Learning to Generate Reviews and Discovering Sentiment

By Alec Radford, Rafal Józefowicz, Ilya Sutskever. 2017

549 citations recorded by the source corpus.

We explore the properties of byte-level recurrent language models. When given sufficient amounts of capacity, training data, and compute time, the representations learned by these models include disentangled features corresponding to high-level concepts. Specifically, we find a single unit which performs sentiment analysis. These representations, learned in an unsupervised manner, achieve state of the art on the binary subset of the Stanford Sentiment Treebank. They are also very data efficient. When using only a handful of labeled examples, our approach matches the performance of strong baselines trained on full datasets. We also demonstrate the sentiment unit has a direct influence on the generative process of the model. Simply fixing its value to be positive or negative generates samples with the corresponding positive or negative sentiment.

Subfields: Artificial Intelligence, Natural Language Processing, Generative Models

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

In Natural Language Processing

In Generative Models

Citations in the corpus

Character-Level Language Modeling with Deeper Self-Attention (2018, background): “Character-level modeling has shown promise in many areas such as sentiment analysis [35], question answering [21] and classiﬁcation [50], and is an exciting area due to its simplicity and the ability to easily adapt to other languages.”
Universal Language Model Fine-tuning for Text Classification (2018, background): “…modeling can be seen as the ideal source task and a counter-part of ImageNet for NLP: It captures many facets of language relevant for downstream tasks, such as long-term dependencies (Linzen et al., 2016), hierarchical relations (Gulordava et al., 2018), and sentiment (Radford et al., 2017).”
Learned in Translation: Contextualized Word Vectors (2017, neutral): “…89.4 SVM [Van-Tu and Anh-Cuong, 2016] 95.2 TE-LSTM [Huang et al., 2017] 89.6 DSCNN-P [Zhang et al., 2016] 95.6 NSE [Munkhdalai and Yu, 2016a] 89.7 BCN+Char+CoVe (Ours) 95.8 BCN+Char+CoVe (Ours) 90.3 TBCNN [Mou et al., 2015] 96.0 bmLSTM [Radford et al., 2017] 91.8 LSTM-CNN [Zhou et al., 2016] 96.1”
Fine-Tuning Language Models from Human Preferences (2019, methodology): “Since the reward model needs to understand language, we initialize it as a random linear function of the final embedding output of the language model policy ρ following Radford et al. (2018) (see section 4.”
Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism (2019, background): “…language tasks, one can achieve state of the art results as shown in recent work (Devlin et al., 2018; Peters et al., 2018; Howard & Ruder, 2018; Radford et al., 2018; 2017; Ramachandran et al., 2016; Liu et al., 2019b; Dai et al., 2019; Yang et al., 2019; Liu et al., 2019a; Lan et al., 2019).”
The Natural Language Decathlon: Multitask Learning as Question Answering (2018, methodology): “We use the unparsed, binary version (Radford et al., 2017).”
Learning General Purpose Distributed Sentence Representations via Large Scale Multi-task Learning (2018, methodology): “Shi et al. (2016) also present evidence that sequence-to-sequence parsers (Vinyals et al.”
Unsupervised Pretraining for Sequence to Sequence Learning (2016, background): “Dai and Le (2015); Radford et al. (2017) are amongst the rare studies which showed the benefits of pretraining in a semi-supervised learning setting.”
Fine-tuned Language Models for Text Classification (2018, background): “Language modeling can be seen as the prototypical NLP task: Language models capture many facets of language, such as long-term dependencies and sentiment that are relevant for downstream tasks (Radford et al., 2017).”
Representation Learning with Contrastive Predictive Coding (2018, methodology): “Skip-thought vectors [26] and Byte mLSTM [27] are alternatives which go beyond word prediction with a Recurrent Neural Network, and use maximum likelihood over sequences of observations.”
GeDi: Generative Discriminator Guided Sequence Generation (2020, neutral): “Natural language generation has seen great progress with the advent of Transformers (Vaswani et al., 2017) and large scale training (Radford et al., 2017, 2018, 2019; Brown et al., 2020).”
Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding (2022, methodology): “However, we find that human raters prefer T5-XXL over CLIP on DrawBench across all 11 categories.”
Controlling Linguistic Style Aspects in Neural Language Generation (2017, background): “…the content of the generated text (Lipton et al., 2015; Kid-don et al., 2016; Lebret et al., 2016; Kiddon et al., 2016; Tang et al., 2016; Radford et al., 2017), while a few model more stylistic aspects of the generated text such as the identity of the speaker in a dialog setting (Li et…”
Using millions of emoji occurrences to learn any-domain representations for detecting sentiment, emotion and sarcasm (2017, background): “observation that even bag-ofwords classifiers and unsupervised approaches can obtain a high accuracy (Joulin et al., 2016; Radford et al., 2017).”
Federated Learning for Emoji Prediction in a Mobile Keyboard (2019, background): “The Long-Short-Term Memory (LSTM) (Hochre-iter and Schmidhuber, 1997) architecture has been shown to achieve state-of-the art performance of a number of sentiment prediction and language modeling tasks (Radford et al., 2017).”
DeepXplore: Automated Whitebox Testing of Deep Learning Systems (2017, background): “It has recently been shown that each neuron in a DNN tends to independently extract a specific feature of the input instead of collaborating with other neurons for feature extraction [58, 87].”
A Survey of Methods for Explaining Black Box Models (2018, background): “Finally, it is worth mentioning that [81] presents the discovery that a single neuron unit of a DNN can perform alone a sentiment analysis task after the training of the network reaching the same level of performance of strong baselines.”