Generalizing and Hybridizing Count-based and Neural Language Models

By Graham Neubig, Chris Dyer. 2016

33 citations recorded by the source corpus.

Language models (LMs) are statistical models that calculate probabilities over sequences of words or other discrete symbols. Currently two major paradigms for language modeling exist: count-based n-gram models, which have advantages of scalability and test-time speed, and neural LMs, which often achieve superior modeling performance. We demonstrate how both varieties of models can be unified in a single modeling framework that defines a set of probability distributions over the vocabulary of words, and then dynamically calculates mixture weights over these distributions. This formulation allows us to create novel hybrid models that combine the desirable features of count-based and neural LMs, and experiments demonstrate the advantages of these approaches.

Subfields: Artificial Intelligence, Large Language Models, Natural Language Processing

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. Graham Neubig
2. Mohit Bansal
3. Yoav Goldberg
4. Noah A. Smith
5. Luke S. Zettlemoyer
6. Yejin Choi
7. Kyunghyun Cho
8. Jianfeng Gao

In Large Language Models

1. Yoav Goldberg
2. Arthur Szlam
3. Edouard Grave
4. Anton Bakhtin
5. Marc'Aurelio Ranzato
6. Yangfeng Ji
7. Trevor Cohn
8. Yusuke Oda

In Natural Language Processing

1. Graham Neubig
2. Yoav Goldberg
3. Noah A. Smith
4. Chris Dyer
5. Kyunghyun Cho
6. Mohit Bansal
7. Luke S. Zettlemoyer
8. Yejin Choi

Citations in the corpus

• Lightweight Adaptive Mixture of Neural and N-gram Language Models (2018, methodology): “To compare our approach with the method proposed in Neubig and Dyer (2016), we trained a gating model using the last hidden layer of the neural language model as features (HIDDEN; 500 features for WSJ and 1024 for 1B Word).”
• DyNet: The Dynamic Neural Network Toolkit (2017, methodology): “Language Modeling: DyNet has been used in the development of hybrid neural/ n -gram language models [47] ( https://github.com/neubig/modlm ), and generative syntactic language models [18] ( https://github.com/clab/rnng ).”
• Latent Relation Language Models (2019, background): “…that while neural LMs greatly out-perform alternatives such as n -gram models on frequent words, they often under-perform on these rare words due to their limited parameter budget, which puts them at a disadvantage compared to non-parametric models like standard n -grams (Neubig and Dyer, 2016).”