minicons: Enabling Flexible Behavioral and Representational Analyses of Transformer Language Models

Misra, Kanishka

doi:10.48550/arxiv.2203.13112

Cited by 6 publications

(3 citation statements)

References 30 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Therefore, in the following we primarily focus on the surprisal values obtained from GPT-3.5. For calculating surprisal with GPT-2, we utilized the implementation by Misra ( 2022 ). The detailed GPT-2 results, along with a corresponding plot, can be accessed on GitHub ( https://www.github.com/tjuzek/om-uid ).…”

Section: More Quantitative Datamentioning

confidence: 99%

Signal Smoothing and Syntactic Choices: A Critical Reflection on the UID Hypothesis

Juzek

2024

Open Mind

View full text Add to dashboard Cite

The Smooth Signal Redundancy Hypothesis explains variations in syllable length as a means to more uniformly distribute information throughout the speech signal. The Uniform Information Density hypothesis seeks to generalize this to choices on all linguistic levels, particularly syntactic choices. While there is some evidence for the Uniform Information Density hypothesis, it faces several challenges, four of which are discussed in this paper. First, it is not clear what exactly counts as uniform. Second, there are syntactic alternations that occur systematically but that can cause notable fluctuations in the information signature. Third, there is an increasing body of negative results. Fourth, there is a lack of large-scale evidence. As to the fourth point, this paper provides a broader array of data—936 sentence pairs for nine syntactic constructions—and analyzes them in a test setup that treats the hypothesis as a classifier. For our data, the Uniform Information Density hypothesis showed little predictive capacity. We explore ways to reconcile our data with theory.

show abstract

Section: More Quantitative Datamentioning

confidence: 99%

Signal Smoothing and Syntactic Choices: A Critical Reflection on the UID Hypothesis

Juzek

2024

Open Mind

View full text Add to dashboard Cite

show abstract

“…Each image is presented on a white background (see Figure 1). For LLaVA, we compare the log-probabilities of the model using the basic or subordinate label as a continuation to the prompt using the minicons library (Misra, 2022). We code a response as basic if P(basic|prompt, image) > P(subordinate|prompt, image) and subordinate otherwise.…”

Section: ✓ ✓mentioning

confidence: 99%

Prompting sometimes invokes expert-like downward shifts in multimodal models’ conceptual hierarchies

Leong,

Lake

2024

Preprint

View full text Add to dashboard Cite

Humans tend to privilege an intermediate level of categorization, known as the basic level, when categorizing objects that exist in a conceptual hierarchy (e.g. choosing to call a Labrador a dog instead of Labrador or animal). Domain experts demonstrate a downward shift in their object categorization behaviour, recruiting subordinate levels in a conceptual hierarchy as readily as conventionally basic categories (Tanaka & Philibert, 2022; Tanaka & Taylor, 1991). Do multimodal large language models show similar behavioural changes when prompted to behave in an expert-like way? We test whether GPT-4 with Vision (GPT-4V, OpenAI, 2023a) and LLaVA-1.5 (Liu, Li, Wu, & Lee, 2023; Liu, Li, Li, & Lee, 2023) demonstrate downward shifts using an object naming task and eliciting expert-like personas by altering the model’s system prompt. We find evidence of downward shifts in GPT-4V when expert system prompts are used, suggesting that human expert-like behaviour can be elicited from GPT-4V using prompting, but find no evidence of downward shift in LLaVA. We also find that there is an unpredicted upward shift in areas of non-expertise in some cases. These findings suggest that in the default case, GPT-4V is not a novice: instead, it behaves at default with a median level of expertise, while further expertise can be primed or forgotten through textual prompts. These results open the door for GPT-4V and similar models to be used as tools for studying differences in the behaviour of experts and novices, and even comparing contrasting levels of expertise within the same large language model.

show abstract

“…Surprisal for each non-initial word in the sentences was estimated using probabilities generated by the open source transformer language model GPT-2 (Radford et al 2019), as shown for the regions surrounding the verb in Figure 2. These surprisals were calculated in Python using the minicons package (Misra 2022), which provides convenience wrappers for the Hugging Face transformers library (Wolf et al 2020). 3.1.4.…”

Section: Tasks For Investigating the Role Of Predictionmentioning

confidence: 99%

Beyond Surprising: English Event Structure in the Maze

Levinson

2023

ELM

View full text Add to dashboard Cite

To what extent can we tease apart semantic representations and processes from other influences on processing such as probabilistic prediction? In this paper I detail two experiments testing the hypothesis that there are semantic complexity in the lexical representations of result verbs that influences reaction times above and beyond probabilistic distributions. This is done by replicating a self-paced reading study from Levinson & Brennan (2016) while also modelling lexical surprisal. Experiment 1 replicates the original result, but only in experiment 2 using the maze task does the effect emerge beyond surprisals. The more focal maze task results suggest that processing costs associated with bieventive result verbs should be accounted for by grammatical factors, in addition to probabilistic prediction.

show abstract

minicons: Enabling Flexible Behavioral and Representational Analyses of Transformer Language Models

Cited by 6 publications

References 30 publications

Signal Smoothing and Syntactic Choices: A Critical Reflection on the UID Hypothesis

Signal Smoothing and Syntactic Choices: A Critical Reflection on the UID Hypothesis

Prompting sometimes invokes expert-like downward shifts in multimodal models’ conceptual hierarchies

Beyond Surprising: English Event Structure in the Maze

Contact Info

Product

Resources

About