Recent advances in large language models (LLMs), such as GPT-4, have spurred interest in their potential applications across various fields, including actuarial work. This paper introduces the use of LLMs in actuarial and insurance-related tasks, both as direct contributors to actuarial modelling and as workflow assistants. It provides an overview of LLM concepts and their potential applications in actuarial science and insurance, examining specific areas where LLMs can be beneficial, including a detailed assessment of the claims process. Additionally, a decision framework for determining the suitability of LLMs for specific tasks is presented. Case studies with accompanying code showcase the potential of LLMs to enhance actuarial work. Overall, the results suggest that LLMs can be valuable tools for actuarial tasks involving natural language processing or structuring unstructured data and as workflow and coding assistants. However, their use in actuarial work also presents challenges, particularly regarding professionalism and ethics, for which high-level guidance is provided.

Scientists have started to explore whether novel artificial intelligence (AI) tools based on large language models, such as GPT-4, could support the scientific peer review process. We sought to understand (i) whether AI versus human reviewers are able to distinguish between made-up AI-generated and human-written conference abstracts reporting on actual research, and (ii) how the quality assessments by AI versus human reviewers of the reported research correspond to each other. We conducted a large-scale field experiment during a medium-sized scientific conference, relying on 305 human-written and 20 AI-written abstracts that were reviewed either by AI or 217 human reviewers. The results show that human reviewers and GPTZero were better in discerning (AI vs. human) authorship than GPT-4. Regarding quality assessments, there was rather low agreement between both human–human and human–AI reviewer pairs, but AI reviewers were more aligned with human reviewers in classifying the very best abstracts. This indicates that AI could become a prescreening tool for scientific abstracts. The results are discussed with regard to the future development and use of AI tools during the scientific peer review process.

This paper addresses the optimization of retrieval-augmented generation (RAG) processes by exploring various methodologies, including advanced RAG methods. The research, driven by the need to enhance RAG processes as highlighted by recent studies, involved a grid-search optimization of 23,625 iterations. We evaluated multiple RAG methods across different vectorstores, embedding models, and large language models, using cross-domain datasets and contextual compression filters. The findings emphasize the importance of balancing context quality with similarity-based ranking methods, as well as understanding tradeoffs between similarity scores, token usage, runtime, and hardware utilization. Additionally, contextual compression filters were found to be crucial for efficient hardware utilization and reduced token consumption, despite the evident impacts on similarity scores, which may be acceptable depending on specific use cases and RAG methods.

Hierarchical text classification (HTC) is a natural language processing task which aims to categorise a text document into a set of classes from a hierarchical class structure. Recent approaches to solve HTC tasks focus on leveraging pre-trained language models (PLMs) and the hierarchical class structure by allowing these components to interact in various ways. Specifically, the Hierarchy-aware Prompt Tuning (HPT) method has proven to be effective in applying the prompt tuning paradigm to Bidirectional Encoder Representations from Transformers (BERT) models for HTC tasks. Prompt tuning aims to reduce the gap between the pre-training and fine-tuning phases by transforming the downstream task into the pre-training task of the PLM. Discriminative PLMs, which use a replaced token detection (RTD) pre-training task, have also shown to perform better on flat text classification tasks when using prompt tuning instead of vanilla fine-tuning. In this paper, we propose the Hierarchy-aware Prompt Tuning for Discriminative PLMs (HPTD) approach which injects the HTC task into the RTD task used to pre-train discriminative PLMs. Furthermore, we make several improvements to the prompt tuning approach of discriminative PLMs that enable HTC tasks to scale to much larger hierarchical class structures. Through comprehensive experiments, we show that our method is robust and outperforms current state-of-the-art approaches on two out of three HTC benchmark datasets.

The English Preposing in PP construction (PiPP; e.g., Happy though/as we were) is extremely rare but displays an intricate set of stable syntactic properties. How do people become proficient with this construction despite such limited evidence? It is tempting to posit innate learning mechanisms, but present-day large language models seem to learn to represent PiPPs as well, even though such models employ only very general learning mechanisms and experience very few instances of the construction during training. This suggests an alternative hypothesis on which knowledge of more frequent constructions helps shape knowledge of PiPPs. I seek to make this idea precise using model-theoretic syntax (MTS). In MTS, a grammar is essentially a set of constraints on forms. In this context, PiPPs can be seen as arising from a mix of construction-specific and general-purpose constraints, all of which seem inferable from general linguistic experience.

