the Creative Commons Attribution 2.0 International License (, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Reprints and permissions

About this article

Cite this article

Shi, Y., Zhang, WQ., Liu, J. et al. RNN language model with word clustering and class-based output layer. J AUDIO SPEECH MUSIC PROC. 2013, 22 (2013).

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: