Natural Language Research Articles

An Attempt to Generate Mozart's Piano Compositions Based on an LSTM Model

LSTM, proposed by German computer scientists Sepp Hochreiter and Jürgen Schmidhuber in 1997 [3], has shown outstanding performance in various tasks, especially in tasks that require capturing long-term dependencies. These tasks include language modeling and machine translation in the field of natural language processing, as well as stock price prediction. This project aims to explore the potential of LSTM models in the field of music generation by analyzing Mozart's classical piano works using music21 and creating a training dataset. The project uses LSTM to learn this dataset, adjusting the forget gate, input gate, and different training iteration times to generate diversified outputs. Finally, the generated music data was manually selected and labeled.

The Use of Natural Language Processing Methods in Reddit to Investigate Opioid Use: Scoping Review.

The growing availability of big data spontaneously generated by social media platforms allows us to leverage natural language processing (NLP) methods as valuable tools to understand the opioid crisis. We aimed to understand how NLP has been applied to Reddit (Reddit Inc) data to study opioid use. We systematically searched for peer-reviewed studies and conference abstracts in PubMed, Scopus, PsycINFO, ACL Anthology, IEEE Xplore, and Association for Computing Machinery data repositories up to July 19, 2022. Inclusion criteria were studies investigating opioid use, using NLP techniques to analyze the textual corpora, and using Reddit as the social media data source. We were specifically interested in mapping studies' overarching goals and findings, methodologies and software used, and main limitations. In total, 30 studies were included, which were classified into 4 nonmutually exclusive overarching goal categories: methodological (n=6, 20% studies), infodemiology (n=22, 73% studies), infoveillance (n=7, 23% studies), and pharmacovigilance (n=3, 10% studies). NLP methods were used to identify content relevant to opioid use among vast quantities of textual data, to establish potential relationships between opioid use patterns or profiles and contextual factors or comorbidities, and to anticipate individuals' transitions between different opioid-related subreddits, likely revealing progression through opioid use stages. Most studies used an embedding technique (12/30, 40%), prediction or classification approach (12/30, 40%), topic modeling (9/30, 30%), and sentiment analysis (6/30, 20%). The most frequently used programming languages were Python (20/30, 67%) and R (2/30, 7%). Among the studies that reported limitations (20/30, 67%), the most cited was the uncertainty regarding whether redditors participating in these forums were representative of people who use opioids (8/20, 40%). The papers were very recent (28/30, 93%), from 2019 to 2022, with authors from a range of disciplines. This scoping review identified a wide variety of NLP techniques and applications used to support surveillance and social media interventions addressing the opioid crisis. Despite the clear potential of these methods to enable the identification of opioid-relevant content in Reddit and its analysis, there are limits to the degree of interpretive meaning that they can provide. Moreover, we identified the need for standardized ethical guidelines to govern the use of Reddit data to safeguard the anonymity and privacy of people using these forums.

Distinguishing word identity and sequence context in DNA language models

Transformer-based large language models (LLMs) are very suited for biological sequence data, because of analogies to natural language. Complex relationships can be learned, because a concept of "words" can be generated through tokenization. Training the models with masked token prediction, they learn both token sequence identity and larger sequence context. We developed methodology to interrogate model learning, which is both relevant for the interpretability of the model and to evaluate its potential for specific tasks. We used DNABERT, a DNA language model trained on the human genome with overlapping k-mers as tokens. To gain insight into the model′s learning, we interrogated how the model performs predictions, extracted token embeddings, and defined a fine-tuning benchmarking task to predict the next tokens of different sizes without overlaps. This task evaluates foundation models without interrogating specific genome biology, it does not depend on tokenization strategies, vocabulary size, the dictionary, or the number of training parameters. Lastly, there is no leakage of information from token identity into the prediction task, which makes it particularly useful to evaluate the learning of sequence context. We discovered that the model with overlapping k-mers struggles to learn larger sequence context. Instead, the learned embeddings largely represent token sequence. Still, good performance is achieved for genome-biology-inspired fine-tuning tasks. Models with overlapping tokens may be used for tasks where a larger sequence context is of less relevance, but the token sequence directly represents the desired learning features. This emphasizes the need to interrogate knowledge representation in biological LLMs.