Growth in the complexity of advanced systems is mirrored by a growth in the number of engineering requirements and related upstream and downstream tasks. These requirements are typically expressed in natural language and require human expertise to manage. Natural language processing (NLP) technology has long been seen as promising to increase requirements engineering (RE) productivity but has yet to demonstrate substantive benefits. The recent addition of large language models (LLMs) to the NLP toolbox is now generating renewed enthusiasm in the hope that it will overcome past shortcomings. This article scrutinizes this claim by reviewing the application of LLMs for engineering requirements tasks. We survey the success of applying LLMs and the scale to which they have been used. We also identify groups of challenges shared across different engineering requirement tasks. These challenges show how this technology has been applied to RE tasks that need reassessment. We finalize by drawing a parallel to other engineering fields with similar challenges and how they have been overcome in the past – and suggest these as future directions to be investigated.

This article examines the idea of mind-reading technology by focusing on an interesting case of applying a large language model (LLM) to brain data. On the face of it, experimental results appear to show that it is possible to reconstruct mental contents directly from brain data by processing via a chatGPT-like LLM. However, the author argues that this apparent conclusion is not warranted. Through examining how LLMs work, it is shown that they are importantly different from natural language. The former operates on the basis of nonrational data transformations based on a large textual corpus. The latter has a rational dimension, being based on reasons. Using this as a basis, it is argued that brain data does not directly reveal mental content, but can be processed to ground predictions indirectly about mental content. The author concludes that this is impressive but different in principle from technology-mediated mind reading. The applications of LLM-based brain data processing are nevertheless promising for speech rehabilitation or novel communication methods.

To what extent can statistical language knowledge account for the effects of world knowledge in language comprehension? We address this question by focusing on a core aspect of language understanding: pronoun resolution. While existing studies suggest that comprehenders use world knowledge to resolve pronouns, the distributional hypothesis and its operationalization in large language models (LLMs) provide an alternative account of how purely linguistic information could drive apparent world knowledge effects. We addressed these confounds in two experiments. In Experiment 1, we found a strong effect of world knowledge plausibility (measured using a norming study) on responses to comprehension questions that probed pronoun interpretation. In experiment 2, participants were slower to read continuations that contradicted world knowledge-consistent interpretations of a pronoun, implying that comprehenders deploy world knowledge spontaneously. Both effects persisted when controlling for the predictions of GPT-3, an LLM, suggesting that pronoun interpretation is at least partly driven by knowledge about the world and not the word. We propose two potential mechanisms by which knowledge-driven pronoun resolution occurs, based on validation- and expectation-driven discourse processes. The results suggest that while distributional information may capture some aspects of world knowledge, human comprehenders likely draw on other sources unavailable to LLMs.

A recent paper by Häffner et al. (2023, Political Analysis 31, 481–499) introduces an interpretable deep learning approach for domain-specific dictionary creation, where it is claimed that the dictionary-based approach outperforms finetuned language models in predictive accuracy while retaining interpretability. We show that the dictionary-based approach’s reported superiority over large language models, BERT specifically, is due to the fact that most of the parameters in the language models are excluded from finetuning. In this letter, we first discuss the architecture of BERT models, then explain the limitations of finetuning only the top classification layer, and lastly we report results where finetuned language models outperform the newly proposed dictionary-based approach by 27% in terms of $R^2$ and 46% in terms of mean squared error once we allow these parameters to learn during finetuning. Researchers interested in large language models, text classification, and text regression should find our results useful. Our code and data are publicly available.

Large language models (LLMs) have achieved amazing successes. They have done well on standardized tests in medicine and the law. That said, the bar has been raised so high that it could take decades to make good on expectations. To buy time for this long-term research program, the field needs to identify some good short-term applications for smooth-talking machines that are more fluent than trustworthy.

Speaking with conversational AIs, technologies whose interfaces enable human-like interaction based on natural language, has become a common phenomenon. During these interactions, people form their beliefs due to the say-so of conversational AIs. In this paper, I consider, and then reject, the concepts of testimony-based beliefs and instrument-based beliefs as suitable for analysis of beliefs acquired from these technologies. I argue that the concept of instrument-based beliefs acknowledges the non-human agency of the source of the belief. However, the analysis focuses on perceiving signs and indicators rather than content expressed in natural language. At the same time, the concept of testimony-based beliefs does refer to natural language propositions, but there is an underlying assumption that the agency of the testifier is human. To fill the lacuna of analyzing belief acquisition from conversational AIs, I suggest a third concept: technology-based beliefs. It acknowledges the non-human agency-status of the originator of the belief. Concurrently, the focus of analysis is on the propositional content that forms the belief. Filling the lacuna enables analysis that considers epistemic, ethical, and social issues of conversational AIs without excluding propositional content or compromising accepted assumptions about the agency of technologies.