A Batch Noise Contrastive Estimation Approach for Training Large Vocabulary Language Models

By Youssef Oualil, Dietrich Klakow. 2017

6 citations recorded by the source corpus.

Training large vocabulary Neural Network Language Models (NNLMs) is a difficult task due to the explicit requirement of the output layer normalization, which typically involves the evaluation of the full softmax function over the complete vocabulary. This paper proposes a Batch Noise Contrastive Estimation (B-NCE) approach to alleviate this problem. This is achieved by reducing the vocabulary, at each time step, to the target words in the batch and then replacing the softmax by the noise contrastive estimation approach, where these words play the role of targets and noise samples at the same time. In doing so, the proposed approach can be fully formulated and implemented using optimal dense matrix operations. Applying B-NCE to train different NNLMs on the Large Text Compression Benchmark (LTCB) and the One Billion Word Benchmark (OBWB) shows a significant reduction of the training time with no noticeable degradation of the models performance. This paper also presents a new baseline comparative study of different standard NNLMs on the large OBWB on a single Titan-X GPU.

Subfields: Artificial Intelligence, Large Language Models, Optimization & Training

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Erik Nijkamp
2. Ying Nian Wu
3. Song-Chun Zhu
4. Ruiqi Gao
5. Zhijian Ou
6. Quoc V. Le
7. Andrew M. Dai
8. Myle Ott

In Large Language Models

1. Zhijian Ou
2. Diederik P. Kingma
3. Erik Nijkamp
4. Andrew M. Dai
5. Arthur Szlam
6. Anton Bakhtin
7. Yuntian Deng
8. Myle Ott

In Optimization & Training

1. Zhijian Ou
2. Ying Nian Wu
3. Ruiqi Gao
4. Andrew M. Dai
5. Erik Nijkamp
6. Zhen Xu
7. Diederik P. Kingma
8. Yoshua Bengio

Citations in the corpus

• Flow Contrastive Estimation of Energy-Based Models (2019, background): “NCE and it variants has gained popularity in natural language processing (NLP) [23, 57, 2, 4].”
• Learning Neural Trans-Dimensional Random Field Language Models with Noise-Contrastive Estimation (2017, methodology): “Then [12] further omits the sampling process by taking the other target words in current mini-batch as the noise samples.”
• Improved Training Of Neural Trans-Dimensional Random field Language Models with Dynamic Noise-Contrastive Estimation (2018, background): “Compared with using hundreds of noise samples per data sample in [16], DNCE uses at most 4 noise samples per data sample in our experiments.”