Sharp Nearby, Fuzzy Far Away: How Neural Language Models Use Context

Abstract

We know very little about how neural language models (LM) use prior linguistic context. In this paper, we investigate the role of context in an LSTM LM, through ablation studies. Specifically, we analyze the increase in perplexity when prior context words are shuffled, replaced, or dropped. On two standard datasets, Penn Treebank and WikiText-2, we find that the model is capable of using about 200 tokens of context on average, but sharply distinguishes nearby context (recent 50 tokens) from the distant history. The model is highly sensitive to the order of words within the most recent sentence, but ignores word order in the long-range context (beyond 50 tokens), suggesting the distant past is modeled only as a rough semantic field or topic. We further find that the neural caching model (Grave et al., 2017b) especially helps the LSTM to copy words from within this distant context. Overall, our analysis not only provides a better understanding of how neural LMs use their context, but also sheds light on recent success from cache-based models.

ACL 2018

References in corpus (3)

• Language Modeling with Gated Convolutional Networks
• Tying Word Vectors and Word Classifiers: A Loss Framework for Language Modeling
• N-gram Language Modeling using Recurrent Neural Network Estimation

Cited by in corpus (28)

• Enhancing the Locality and Breaking the Memory Bottleneck of Transformer on Time Series Forecasting
• Calibrate Before Use: Improving Few-Shot Performance of Language Models
• MMM : Exploring Conditional Multi-Track Music Generation with the Transformer
• Dynamic Evaluation of Transformer Language Models
• Modeling Vocabulary for Big Code Machine Learning
• Neural Language Generation: Formulation, Methods, and Evaluation
• Recurrent Quantum Neural Networks
• Can LSTM Learn to Capture Agreement? The Case of Basque
• A Self-Attentional Neural Architecture for Code Completion with Multi-Task Learning
• Effective Low-Cost Time-Domain Audio Separation Using Globally Attentive Locally Recurrent Networks
• Lattice Transformer for Speech Translation
• Bidirectional Scene Text Recognition with a Single Decoder
• Hard but Robust, Easy but Sensitive: How Encoder and Decoder Perform in Neural Machine Translation
• Data Augmentation for Neural Online Chat Response Selection
• Dissecting Contextual Word Embeddings: Architecture and Representation
• On Training Recurrent Networks with Truncated Backpropagation Through Time in Speech Recognition
• Neural Academic Paper Generation
• You May Not Need Order in Time Series Forecasting
• Relevance-Promoting Language Model for Short-Text Conversation
• Which Kind Is Better in Open-domain Multi-turn Dialog,Hierarchical or Non-hierarchical Models? An Empirical Study
• H-Transformer-1D: Fast One-Dimensional Hierarchical Attention for Sequences
• Post-Processing of Word Representations via Variance Normalization and Dynamic Embedding
• Temporal Action Localization using Long Short-Term Dependency
• Context Dependent Semantic Parsing: A Survey
• An Augmented Transformer Architecture for Natural Language Generation Tasks
• Transformer-based language modeling and decoding for conversational speech recognition
• Sparsity Emerges Naturally in Neural Language Models
• DCT: Dynamic Compressive Transformer for Modeling Unbounded Sequence