The influence of follower reactions on political leaders’ environmental policy communication: toward a social-mediated dynamic influence framework

ABSTRACT This study proposes that social-mediated leadership communication is a dynamic influence process, involving visible leader communications, follower emotional responses, and subsequent leader actions. These interactions form a feedback loop, where follower reactions influence leaders' strategic communication. We tested this framework by analyzing U.S. political leaders' engagement with the Green New Deal during the 2020 Presidential campaign. Using Facebook posts from congressional politicians and data on follower reactions, we applied longitudinal network models, natural language processing, and multilevel event history models to examine factors behind leaders' decisions to co-share content and discuss specific policy topics. Our findings support the dynamic influence framework, showing that follower reactions, especially those indicating support and boosting message virality, are most likely to prompt shifts in leaders’ communication strategies.

Tuning Vision-Language Models with Multiple Prototypes Clustering.

Benefiting from advances in large-scale pre-training, foundation models, have demonstrated remarkable capability in the fields of natural language processing, computer vision, among others. However, to achieve expert-level performance in specific applications, such models often need to be fine-tuned with domain-specific knowledge. In this paper, we focus on enabling vision-language models to unleash more potential for visual understanding tasks under few-shot tuning. Specifically, we propose a novel adapter, dubbed as lusterAdapter, which is based on trainable multiple prototypes clustering algorithm, for tuning the CLIP model. It can not only alleviate the concern of catastrophic forgetting of foundation models by introducing anchors to inherit common knowledge, but also improve the utilization efficiency of few annotated samples via bringing in clustering and domain priors, thereby improving the performance of few-shot tuning. We have conducted extensive experiments on 11 common classification benchmarks. The results show our method significantly surpasses the original CLIP and achieves state-of-the-art (SOTA) performance under all benchmarks and settings. For example, under the 16-shot setting, our method exhibits a remarkable improvement over the original CLIP by 19.6%, and also surpasses TIP-Adapter and GraphAdapter by 2.7% and 2.2%, respectively, in terms of average accuracy across the 11 benchmarks. Code is available at https://github.com/uyzhang/Cluster-Adapter.

Augmented dialectal speech recognition for AI-based neuropsychological scale assessment in Alzheimer’s disease

Alzheimer’s disease (AD) is a prevalent and widespread neurodegenerative disorder among the older adult population worldwide. Among the numerous cognitive screening methods, neuropsychological scale assessments (NSA) are the widely utilized screening tools in clinical practice. The NSA places significant emphasis on speech-related questions, and thus the role of Automatic Speech Recognition (ASR) technology in AI-based NSA evaluations becomes particularly critical. However, the majority of ASR research pays limited attention to the study of speech recognition for various dialects. One of the primary reasons and main challenges for the scarcity of research in dialectal speech recognition is the limited availability of annotated dialectal data. Furthermore, the older adults’ unclear pronunciation leads to erroneous recognition of similar phonetic words by ASR. To overcome the challenges of limited annotated dialect data in various regions and the issues related to confusable pronunciation of older adults, we propose a dialectal collaborative encoder mechanism and a machine reading comprehension augmented model (MRCAM) that automatically corrects the texts in the NSA. The dialectal collaborative encoder mechanism utilizes layer frozenness to transfer recognition ability to the dialectal model. Meanwhile, the MRCAM model fills the gap by combining speech recognition algorithms with natural language processing (NLP) technology to recognize the dialects spoken by older adults. The experimental evaluation results indicate that the proposed augmented model achieves over 90% consistency with the physician’s judgment, providing the premise of the proposed methods for the task of automated screening for AD based on ASR.

Advanced Sarcasm Detection in Newspaper Headlines: A Comparative Study of Machine Learning Approaches

This study presents a comprehensive analysis of sarcasm detection in newspaper headlines using various machine learning techniques. We explore the effectiveness of logistic regression, Naïve Bayes (multinomial and Gaussian), random forest, and support vector machine (SVM) models in identifying sarcastic content within the concise format of news titles. Our research addresses the unique challenges posed by the brevity and impactful language characteristic of headlines. The study utilizes a curated dataset of sarcastic and non-sarcastic headlines, employing natural language processing techniques for preprocessing and feature extraction. Performance evaluation metrics include accuracy, precision, recall, and F1-score. Results indicate that the random forest model outperforms other approaches, achieving 94% accuracy in sarcasm detection. This research contributes to the growing field of sentiment analysis and offers insights into the nuanced task of decoding sarcasm in condensed textual formats. KEYWORDS: Random Forest, Support Vector Machine (SVM), TF-IDF, Natural Language Processing (NLP), Subtle Sarcasm

