Standard language model training employs gold human documents or human-human interaction data, and treats all training data as positive examples. Growing evidence shows that even with very large amounts of positive training data, issues remain that can be alleviated with relatively small amounts of negative data – examples of what the model should not do. In this work, we propose a novel procedure to train with such data called the “CRINGE” loss (ContRastive Iterative Negative GEneration). We show the effectiveness of this approach across three different experiments on the tasks of safe generation, contradiction avoidance, and open-domain dialogue. Our models outperform multiple strong baselines and are conceptually simple, easy to train and implement.

Current language models achieve low perplexity but their resulting generations still suffer from toxic responses, repetitiveness, and contradictions. The standard language modeling setup fails to address these issues. In this paper, we introduce a new architecture, Director, that consists of a unified generator-classifier with both a language modeling and a classification head for each output token. Training is conducted jointly using both standard language modeling data, and data labeled with desirable and undesirable sequences. Experiments in several settings show that the model has competitive training and decoding speed compared to standard language models while yielding superior results, avoiding undesirable behaviors while maintaining generation quality. It also outperforms existing model guiding approaches in terms of both accuracy and efficiency. Our code is made publicly available.

We propose a novel self-attention mechanism that can learn its optimal attention span. This allows us to extend significantly the maximum context size used in Transformer, while maintaining control over their memory footprint and computational time. We show the effectiveness of our approach on the task of character level language modeling, where we achieve state-of-the-art performances on text8 and enwiki8 by using a maximum context of 8k characters.

In this paper, we study the problem of hybrid language modeling, that is using models which can predict both characters and larger units such as character ngrams or words. Using such models, multiple potential segmentations usually exist for a given string, for example one using words and one using characters only. Thus, the probability of a string is the sum of the probabilities of all the possible segmentations. Here, we show how it is possible to marginalize over the segmentations efficiently, in order to compute the true probability of a sequence. We apply our technique on three datasets, comprising seven languages, showing improvements over a strong character level language model.