Learning to Generate Reviews and Discovering Sentiment

2017/04/05 by Alec Radford, Rafal Jozefowicz, Radford, Alec +4 · 79 citations
Computer Science · #Advanced Text Analysis Techniques #Natural Language Processing Techniques #Topic Modeling #cs.CL #cs.LG #cs.NE

paper · pdf · doi:10.48550/arxiv.1704.01444

arxiv created 2017/04/06 · arxiv updated 2017/04/07

Abstract

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.

Citations

• Ask Me Anything: Dynamic Memory Networks for Natural Language Processing
• Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation
• Adversarial Training Methods for Semi-Supervised Text Classification

Cited by

• Representational Curvature Modulates Behavioral Uncertainty in Large Language Models
• PPSEBM: An Energy-Based Model with Progressive Parameter Selection for Continual Learning
• Unsupervised Pretraining for Sequence to Sequence Learning
• Persuasion for Good: Towards a Personalized Persuasive Dialogue System for Social Good
• Augmenting Neural Networks with First-order Logic
• Auxiliary Metrics Help Decoding Skill Neurons in the Wild
• The geometry of integration in text classification RNNs
• LAMOL: LAnguage MOdeling for Lifelong Language Learning
• Fine-Tuning Language Models from Human Preferences
• Selectivity considered harmful: evaluating the causal impact of class selectivity in DNNs
• Disentangled Representations for Manipulation of Sentiment in Text
• From attribution maps to human-understandable explanations through Concept Relevance Propagation
• A Mathematical Exploration of Why Language Models Help Solve Downstream Tasks
• Towards Language Agnostic Universal Representations
• Senti-Attend: Image Captioning using Sentiment and Attention
• GeDi: Generative Discriminator Guided Sequence Generation
• Towards Controllable and Personalized Review Generation
• Do Convolutional Networks need to be Deep for Text Classification ?
• Objective-Reinforced Generative Adversarial Networks (ORGAN) for Sequence Generation Models
• Atomic Literary Styling: Mechanistic Manipulation of Prose Generation in Neural Language Models
• Learning to Generate Music With Sentiment
• Assessing the Helpfulness of Review Content for Explaining Recommendations
• Decoding Emotion in the Deep: A Systematic Study of How LLMs Represent, Retain, and Express Emotion
• Nonparametric Identification of Latent Concepts
• MindCraft: How Concept Trees Take Shape In Deep Models
• Latent Activation Editing: Inference-Time Refinement of Learned Policies for Safer Multirobot Navigation
• AI-Powered Text Generation for Harmonious Human-Machine Interaction: Current State and Future Directions
• Generating and designing DNA with deep generative models
• A journey in ESN and LSTM visualisations on a language task
• Style Example-Guided Text Generation using Generative Adversarial Transformers
• Object category learning and retrieval with weak supervision
• Distributional Generalization: A New Kind of Generalization
• Redefining Experts: Interpretable Decomposition of Language Models for Toxicity Mitigation
• On Completeness-aware Concept-Based Explanations in Deep Neural Networks
• Statistical Methods in Generative AI
• The Low-Dimensional Linear Geometry of Contextualized Word Representations
• What made you do this? Understanding black-box decisions with sufficient input subsets
• Large Scale Language Modeling: Converging on 40GB of Text in Four Hours
• Exploring Deep Neural Networks and Transfer Learning for Analyzing Emotions in Tweets
• Understanding Pre-trained BERT for Aspect-based Sentiment Analysis
• Learning to Generate Multiple Style Transfer Outputs for an Input Sentence
• Emergent Properties of Finetuned Language Representation Models
• Image Captioning at Will: A Versatile Scheme for Effectively Injecting Sentiments into Image Descriptions
• Authorship Attribution in Bangla literature using Character-level CNN
• Cooperative Learning of Disjoint Syntax and Semantics
• What Does it Mean for a Neural Network to Learn a "World Model"?
• Using General Adversarial Networks for Marketing: A Case Study of Airbnb
• DONUT: CTC-based Query-by-Example Keyword Spotting
• Towards falsifiable interpretability research
• Toward estimating personal well-being using voice
• Mining Program Properties From Neural Networks Trained on Source Code Embeddings
• Linking average- and worst-case perturbation robustness via class selectivity and dimensionality
• The emergence of number and syntax units in LSTM language models
• Under the Hood of Neural Networks: Characterizing Learned Representations by Functional Neuron Populations and Network Ablations
• Cell-aware Stacked LSTMs for Modeling Sentences
• Dynamic Compositionality in Recursive Neural Networks with Structure-aware Tag Representations
• Towards Controlled Transformation of Sentiment in Sentences
• Improving Context Aware Language Models
• Multi-Modal Emotion recognition on IEMOCAP Dataset using Deep Learning
• Unsupervised Representation Learning for Time Series with Temporal Neighborhood Coding
• Efficient Contextualized Representation: Language Model Pruning for Sequence Labeling
• Persona Features Control Emergent Misalignment
• How Do You Act? An Empirical Study to Understand Behavior of Deep Reinforcement Learning Agents
• Robustly Improving LLM Fairness in Realistic Settings via Interpretability
• Machine Learning Techniques for Software Quality Assurance: A Survey
• Discovery of Natural Language Concepts in Individual Units of CNNs
• Detecting Parkinson's Disease from interactions with a search engine: Is expert knowledge sufficient?
• Explainable Software Defect Prediction: Are We There Yet?
• AspeRa: Aspect-based Rating Prediction Model
• Semi-Supervised Self-Growing Generative Adversarial Networks for Image Recognition
• Explaining Neural Networks with Reasons
• Meta-Learning a Dynamical Language Model
• Concept-Guided Interpretability via Neural Chunking
• Adaptive Noise Injection: A Structure-Expanding Regularization for RNN
• Understanding the Importance of Single Directions via Representative Substitution
• Generating Sentiment-Preserving Fake Online Reviews Using Neural Language Models and Their Human- and Machine-based Detection
• On Variational Learning of Controllable Representations for Text without Supervision
• Neuron Populations Exhibit Divergent Selectivity with Scale
• Controlling Repetition in Protein Language Models
• Affect in Tweets Using Experts Model

Related

• Distilling Fine-grained Sentiment Understanding from Large Language Models
• Assessing State-of-the-Art Sentiment Models on State-of-the-Art Sentiment Datasets
• Ensemble of Generative and Discriminative Techniques for Sentiment Analysis of Movie Reviews
• Linear Representations of Sentiment in Large Language Models
• Unpaired Sentiment-to-Sentiment Translation: A Cycled Reinforcement Learning Approach