Transformer-Based Video Comment Analysis

In this work, we conduct a comprehensive analysis of sentiment in Bilibili comments using a Transformer-based model. We employ the mT5_m2o_chinese_simplified_crossSum model, a multitasking Transformer model specifically designed for summarizing Chinese text. The study involves scraping comments from a video related to the "trolley problem," followed by data cleaning and preprocessing. The preprocessed data is fed into the Transformer model, which generates summaries that accurately reflect the main viewpoints and discussions in the comments. Our results show that the model effectively captures key ethical dilemmas, personal opinions, and emotional responses, providing a nuanced understanding of public sentiment. This research highlights the potential of advanced NLP techniques in processing large volumes of user-generated content, offering valuable insights for businesses and researchers in understanding public opinion and facilitating ethical discussions. The study underscores the importance of optimizing Transformer models for specific tasks, demonstrating their flexibility and performance in text summarization. These findings provide a robust foundation for future applications and innovations in natural language processing.

Natural language processing for healthcare: Applications, progress, and future directions

Natural Language Processing (NLP) is causing a significant change in the healthcare industry. This state-of-the-art technology is transforming the processing, analysis, and utilisation of healthcare data by improving patient care and clinical decision-making. This article aims to analyse the current and prospective uses of natural language processing (NLP) in the healthcare industry. We examine the tangible advancements that have been made and shed light on the bright prospects that lie ahead. The approach for early sickness diagnosis involves the use of advanced predictive analytics based on natural language processing (NLP). By analysing unstructured patient narratives, models may accurately identify potential health issues with exceptional precision, sometimes exceeding 90% accuracy. This article discusses the benefits of using voice recognition and chatbots, such as improved patient-provider communication and decreased administrative tasks. Systems powered by natural language processing have made significant advancements, as shown by statistical data indicating that they attain an accuracy rate of 80% or more on tasks like clinical text classification. Consequently, medical record review and data extraction have been automated, hence alleviating the burden on healthcare personnel. Hospitals may use sentiment analysis on online reviews to assess patient satisfaction levels and implement targeted improvements, therefore enhancing the overall patient experience. Undoubtedly, natural language processing (NLP) is revolutionizing the healthcare industry by enabling data-driven insights to optimise operations, personalize patient therapy, and enhance overall patient care quality. In order to fully harness the capabilities of natural language processing, we must address existing challenges and promote continuous research in areas such as deep learning and explainability. Only by doing so can we establish a healthcare system that is both more health-oriented and reliant on data.

Teaching AI to Summarize Like a Human: AReinforcement Learning Experiment

Text summarization is a crucial task in natural language processing (NLP), aiming to distill extensive information into concise and coherent summaries. Traditional summarization methods, including both extractive and abstractive techniques, face challenges in generating summaries that balance brevity and informativeness. This paper explores the application of Reinforce- ment Learning with Human Feedback (RLHF) to address these challenges and enhance the quality of text summarization.We introduce an RLHF-based approach using the FLAN-T5-small model, which integrates human feedback into the reinforcement learning framework to refine summary generation. Our method leverages a dataset from the Hugging Face datasets library, consisting of diverse document-summary pairs. The model is pre-trained on a large corpus and fine-tuned using human feedback, which serves as a reward signal to guide the model towards generating more relevant and coherent summaries.Our experimental results demonstrate that the RLHF-enhanced model significantly outperforms traditional summarization methods. Quantitative evaluations using ROUGE and BLEU metrics reveal substantial improvements in summary quality, with increases of up to 12.5% in ROUGE- 1 and 9.8% in BLEU scores over baseline methods. Qualitative assessments by human evaluators further confirm that the RLHF-based model produces summaries that are more aligned with human expectations in terms of coherence and relevance.This study highlights the potential of RLHF to overcome the limitations of conventional summarization tech- niques, offering a robust framework for generating high-quality summaries across various domains. Future work will explore the scalability of this approach to more complex summarization tasks and the integration of additional feedback mechanisms to further enhance performance.

