Representation Degeneration Problem in Training Natural Language Generation Models

By Jun Gao, Di He, Xu Tan, Tao Qin, Liwei Wang, Tie-Yan Liu. 2019

386 citations recorded by the source corpus.

We study an interesting problem in training neural network-based models for natural language generation tasks, which we call the \emph{representation degeneration problem}. We observe that when training a model for natural language generation tasks through likelihood maximization with the weight tying trick, especially with big training datasets, most of the learnt word embeddings tend to degenerate and be distributed into a narrow cone, which largely limits the representation power of word embeddings. We analyze the conditions and causes of this problem and propose a novel regularization method to address it. Experiments on language modeling and machine translation show that our method can largely mitigate the representation degeneration problem and achieve better performance than baseline algorithms.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

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 Computer Vision and Pattern Recognition

Citations in the corpus

The case for 4-bit precision: k-bit Inference Scaling Laws (2022, background): “Outlier-dependent Quantization Through Proxy Quantization Outlier features that emerge in large language models (Gao et al., 2019; Timkey & van Schijndel, 2021; Bondarenko et al., 2021; Wei et al., 2022; Luo et al., 2021; Kovaleva et al., 2021; Puccetti et al., 2022) can cause large quantization…”
LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale (2022, neutral): “Previous work proved the theoretical relationship between outlier appearance in transformers and how it relates to layer normalization and the token frequency distribution (Gao et al., 2019).”
Improving Neural Language Modeling via Adversarial Training (2019, methodology): “In the particular setting of enforcing diversity of word embeddings, Gao et al. (2019) show that adding a cosine similarity reg-ularizer improves language modeling performance, which has the form However, in language modeling, one disadvantage of the direct diversity regularization approach is that…”
SimCSE: Simple Contrastive Learning of Sentence Embeddings (2021, background): “…we also try taking neutral and contradiction pairs from the NLI datasets as positive pairs. in analyzing word embeddings (Mu and Viswanath, 2018; Gao et al., 2019; Wang et al., 2020), and show that the contrastive objective can “flatten” the singular value distribution of sentence embeddings…”
COCO-LM: Correcting and Contrasting Text Sequences for Language Model Pretraining (2021, background): “Similar to empirical observations and theoretical analyses in recent research [14, 16, 55], a more regularized representation space results in better generalization ability in scenarios with limited labels.”
Adapting Large Language Models by Integrating Collaborative Semantics for Recommendation (2023, methodology): “Following previous work [56], [57], we employ Principal Component Analysis (PCA) to visualize the embeddings corresponding to different tokens.”
PowerNorm: Rethinking Batch Normalization in Transformers (2020, methodology): “Representation Power of learned Embedding To investigate further the performance gain of PN, we compute the Singular Value Decomposition of the embedding layers, as proposed by (Gao et al., 2019), which argued that the singular value distribution could be used as a proxy for measuring representation”
Taking Notes on the Fly Helps BERT Pre-training (2020, positive): “Gao et al. (2019) provides a theoretical understanding of this problem, which illustrates that the problem lies in the sparse (and inaccurate) stochastic optimization of neural networks.”
Outliers Dimensions that Disrupt Transformers Are Driven by Frequency (2022, background): “Transformer-based language models (LMs) have been shown to exhibit anisotropic behavior in their representations of both tokens and sentences (Etha-yarajh, 2019; Gao et al., 2019; Rajaee and Pilehvar, 2021; Timkey and van Schijndel, 2021).”
On the Sentence Embeddings from BERT for Semantic Textual Similarity (2020, background): “Gao et al. (2019) and Wang et al. (2020) have pointed out that, for language modeling, the maximum likelihood training with Equation 1 usually produces an anisotropic word embedding space.”
Mind the Gap: Understanding the Modality Gap in Multi-modal Contrastive Representation Learning (2022, methodology): “A common explanation is that the unbalanced distribution of word frequencies biased the optimization [15, 33].”
Analyzing Transformers in Embedding Space (2022, background): “While a substantive line of work [Ethayarajh, 2019; Gao et al., 2019; Wang et al., 2020; Rudman et al., 2021] has shown that embedding matrices are not isotropic (an isotropic matrix E has to satisfy EE T = αI for some scalar α ), we show that it is isotropic enough to make E T a legitimate…”
Positional Artefacts Propagate Through Masked Language Model Embeddings (2020, background): “Per the insights of Gao et al. (2019), it is very likely that the representational spaces of such models are anisotropic, but it is important to gauge the extent to which this can be traced to positional artefacts.”
Improve Vision Transformers Training by Suppressing Over-smoothing (2021, methodology): “Similar approaches have also been adopted for training improved representations in graph neural networks (Chen et al., 2020) and diversiﬁed word embeddings in language models (Gao et al., 2019).”
All Bark and No Bite: Rogue Dimensions in Transformer Language Models Obscure Representational Quality (2021, background): “Gao et al. (2019) introduce a cosine-regularization term during pretraining which improved the performance of transformer models on machine translation.”
Whitening Sentence Representations for Better Semantics and Faster Retrieval (2021, methodology): “(Gao et al., 2019) designs a novel way to mitigate the degeneration problem by regularizing the word embedding matrix.”
ConSERT: A Contrastive Framework for Self-Supervised Sentence Representation Transfer (2021, result): “Such phenomenon is also observed in several previous works (Gao et al., 2019; Wang et al., 2019; Li et al., 2020).”