Learning Non-linguistic Skills without Sacrificing Linguistic Proficiency

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 6178–6191 July 9-14, 2023 ©2023 Association for Computational Linguistics Learning Non-linguistic Skills without Sacrificing Linguistic Proficiency Mandar Sharma Virginia Tech mandarsharma@vt.edu Nikhil Muralidhar Stevens Institute of Technology nmurali1@stevens.edu Naren Ramakrishnan Virginia Tech naren@cs.vt.edu Abstract The field of Math-NLP has witnessed signifi- cant growth in recent years, motivated by the desire to expand LLM performance to the learn- ing of non-linguistic notions (numerals, and subsequently, arithmetic reasoning). However, non-linguistic skill injection typically comes at a cost for LLMs: it leads to catastrophic forgetting of core linguistic skills, a conse- quence that often remains unaddressed in the literature. As Math-NLP has been able to cre- ate LLMs that can closely approximate the mathematical skills of a grade-schooler or the arithmetic reasoning skills of a calculator, the practicality of these models fail if they con- comitantly shed their linguistic capabilities. In this work, we take a closer look into the phe- nomena of catastrophic forgetting as it per- tains to LLMs and subsequently offer a novel framework for non-linguistic skill injection for LLMs based on information-theoretic interven- tions and skill-specific losses that enable the learning of strict arithmetic reasoning. Our model outperforms the state-of-the-art both on injected non-linguistic skills and on linguistic knowledge retention, and does so with a frac- tion of the non-linguistic training data (1/4) and zero additional synthetic linguistic training data. Our pre-trained models and experimenta- tion codebases are hosted online1. 1 Introduction Numeracy, involving the comprehension of sizes, magnitudes, and order, is the most prevalent form of non-linguistic information embedded in tex- tual corpora (Joram et al., 1995). Thus, the case for numerically-capable LLMs is rather easy to make: as numerals grant objectivity to language (Porter, 1996), numerically-capable language mod- els are key to optimal performance in a host of downstream tasks such as information extraction (Madaan et al., 2016), inference (Naik et al., 2018), 1https://github.com/Mandar-Sharma/Skill-LM Figure 1: LLMs trained for dissimilar skillsets have different convergence points for their parameters - the parameterization space for an LLM trained for linguistic skills θLinguistics lives in the green space while the pa- rameterization space for an LLM trained for quantitative reasoning θQuant lives in the red space. The goal of this work is to approximate a locality of parameterization θSkill−LM (yellow) where the model reliably learns a non-linguistic skill (quantitative reasoning) without sac- rificing its linguistic proficiency. and data-to-text generation (Sharma et al., 2021, 2022a). 1.1 Re-thinking the Objective of Math-NLP Progress in Math-NLP: Several notable publi- cations in the Math-NLP space have made rapid strides in numeracy-tinged language-modeling (Thawani et al., 2021) - from investigations of the inherent deficiency of numerical reasoning skills in LLMs induced through unsupervised training, both for numerals that appear in the training corpus (Zhang et al., 2020) and OOD (out-of-domain) nu- merals (Wallace et al., 2019; Razeghi et al., 2022), to interventions that strengthen the numerical rea- soning skills of these models (Spithourakis and Riedel, 2018; Jiang et …

Content truncated to 500 words