An effective framework for measuring the novelty of scientific articles through integrated topic modeling and cloud model

Novelty is a critical characteristic of innovative scientific articles, and accurately identifying novelty can facilitate the early detection of scientific breakthroughs. However, existing methods for measuring novelty have two main limitations: (1) Metadata-based approaches, such as citation analysis, are retrospective and do not alleviate the pressures of the peer review process or enable timely tracking of scientific progress; (2) Content-based methods have not adequately addressed the inherent uncertainty between the qualitative concept of novelty and the textual representation of papers. To address these issues, we propose a practical and effective framework for measuring the novelty of scientific articles through integrated topic modeling and cloud model, referred to as MNSA-ITMCM. In this framework, papers are represented as topic combinations, and novelty is reflected in the organic reorganization of these topics. We use the BERTopic model to generate semantically informed topics, and then apply a topic selection algorithm based on maximum marginal relevance to obtain a topic combination that balances similarity and diversity. Furthermore, we leverage the cloud model from fuzzy mathematics to quantify novelty, overcoming the uncertainty inherent in natural language expression and topic modeling to improve the accuracy of novelty measurement. To validate the effectiveness of our framework, we conducted empirical evaluations on papers from the Cell 2021 journal (biomedical domain) and the ICLR 2023 conference (computer science domain). Through correlation analysis and prediction error analysis, our framework demonstrated the ability to identify different types of novel papers and accurately predict their novelty levels. The proposed framework is applicable across diverse scientific disciplines and publication venues, benefiting researchers, librarians, science evaluation agencies, policymakers, and funding organizations by improving the efficiency and comprehensiveness of identifying novelty research.

An exploration into CTEPH medications: Combining natural language processing, embedding learning, in vitro models, and real-world evidence for drug repurposing.

In the modern era, the growth of scientific literature presents a daunting challenge for researchers to keep informed of advancements across multiple disciplines. We apply natural language processing (NLP) and embedding learning concepts to design PubDigest, a tool that combs PubMed literature, aiming to pinpoint potential drugs that could be repurposed. Using NLP, especially term associations through word embeddings, we explored unrecognized relationships between drugs and diseases. To illustrate the utility of PubDigest, we focused on chronic thromboembolic pulmonary hypertension (CTEPH), a rare disease with an overall limited number of scientific publications. Our literature analysis identified key clinical features linked to CTEPH by applying term frequency-inverse document frequency (TF-IDF) scoring, a technique measuring a term's significance in a text corpus. This allowed us to map related diseases. One standout was venous thrombosis (VT), which showed strong semantic links with CTEPH. Looking deeper, we discovered potential repurposing candidates for CTEPH through large-scale neural network-based contextualization of literature and predictive modeling on both the CTEPH and the VT literature corpora to find novel, yet unrecognized associations between the two diseases. Alongside the anti-thrombotic agent caplacizumab, benzofuran derivatives were an intriguing find. In particular, the benzofuran derivative amiodarone displayed potential anti-thrombotic properties in the literature. Our in vitro tests confirmed amiodarone's ability to reduce platelet aggregation significantly by 68% (p = 0.02). However, real-world clinical data indicated that CTEPH patients receiving amiodarone treatment faced a significant 15.9% higher mortality risk (p<0.001). While NLP offers an innovative approach to interpreting scientific literature, especially for drug repurposing, it is crucial to combine it with complementary methods like in vitro testing and real-world evidence. Our exploration with benzofuran derivatives and CTEPH underscores this point. Thus, blending NLP with hands-on experiments and real-world clinical data can pave the way for faster and safer drug repurposing approaches, especially for rare diseases like CTEPH.

A comprehensive data network for data-driven study of battery materials

ABSTRACT Data-driven material research for property prediction and material design using machine learning methods requires a large quantity, wide variety, and high-quality materials data. For battery materials, which are commonly polycrystalline, ceramics, and composites, multiscale data on substances, materials, and batteries are required. In this work, we develop a data network composed of three interlinked databases, from which we can obtain comprehensive data on substances such as crystal structures and electronic structures, data on materials such as chemical composition, structure, and properties, and data on batteries such as battery composition, operation conditions, and capacity. The data are extracted from research papers on solid electrolytes and cathode materials, selected by screening more than 330 thousand papers using natural language processing tools. Data extraction and curation are carried out by editors specialized in material science and trained in data standardization.

Artificial intelligence-based suicide prevention and prediction: A systematic review (2019–2023)

Suicide is a major global public health concern, and the application of artificial intelligence (AI) methods, such as natural language processing (NLP), machine learning (ML), and deep learning (DL), has shown promise in advancing suicide prediction and prevention efforts. Recent advancements in AI – particularly NLP and DL have opened up new avenues of research in suicide prediction and prevention. While several papers have reviewed specific detection techniques like NLP or DL, there has been no recent study that acts as a one-stop-shop, providing a comprehensive overview of all AI-based studies in this field. In this work, we conduct a systematic literature review to identify relevant studies published between 2019 and 2023, resulting in the inclusion of 156 studies. We provide a comprehensive overview of the current state of research conducted on AI-driven suicide prevention and prediction, focusing on different data types and AI techniques employed. We discuss the benefits and challenges of these approaches and propose future research directions to improve the practical application of AI in suicide research. AI is highly capable of improving the accuracy and efficiency of risk assessment, enabling personalized interventions, and enhancing our understanding of risk and protective factors. Multidisciplinary approaches combining diverse data sources and AI methods can help identify individuals at risk by analyzing social media content, patient histories, and data from mobile devices, enabling timely intervention. However, challenges related to data privacy, algorithmic bias, model interpretability, and real-world implementation must be addressed to realize the full potential of these technologies. Future research should focus on integrating prediction and prevention strategies, harnessing multimodal data, and expanding the scope to include diverse populations. Collaboration across disciplines and stakeholders is essential to ensure that AI-driven suicide prevention and prediction efforts are ethical, culturally sensitive, and person-centered.

Glycemic Control in Diabetic Patients Improved Overall Lung Cancer Survival across Diverse Populations.

The consequence of diabetes on lung cancer overall survival (OS) is debated. This retrospective study used two large lung cancer databases to assess comprehensively diabetes effects on lung cancer OS in diverse demographic populations, including health disparity. The University of Texas MD Anderson Cancer Center database (32,643 lung cancer cases with 11,973 diabetics) was extracted from electronic health records (EHRs) using natural language processing (NLP). Associations were between diabetes and lung cancer prognostic features [age, sex, race, body mass index (BMI), insurance status, smoking, stage, and histopathology]. Hemoglobin A1C (HgbA1c) and glucose levels assessed glycemic control. Validation was with a Louisiana cohort (17,768 lung cancer cases with 4,746 diabetics) enriched for health disparity cases. Kaplan-Meier analysis, log-rank test, multivariable Cox proportional hazard models, and survival tree analyses were employed. Lung cancer patients with diabetes exhibited marginally elevated OS or no statistically-significant difference versus non-diabetic patients. When examining OS for two glycemic levels (HgbA1c > 7.0 or glucose > 154 mg/dL versus HgbA1c > 9.0 or glucose > 215 mg/dL), a statistically significant improvement in OS occurred in lung cancers with controlled versus uncontrolled glycemia (P < 0.0001). This improvement spanned gender, age, smoking status, insurance status, stage, race, BMI, histopathology and therapy. Survival tree analysis revealed that obese and morbidly obese patients with controlled glycemia or no known diabetes had higher lung cancer OS than comparison groups. These findings indicate a need for optimal glycemic control to improve lung cancer OS in diverse populations with diabetes.

Temporal Coordination of Articulatory Gestures in Nonnative Onset Clusters: Evidence From American English Speakers Using Electromagnetic Articulography.

Research in cross-language speech production indicates that, although the production of nonnative consonant clusters is often difficult, speakers of American English can produce some nonnative clusters (e.g., /fn/) with high accuracy. This ease of production for select nonnative clusters may occur due to similarity of phonetic structure with native clusters (e.g., nonnative /fn/ and native /sm/ are both fricative-nasal sequences). The current study tested this hypothesis by examining the extent of transfer of articulatory coordination from phonetically similar native onset clusters (i.e., /fl/, /sm/) to nonnative /fn/ clusters. Using electromagnetic articulography, lip, tongue, and jaw movements were recorded in nine participants during the production of 22 nonwords (eight tokens per nonword) containing the native and nonnative clusters in different carrier phrases. We examined the temporal lags between each consonantal gesture in a cluster and the flanking vowel gesture, which were compared to the matched singleton conditions. Analyses revealed that, as in native speech, when the syllable onset became more complex (i.e., CV ➔ CCV [C as consonant, V as vowel]), there was an increase in lag (less temporal overlap) between the leftmost consonantal gesture and the vocalic gesture, whereas there was a decrease in lag (more temporal overlap) between the rightmost consonant and the vocalic gesture (i.e., C-center timing). However, the segmental makeup of the cluster and type of carrier phrase used were also found to influence this change in temporal organization, raising new questions for future research. By and large, the findings are in agreement with the idea that the temporal coordination of articulator movements may be transferred from native clusters to phonetically similar nonnative clusters. However, kinematic measures of a broader range of nonnative clusters in different contexts are needed to fully explore this position.

