minicons: Enabling Flexible Behavioral and Representational Analyses of Transformer Language Models (2203.13112v1)

Abstract: We present minicons, an open source library that provides a standard API for researchers interested in conducting behavioral and representational analyses of transformer-based LLMs (LMs). Specifically, minicons enables researchers to apply analysis methods at two levels: (1) at the prediction level -- by providing functions to efficiently extract word/sentence level probabilities; and (2) at the representational level -- by also facilitating efficient extraction of word/phrase level vectors from one or more layers. In this paper, we describe the library and apply it to two motivating case studies: One focusing on the learning dynamics of the BERT architecture on relative grammatical judgments, and the other on benchmarking 23 different LMs on zero-shot abductive reasoning. minicons is available at https://github.com/kanishkamisra/minicons

• The paper presents a standardized API that enables flexible behavioral and representational analyses on transformer language models.
• It validates minicons through case studies on grammatical judgments and unsupervised abductive reasoning to reveal model intricacies.
• Its integration with the HuggingFace hub and support for CPU/GPU setups paves the way for scalable and reproducible NLP research.

Analyzing Transformer LLMs with Minicons

The paper "minicons: Enabling Flexible Behavioral and Representational Analyses of Transformer LLMs" introduces the minicons library, a tool aimed at facilitating research on transformer-based LLMs (LMs) for NLP. The library presents a standardized API that empowers researchers to conduct detailed behavioral and representational analyses of transformer LMs, handling both prediction-level and representation-level analyses without necessitating additional supervised learning or fine-tuning.

Overview of Minicons

Minicons builds upon the widely used transformers library and focuses on two primary modes of analysis. At the prediction level, researchers can delve into the LLMs' abilities in word prediction tasks, applying methods to evaluate linguistic capacities, commonsense reasoning, and biases. Representational analysis, on the other hand, involves the extraction of token and phrase embeddings from various layers of the model, inviting the investigation of the information encoded in these internal activations.

Minicons consists of two core modules:

• Scorer module: This is designed for tasks that involve estimating word probabilities in context, enabling detailed word-level and sequence-level analyses. The module includes functionality for masked LLMs (MLMs) and autoregressive LMs.
• CWE (contextual word embeddings) module: This module helps researchers extract contextual embeddings from different layers of a LLM, supporting analysis methods such as probing linguistic competencies and comparing model representations.

Case Studies

The paper provides two case studies to demonstrate the utility of minicons. The first paper investigates the learning dynamics of BERT variations concerning grammatical judgments using the BLiMP benchmark. By analyzing how BERT models learn various linguistic phenomena over the training steps, the paper reveals insights into their performance on tasks like agreement, scope, and binding phenomena. Interestingly, while some competencies are acquired early in training, others such as island effects are learned more gradually.

The second case paper examines the efficacy of LLMs in unsupervised abductive reasoning, as evaluated with the Abductive Natural Language Inference dataset. The performance of various models, including BERT, RoBERTa, ALBERT, GPT, and others, is measured to see how well they discern plausible explanations given partial observations. The findings indicate that, while some models slightly surpass chance performance, they generally fall short of fine-tuned versions.

Implications and Prospects

These analyses underscore the potential of minicons as a versatile tool for probing into the operational and representational capabilities of transformer LMs. By providing an accessible framework for deploying varied NLP analysis strategies, it enhances the ability to benchmark and enhance LLMs across a spectrum of linguistic and reasoning tasks.

Minicons' integration with the HuggingFace model hub renders it a pivotal tool for large-scale analysis, aligning seamlessly with existing model infrastructures and supporting both CPU and GPU computations. The insights derived from employing minicons can influence future research trajectories, potentially steering toward the incorporation of structured knowledge or enhanced architectural designs to overcome current limitations in language understanding and reasoning.

In conclusion, the minicons library constitutes a significant contribution to the NLP toolkit, offering extensive capabilities for researchers dedicated to uncovering the intricacies of transformer LLMs. Its development and open-source nature present opportunities for continuous community contributions and innovative applications in unraveling the complexities of modern linguistic models.