Language Model Evaluation Beyond Perplexity

Abstract

We propose an alternate approach to quantifying how well language models learn natural language: we ask how well they match the statistical tendencies of natural language.To answer this question, we analyze whether text generated from language models exhibits the statistical tendencies present in the humangenerated text on which they were trained. We provide a framework-paired with significance tests-for evaluating the fit of language models to these trends. We find that neural language models appear to learn only a subset of the tendencies considered, but align much more closely with empirical trends than proposed theoretical distributions (when present). Further, the fit to different distributions is highly-dependent on both model architecture and generation strategy. As concrete examples, text generated under the nucleus sampling scheme adheres more closely to the typetoken relationship of natural language than text produced using standard ancestral sampling; text from LSTMs reflects the natural language distributions over length, stopwords, and symbols surprisingly well.

Modeling the Unigram Distribution

“…As neural networks yield stateof-the-art performance in language modeling tasks, we expect them to also do well with the unigram distribution. In fact, pseudo-text generated by LSTM-based language models reproduces Zipf's law to some extent (Takahashi and Tanaka-Ishii, 2017;Meister and Cotterell, 2021). Thus, we view state-of-the-art LSTM models as a strong baseline.…”

Probing BERT's priors with serial reproduction chains

“…In this section, we examine the extent to which higher-order statistics of sentences from BERT's prior are well-calibrated to the data it was trained on. This kind of comparison provides a richer sense of what the model has learned or failed to learn than traditional scalar metrics like perplexity (Takahashi and Tanaka-Ishii, 2017; Meister and Cotterell, 2021;Takahashi and Tanaka-Ishii, 2019).…”

Evaluating Distributional Distortion in Neural Language Modeling

“…ing frameworks (Meister & Cotterell, 2021) to better understand whether the large-scale statistical tendencies of natural language, such as Zipf's law (Zipf, 1949), are captured by LMs. We take a more fine-grained approach, proposing a methodology which draws off of instance-level evaluation schemes (Zhong et al, 2021) and the experimental control afforded by artificial corpora (White & Cotterell, 2021;Papadimitriou & Jurafsky, 2020).…”

“…Recently, a growing body work has sought to understand how these language models (LM) fit the distribution of a language beyond standard measures such as perplexity. Meister & Cotterell (2021), for example, investigated the statistical tendencies of the distribution defined by neural LMs, whereas Kulikov et al (2021) explored whether they adequately capture the modes of the distribution they attempt to model. At the same time, increased focus has been given to performance on rare or novel events in the data distribution, both for models of natural language (McCoy et al, 2021;Lent et al, 2021;Dudy & Bedrick, 2020;Oren et al, 2019) and neural models more generally (see, for example Sagawa et al, 2020;D'souza et al, 2021;Blevins & Zettlemoyer, 2020;Czarnowska et al, 2019;Horn & Perona, 2017;Ouyang et al, 2016;Bengio, 2015;Zhu et al, 2014).…”