Natural Language Processing to assess structure and complexity of system requirements

AbstractThe development process of a system is shaped by numerous variables that influence its progress and outcome. As a result, complexity can increase throughout the development process, potentially leading to negative consequences, which makes the management of complexity critical. Most development processes begin with the definition of needs and requirements. In this paper, the authors present a novel approach that enables the automated extraction of structure from requirements specifications. The approach uses Natural Language Processing to elicit three structural layers from a set of requirements, which are subsequently analyzed using metrics to assess complexity. In a case study, the approach is demonstrated using a set of 79 requirements, within which 246 individual entities are identified. These entities and the requirements are structured and analyzed using network density and spectral entropy. The metrics allow for interpretation and insight generation, such as identifying an increase in the number of potentially problematic loops. The approach achieved a detection and structural accuracy of over 98% in the given case study and is planned to be expanded with future cases.

Linear symmetric self-selecting 14-bit kinetic molecular memristors.

Artificial Intelligence (AI) is the domain of large resource-intensive data centres that limit access to a small community of developers1,2. Neuromorphic hardware promises greatly improved space and energy efficiency for AI but is presently only capable of low-accuracy operations, such as inferencing in neural networks3-5. Core computing tasks of signal processing, neural network training and natural language processing demand far higher computing resolution, beyond that of individual neuromorphic circuit elements6-8. Here we introduce an analog molecular memristor based on a Ru-complex of an azo-aromatic ligand with 14-bit resolution. Precise kinetic control over a transition between two thermodynamically stable molecular electronic states facilitates 16,520 distinct analog conductance levels, which can be linearly and symmetrically updated or written individually in one time step, substantially simplifying the weight update procedure over existing neuromorphic platforms3. The circuit elements are unidirectional, facilitating a selector-less 64 × 64 crossbar-based dot-product engine that enables vector-matrix multiplication, including Fourier transform, in a single time step. We achieved more than 73 dB signal-to-noise-ratio, four orders of magnitude improvement over the state-of-the-art methods9-11, while consuming 460× less energy than digital computers12,13. Accelerators leveraging these molecular crossbars could transform neuromorphic computing, extending it beyond niche applications and augmenting the core of digital electronics from the cloud to the edge12,13.

Le Petit Prince Hong Kong (LPPHK): Naturalistic fMRI and EEG data from older Cantonese speakers

Currently, the field of neurobiology of language is based on data from only a few Indo-European languages. The majority of this data comes from younger adults neglecting other age groups. Here we present a multimodal database which consists of task-based and resting state fMRI, structural MRI, and EEG data while participants over 65 years old listened to sections of the story The Little Prince in Cantonese. We also provide data on participants’ language history, lifetime experiences, linguistic and cognitive skills. Audio and text annotations, including time-aligned speech segmentation and prosodic information, as well as word-by-word predictors such as frequency and part-of-speech tagging derived from natural language processing (NLP) tools are included in this database. Both MRI and EEG data diagnostics revealed that the data has good quality. This multimodal database could advance our understanding of spatiotemporal dynamics of language comprehension in the older population and help us study the effects of healthy aging on the relationship between brain and behaviour.

Synergizing human insight and machine learning: A dual-lens approach to uncovering healthcare research and innovation outcomes

Many healthcare organisations have extensive documentation detailing the processes behind their various research and innovation projects. Analysing this data can provide valuable insights into why some projects succeed without major issues, others encounter and overcome problems, and some ultimately fail. This study introduces an approach that combines narrative interviews and Natural Language Processing (NLP) to identify patterns associated with innovation project outcomes. We analysed 618 documents from 67 projects provided by ZonMw, a major Dutch healthcare research funder, and conducted 32 narrative interviews across seven cases of healthcare innovation projects. By using narrative interviews to inform and pre-train a text embedding model, we demonstrate the potential to create a proxy for human judgement, allowing for a more natural identification of contextual patterns in project documentation. The findings indicate that successful projects are more likely to adopt a proactive approach to role changes and uncertainty (due to ambiguous laws and regulations) and to allow flexibility, which enhances stakeholder engagement, compared to failed projects. However, while we were able to conduct descriptive analysis to gain these insights, significant interpretation is still required to fully understand the findings. Our study makes two primary contributions: first, it offers a new approach for future research on the factors that determine project success or failure, closely aligning with Structuration Theory. Additionally, it suggests potential efficiency improvements in theory development by enabling multiple pattern configurations within Grounded Theory. Second, it offers practical strategies for organisations to more effectively capture and use contextual information in their project documentation for future success.