Improved Language Modeling by Decoding the Past (P19-1)

Copied to clipboard

Challenge: Existing methods to improve language modeling performance are based on regularized LSTMs with a large number of parameters and training time.
Approach: They propose a method that decodes the last token in context using the predicted distribution of the next token.
Outcome: The proposed method improves perplexity on the Penn Treebank dataset by 1.8 points and 2.3 points on the WikiText-2 datasets.

Similar Papers

An Embarrassingly Simple Method to Mitigate Undesirable Properties of Pretrained Language Model Tokenizers (2022.acl-short)

Copied to clipboard

Challenge: a standard tokenizer does not cover all characters of a word but preserves key aspects of its morphological structure . a novel method to improve tokenization of pretrained language models is proposed .
Approach: They propose a method to improve the tokenization of pretrained language models . they use the vocabulary of a standard tokenizer but preserves morphological structure .
Outcome: The proposed method improves tokenization of pretrained language models on morphological gold segmentations and text classification tasks.
Sharp Nearby, Fuzzy Far Away: How Neural Language Models Use Context (P18-1)

Copied to clipboard

Challenge: Recent studies have shed light on the information encoded by long-term memory networks.
Approach: They propose to use a neural caching model to model the role of context in an LSTM LM . they analyze the increase in perplexity when prior context words are shuffled, replaced, or dropped .
Outcome: The proposed model is highly sensitive to the order of words within the most recent sentence, but ignores word order in the long-range context, suggesting the distant past is modeled only as a rough semantic field or topic.
Efficient Training of Language Models with Compact and Consistent Next Token Distributions (2024.findings-acl)

Copied to clipboard

Challenge: Existing methods to train language models have focused on maximizing the likelihood of the next token . however, the construction and querying of such n-grams can be costly and impede training speed.
Approach: They propose a method to train language models faster by pre-aggregating corpus with collapsed n-gram distribution.
Outcome: The proposed model improves model quality and convergence rate while reducing variance across mini-batches compared to the standard next-token loss method.
Learn Your Tokens: Word-Pooled Tokenization for Language Modeling (2023.findings-emnlp)

Copied to clipboard

Challenge: Language models typically tokenize text into subwords, using a deterministic, hand-engineered heuristic of combining characters into longer surface-level strings such as ‘ing’ or whole words.
Approach: They propose a 'learn your tokens' scheme which pooles bytes/characters into word representations and decodes individual characters/bytes per word in parallel.
Outcome: The proposed tokenizer outperforms subword models and byte/character models over the word boundary and outperformed on rare words by a factor of 30!
DecoCal: Decoding with Calibration in Diffusion Large Language Models (2026.acl-long)

Copied to clipboard

Challenge: Diffusion Large Language Models (DLLMs) generate text via iterative token denoising . but decoding is challenging, with many tokens appearing predictable early .
Approach: They propose a Decoding framework that performs Calibration of token-level confidence across diffusion steps and leverages the calibrated results to guide decoding decisions.
Outcome: Experiments on multiple DLLMs and benchmarks show that DecoCal improves generation accuracy compared to existing strategies.
Token-wise Decomposition of Autoregressive Language Model Hidden States for Analyzing Model Predictions (2023.acl-long)

Copied to clipboard

Challenge: Recent work on why Transformer-based large language models make predictions has made their behavior opaque due to the complexity of the computations performed within each layer.
Approach: They propose a linear decomposition of final hidden states from autoregressive language models based on each initial input token, which is exact for virtually all contemporary Transformer architectures.
Outcome: The proposed method analyzes the influence of input tokens on model probabilities over a sequence of upcoming words with only one forward pass from the model.
Mitigating the Learning Bias towards Repetition by Self-Contrastive Training for Open-Ended Generation (2023.findings-acl)

Copied to clipboard

Challenge: Existing language models generate repetitive texts with greedy decoding or beam search.
Approach: They propose a self-contrastive training technique to penalize the output of a premature checkpoint of the same model when it incorrectly predicts repetition.
Outcome: The proposed training mitigates repetition while maintaining fluency while minimizing the overestimation of token-level repetition probabilities.
Improving Temporal Generalization of Pre-trained Language Models with Lexical Semantic Change (2022.emnlp-main)

Copied to clipboard

Challenge: Existing methods to improve neural language models perform poorly on emerging data.
Approach: They propose a lexical-level masking strategy to post-train a neural language model using static data from past years.
Outcome: The proposed method outperforms existing methods on two pre-trained language models, two classification tasks, and four benchmark datasets.
Look-back Decoding for Open-Ended Text Generation (2023.emnlp-main)

Copied to clipboard

Challenge: Existing approaches to decode open-ended text have addressed degeneration problems in large-scale language models (LLMs)
Approach: They propose an improved decoding algorithm that leverages the Kullback–Leibler divergence to track the distribution distance between current and historical decoding steps.
Outcome: The proposed algorithm outperforms existing methods in document continuation and story generation.
Nonparametric Masked Language Modeling (2023.findings-acl)

Copied to clipboard

Challenge: Existing language models (LMs) predict tokens with a softmax over a finite vocabulary, which can make it difficult to predict rare tokens or phrases.
Approach: They introduce a nonparametric masked language model that replaces a softmax with a distribution over every phrase in a reference corpus and uses an in-batch approximation to train it.
Outcome: The proposed model outperforms larger parametric models on 16 tasks including classification, fact probing and question answering.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations