Natural Language Text Research Articles

Exploring the Development of Safety Culture among Physicians with Text Mining of Patient Safety Reports: A Retrospective Study.

Safety culture development is essential for patient safety in healthcare institution. Perceptions of patient safety and cultural changes are shown up in patient safety reports, however, were rarely investigated. The aim of this study was to investigate the perception of physicians and to explore the development of safety culture using quantitative content analysis for patient safety reports. A retrospective analysis of free descriptions of harmful patient safety reports submitted by physicians was performed. Natural language processing and text analysis were conducted using the 'KH Coder'. A co-occurrence analysis was performed in each period to identify and analyze the safety concepts. The study period was grouped into three for comparison. The patient safety reports from physicians were collected between April 2004 and March 2020. Of these, 3351 reports were harmful; 839 reports were included in period 1, 1016 reports in period 2, and 1496 reports in period 3. Natural language processing identified 316 307 words in the free descriptions of 3351 reports. We identified seven concepts from the cluster in co-occurrence analysis as follows: 'explanation of adverse event to patients and families,' 'central venous catheter,' 'intraoperative procedure and injury,' 'minimally invasive surgery,' 'life-threatening events,' 'blood loss,' and 'medical emergency team and critical care.' These seven concepts showed significant differences among the three periods, except for 'blood loss'. The 'explanation of adverse event to patients and families' decreased in proportion from 11.3% to 8.8% (p<0.05). The 'central venous catheter' decreased from 17.3% to 11.3% (p<0.01). Meanwhile, 'minimally invasive surgeries' and 'intraoperative procedures' increased from 3.9% to 12.9% (p<0.01) and from 10.8% to 14.6% (p<0.05), respectively. Focusing on patients' events, 'life-threatening events' decreased from 13.0% to 8.1% (p<0.01); however, 'medical emergency teams and critical care' increased from 3.3% to 10.6% (p<0.01). Free description in patient safety reports is useful for evaluating the safety culture. Co-occurrence analysis revealed multiple concepts of physicians' perceptions. Quantitative content analysis revealed changes in perceptions and attitudes, and a disclosure policy of adverse events and priority of patient care appeared with the development of safety culture.

Analysis of Datasets and Large Language Models for Vulnerability Detection in Imperative Programming Language Code

Large language models are machine learning models that enable the classification and generation of both natural language text and code in various programming languages. These models have billions of parameters and are trained on vast datasets. In recent years, such models have been successfully applied to a wide range of tasks in software engineering. The paper presents data on publication activity on the topic under study, which is obtained on the basis of statistical analysis of search results for relevant key queries. In addition, a review of recent publications in the field of using large language models to detect vulnerabilities in program code is carried out, and the results of the analysis of data sets that are used in training neural network models to search for vulnerabilities in program code are presented.

Redescription mining-based business process deviance analysis

AbstractBusiness processes often deviate from their expected or desired behavior. Such deviations can be either positive or negative, depending on whether or not they lead to better process performance. Deviance mining addresses the problem of identifying such deviations and explaining why a process deviates. In this paper, we propose a novel approach to identify and explain the causes of deviant process executions based on the technique of redescription mining, which extracts knowledge in the form of logical rules. By analyzing, comparing, and filtering these rules, the reasons for the deviant behaviors of a business process are identified both in general and for particular process instances. Afterward, the results of this analysis are transformed into a concise and well-readable natural language text that can be used by business analysts and process owners to optimize processes in a reasoned manner. We evaluate our approach from different angles using four process models and provide some advice for further optimization.

Deep Learning-Based Freight Recommendation System for Freight Brokerage Platform

Platform-based businesses in the logistics market are evolving under the influence of digital transformation. Transforming the freight market into an environment where various types of freight can be traded across multiple markets and locations. Freight brokerage platforms have revolutionized the trading relationship between freight owners and vehicle owners. However, this type of system has also introduced inefficiencies, such as unestablished contracts, leading to unnecessary costs and delays. To address this issue, a freight recommendation system can assist users in finding what they are looking for while aiming to reduce failed contracts. With current advances in deep learning, complex patterns based on users’ past behaviors and preferences can be learned, enabling more accurate and personalized recommendations. This study proposes a deep learning-based freight recommendation system to provide personalized services and reduce failed contracts on freight brokerage platforms. The system is built by creating a freight transaction dataset, classifying freight categories through natural language processing and text mining techniques, and incorporating externally derived data on transportation distances. The deep learning model is trained using Autoencoder, Word2Vec, and Graph Neural Networks (GNN), with recommendation logic implemented to suggest suitable freight matches for vehicle owners. This system is expected to increase the market efficiency of the freight logistics industry and is a key step toward improving the long-term profit structure.

Tracking technical change: Past, present and future

AbstractProductivity growth in many countries has remained low for several years. Whether new technologies can reverse the trend depends on the scope of their impact and scale of their adoption—two dimensions of technical change that are historically difficult to measure. Here, we elaborate on the materials and methods presented in Alexopoulos's presidential address at the 2024 Canadian Economics Association meeting. Specifically, we discuss how applying natural language processing and text mining to library collections and cataloguing materials can help: (i) identify new technologies as they come to market and (ii) track their uses and spread over time. We further describe how our insights can be used to uncover general purpose technologies and macro‐innovations in both the past and the present. An application to current data suggests that AI and robotics are responsible for an increasing share of recent technical change. Moreover, they resemble past early‐stage general purpose technologies and thus do promise a reversal in productivity trends as their adoption increases. Going forward, our new methods should be especially useful to economists and policy‐makers who need to track future development and adoption of key technologies—especially during periods of rapid innovation.

Empowering legal justice with AI: A reinforcement learning SAC-VAE framework for advanced legal text summarization.

Automated summarization of legal texts poses a significant challenge due to the complex and specialized nature of legal documentation. Despite the recent progress in reinforcement learning for natural language text summarization, its application in the legal domain has been less effective. This paper introduces SAC-VAE, a novel reinforcement learning framework specifically designed for legal text summarization. We leverage a Variational Autoencoder (VAE) to condense the high-dimensional state space into a more manageable lower-dimensional feature space. These compressed features are subsequently utilized by the Soft Actor-Critic (SAC) algorithm for policy learning, facilitating the automated generation of summaries from legal texts. Through comprehensive experimentation, we have empirically demonstrated the effectiveness and superior performance of the SAC-VAE framework in legal text summarization.

Cancer-related Keywords in 2023: Insights from Text Mining of a Major Consumer Portal.

With the growing importance of monitoring cancer patients' internet usage, there is an increasing need for technology that expands access to relevant information through text mining. This study analyzed internet articles from portal sites in 2023 to identify trends in the information available to cancer patients and to derive meaningful insights. This study analyzed 19,578 news articles published on Naver, a major Korean portal site, from January 1, 2023, to December 31, 2023. Natural language processing, text mining, network analysis, and word cloud analysis were employed. The search term "am" (Korean for "cancer") was used to identify keywords related to cancer. In 2023, an average of 1,631 cancer-related articles were published monthly, with a peak of 1,946 in September and a low of 1,371 in February. A total of 132,456 keywords were extracted, with "cure" (2,218 occurrences), "lung cancer" (1,652), and "breast cancer" (1,235) being the most frequent. Term frequency-inverse document frequency analysis ranked "struggle" (1064.172) as the most significant keyword, followed by "lung cancer" (839.988) and "breast cancer" (744.840). Network analysis revealed four distinct clusters focusing on treatment, celebrity-related issues, major cancer types, and cancer-causing factors. The analysis of cancer-related keywords in 2023 indicates that news articles often prioritize gossip over essential information. These findings provide foundational data for future policy directions and strategies to address misinformation. This study underscores the importance of understanding the nature of cancer-related information consumed by the public and offers insights to guide official policies and healthcare practices.

The Power of Culture: Business Nationalist Culture and ESG Performance

High-quality development is the theme of China’s economic and social development in the new era, and corporate ESG performance is a comprehensive indicator for evaluating the level of corporate environmental responsibility, social responsibility and governance, as well as an important yardstick for identifying the high-quality development of enterprises. This paper takes Chinese non-financial listed companies from 2011 to 2022 as the research sample and empirically examines the impact of corporate nationalism culture on corporate ESG performance and its mechanism by quantifying corporate nationalism culture using the text of corporate annual reports, natural language processing and text analysis methods. The results of the study show that corporate nationalism culture significantly enhances corporate ESG performance. The mechanism analysis suggests that corporate nationalism culture, as an internal informal system, can play a governance role and promote corporate ESG practices by changing attention allocation and mitigating agency problems. The positive effect of corporate nationalism culture on corporate ESG performance is more pronounced in the grouping of firms with lower institutional investor shareholding, fewer analysts’ attention and embedded party organisations. A heterogeneity analysis reveals that the corporate nationalism culture driving effect on corporate ESG performance is more significant in the subsample of firms with weak financing constraints, in the growth period and in the decline period. This study reveals the positive role of soft cultural factors in enhancing corporate ESG performance, providing useful managerial evidence for companies to integrate ESG concepts at the strategic level for high-quality economic development.

Corporate environmental, social, and governance information disclosure and audit governance in the context of green development

The purpose of this paper is to investigate the relationship between Environmental, Social, and Governance (ESG) disclosure and audit governance. The objective is to assess the impact of auditor team independence, auditor professional quality, and transparency of audit procedures on the quality of information disclosure, and to elucidate the role of these factors in advancing sustainable development. Research methodologies include the optimization of information disclosure text analysis using natural language processing (NLP) and text generation techniques, complemented by empirical analysis. Key findings indicate that the maximum error between NLP and manual annotation in the proportion of positive emotions in information disclosure is 0.03, and 0.06 in audit governance. The Rouge index score surpasses the 0.5 threshold, while the Perplexity value varies across three experiments, though remaining generally low. Deep learning model results reveal that auditor team independence, professional quality, and procedural transparency significantly enhance the quality of information disclosure. These findings offer crucial insights for enterprises, regulators, and investors, affirming the rationality and stability of the model, and providing valuable input for future green development strategies and decision-making processes.

Toponym resolution leveraging lightweight and open-source large language models and geo-knowledge

Toponym resolution is crucial for extracting geographic information from natural language texts, such as social media posts and news articles. Despite the advancements in current methods, including state-of-the-art deep learning solutions like GENRE and a sophisticated voting system that integrates seven individual methods, further enhancing their accuracy is essential. To achieve this goal, we propose a novel method that combines lightweight and open-source large language models and geo-knowledge. Specifically, we first fine-tune Mistral (7B), Baichuan2 (7B), Llama2 (7B & 13B), and Falcon (7B) to estimate toponyms’ unambiguous reference (e.g., city, state, country) given their contexts. Subsequently, we correct inaccuracies in generated references and determine their geo-coordinates via sequentially querying GeoNames, Nominatim, and ArcGIS geocoders until a successful geocoding result is achieved. Our methods demonstrate enhanced performance compared to 20 existing methods, as evidenced across seven challenging datasets including 83,365 toponyms worldwide, with the Mistral-based method leading, followed by Baichuan2, Llama2, and Falcon-based methods. Specifically, the Mistral-based method achieves an Accuracy@161km of 0.91, surpassing GENRE, the best individual method, by 17% and the seven-methods composite voting system by 7%. Moreover, our methods are computationally efficient, operable on one general GPU, have modest memory requirements (14 GB for 7B models and 27 GB for 13B models), and exceed both GENRE and the voting system in inferring speed.

TTPXHunter: Actionable Threat Intelligence Extraction as TTPs from Finished Cyber Threat Reports

Understanding the modus operandi of adversaries aids organizations to employ efficient defensive strategies and share intelligence in the community. This knowledge is often present in unstructured natural language text within threat analysis reports. A translation tool is needed to interpret the modus operandi explained in the sentences of the threat report and convert it into a structured format. This research introduces a methodology named TTPXHunter for automated extraction of threat intelligence in terms of Tactics, Techniques, and Procedures (TTPs) from finished cyber threat reports. It leverages cyber domain-specific state-of-the-art natural language model to augment sentences for minority class TTPs and refine pinpointing the TTPs in threat analysis reports significantly. We create two datasets: an augmented sentence-TTP dataset of \(39,296\) sentence samples and a \(149\) real-world cyber threat intelligence report-to-TTP dataset. Further, we evaluate TTPXHunter on the augmented sentence and report datasets. The TTPXHunter achieves the highest performance of \(92.42\%\) f1-score on the augmented dataset, and it also outperforms existing state-of-the-art TTP extraction method by achieving an f1-score of \(97.09\%\) when evaluated over the report dataset. TTPXHunter significantly improves cybersecurity threat intelligence by offering quick, actionable insights into attacker behaviors. This advancement automates threat intelligence analysis and provides a crucial tool for cybersecurity professionals to combat cyber threats.

Development of message passing-based graph convolutional networks for classifying cancer pathology reports

BackgroundApplying graph convolutional networks (GCN) to the classification of free-form natural language texts leveraged by graph-of-words features (TextGCN) was studied and confirmed to be an effective means of describing complex natural language texts. However, the text classification models based on the TextGCN possess weaknesses in terms of memory consumption and model dissemination and distribution. In this paper, we present a fast message passing network (FastMPN), implementing a GCN with message passing architecture that provides versatility and flexibility by allowing trainable node embedding and edge weights, helping the GCN model find the better solution. We applied the FastMPN model to the task of clinical information extraction from cancer pathology reports, extracting the following six properties: main site, subsite, laterality, histology, behavior, and grade.ResultsWe evaluated the clinical task performance of the FastMPN models in terms of micro- and macro-averaged F1 scores. A comparison was performed with the multi-task convolutional neural network (MT-CNN) model. Results show that the FastMPN model is equivalent to or better than the MT-CNN.ConclusionsOur implementation revealed that our FastMPN model, which is based on the PyTorch platform, can train a large corpus (667,290 training samples) with 202,373 unique words in less than 3 minutes per epoch using one NVIDIA V100 hardware accelerator. Our experiments demonstrated that using this implementation, the clinical task performance scores of information extraction related to tumors from cancer pathology reports were highly competitive.

Rethinking Cultural Ecosystem Services in Urban Forest Parks: An Analysis of Citizens’ Physical Activities Based on Social Media Data

Urban forest parks play a vital role in promoting physical activities (PAs) and providing cultural ecosystem services (CESs) that enhance citizens’ well-being. This study aims to reevaluate CESs by focusing on the physical activity experiences of park visitors to optimize park management and enhance citizen satisfaction. This study utilized social media data and employed natural language processing techniques and text analysis tools to examine experiences related to physical activities in Beijing Olympic Forest Park, Xishan Forest Park, and Beigong Forest Park. A specialized sports activity dictionary was developed to filter and analyze comments related to PA, emphasizing the impact of natural environments on enjoyment and participation in PA. The importance–performance analysis (IPA) method was used to assess the service characteristics of each park. The findings reveal that urban forest parks are highly valued by citizens, particularly for their natural landscapes, leisure activities, and the emotional fulfillment derived from PA, with 82.58% of comments expressing positive sentiments. Notably, appreciation for natural landscapes was exceptionally high, as evidenced by the frequent mentions of key terms such as ‘scenery’ (mentioned 2871 times), ‘autumn’ (mentioned 2314 times), and ‘forest’ (mentioned 1439 times), which significantly influence park usage. However, 17.11% of the reviews highlighted dissatisfaction, primarily with the management of facilities and services during sports and cultural activities. These insights underscore the need for performance improvements in ecological environments and sports facilities. This study provides a novel perspective on assessing and optimizing urban forest parks’ functions, particularly in supporting active physical engagement. The rich CESs offered by these parks enhance physical activity experiences and overall satisfaction. The findings offer strategic insights for park managers to better meet citizens’ needs and improve park functionality.

Leveraging Large Language Models to Support Authoring Gamified Programming Exercises

Skilled programmers are in high demand, and a critical obstacle to satisfying this demand is the difficulty of acquiring programming skills. This issue can be addressed with automated assessment, which gives fast feedback to students trying to code, and gamification, which motivates them to intensify their learning efforts. Although some collections of gamified programming exercises are available, producing new ones is very demanding. This paper presents GAMAI, an AI-powered exercise gamifier, enriching the Framework for Gamified Programming Education (FGPE) ecosystem. Leveraging large language models, GAMAI enables teachers to effortlessly apply storytelling to describe a gamified scenario, as GAMAI decorates natural language text with the sentences needed by OpenAI APIs to contextualize the prompt. Once a gamified scenario has been generated, GAMAI automatically produces exercise files in a FGPE-compatible format. According to the presented evaluation results, most gamified exercises generated with AI support were ready to be used, with no or minimum human effort, and were positively assessed by students. The usability of the software was also assessed as high by the users. Our research paves the way for a more efficient and interactive approach to programming education, leveraging the capabilities of advanced language models in conjunction with gamification principles.

Uniqorn: Unified question answering over RDF knowledge graphs and natural language text

Question answering over RDF data like knowledge graphs has been greatly advanced, with a number of good systems providing crisp answers for natural language questions or telegraphic queries. Some of these systems incorporate textual sources as additional evidence for the answering process, but cannot compute answers that are present in text alone. Conversely, the IR and NLP communities have addressed QA over text, but such systems barely utilize semantic data and knowledge. This paper presents a method for complex questions that can seamlessly operate over a mixture of RDF datasets and text corpora, or individual sources, in a unified framework. Our method, called Uniqorn, builds a context graph on-the-fly, by retrieving question-relevant evidences from the RDF data and/or a text corpus, using fine-tuned BERT models. The resulting graph typically contains all question-relevant evidences but also a lot of noise. Uniqorn copes with this input by a graph algorithm for Group Steiner Trees, that identifies the best answer candidates in the context graph. Experimental results on several benchmarks of complex questions with multiple entities and relations, show that Uniqorn significantly outperforms state-of-the-art methods for heterogeneous QA – in a full training mode, as well as in zero-shot settings. The graph-based methodology provides user-interpretable evidence for the complete answering process.

Enhancing knowledge transformation in digital education: an analysis of the SECI model’s application in course design and execution

This study investigates the application of the SECI model of knowledge dimensions in the design and execution of educational courses in “Innovation and Digitization Management” and “Data-Based Decision Making” microdegrees, developed within the rapidly evolving educational landscape of 2023 and 2024. The research incorporates a diverse sample of learners across multiple demographics to enhance generalizability. Leveraging natural language processing (NLP) and text analysis, the study explores patterns in course content that correlate with positive learner feedback and effective knowledge transformation. The methodology, combining both quantitative and qualitative approaches, includes data preprocessing, tokenization, vectorization, and clustering to compare and contrast course elements with learner feedback systematically. Preliminary findings indicate that the integration of the SECI model, emphasizing real-time content sharing and example-based learning, significantly enhances the transfer of knowledge from tacit to explicit forms. This research aims to identify a replicable cadence in content preparation that optimizes learning outcomes, applicable to a broad range of educational contexts including data-related courses, innovation, digital transformation or knowledge-management topics.

Latent Relations at Steady-state with Associative Nets.

Models of word meaning that exploit patterns of word usage across large text corpora to capture semantic relations, like the topic model and word2vec, condense word-by-context co-occurrence statistics to induce representations that organize words along semantically relevant dimensions (e.g., synonymy, antonymy, hyponymy, etc.). However, their reliance on latent representations leaves them vulnerable to interference, makes them slow learners, and commits to a dual-systems account of episodic and semantic memory. We show how it is possible to construct the meaning of words online during retrieval to avoid these limitations. We implement a spreading activation account of word meaning in an associative net, a one-layer highly recurrent network of associations, called a Dynamic-Eigen-Net, that we developed to address the limitations of earlier variants of associative nets when scaling up to deal with unstructured input domains like natural language text. We show that spreading activation using a one-hot coded Dynamic-Eigen-Net outperforms the topic model and reaches similar levels of performance as word2vec when predicting human free associations and word similarity ratings. Latent Semantic Analysis vectors reached similar levels of performance when constructed by applying dimensionality reduction to the Shifted Positive Pointwise Mutual Information but showed poorer predictability for free associations when using an entropy-based normalization. An analysis of the rate at which the Dynamic-Eigen-Net reaches asymptotic performance shows that it learns faster than word2vec. We argue in favor of the Dynamic-Eigen-Net as a fast learner, with a single-store, that is not subject to catastrophic interference. We present it as an alternative to instance models when delegating the induction of latent relationships to process assumptions instead of assumptions about representation.

A universal knowledge model and cognitive architectures for prototyping AGI

The article identified 56 cognitive architectures for creating general artificial intelligence (AGI) and proposed a set of interrelated functional blocks that an agent approaching AGI in its capabilities should possess. Since the required set of blocks is not found in any of the existing architectures, the article proposes a reference cognitive architecture for intelligent systems approaching AGI in their capabilities. As one of the key solutions within the framework of the architecture, a universal method of knowledge representation is proposed, which allows combining various non-formalized, partially and fully formalized methods of knowledge representation in a single knowledge base, such as texts in natural languages, images, audio and video recordings, graphs, algorithms, databases, neural networks, knowledge graphs, ontologies, frames, essence-property-relation models, production systems, predicate calculus models, conceptual models, and others. To combine and structure various fragments of knowledge, archigraph model are used, constructed as a development of annotated metagraphs. As other components, the reference cognitive architecture being developed includes following modules: machine consciousness, machine subconsciousness, interaction with the external environment, a goal management, an emotional control, social interaction, reflection, ethics, worldview, learning, monitoring, statement problems, solving problems, self-organization and meta learning. Based on the composition of the proposed reference architecture modules, existing cognitive architectures containing the following modules were analyzed: machine consciousness, machine subconsciousness, reflection, worldview.

Kyrtos: A methodology for automatic deep analysis of graphic charts with curves in technical documents

Deep Understanding of Technical Documents (DUTD) has become a very attractive field with great potential due to large amounts of accumulated documents and the valuable knowledge contained in them. In addition, the holistic understanding of technical documents depends on the accurate analysis of its particular modalities, such as graphics, tables, diagrams, text, etc. and their associations. In this paper, we introduce the Kyrtos methodology for the automatic recognition and analysis of charts with curves in graphics images of technical documents. The recognition processing part adopts a clustering based approach to recognize middle-points that delimit the line-segments that construct the illustrated curves. The analysis processing part parses the extracted line-segments of curves to capture behavioral features such as direction, trend and etc. These associations assist the conversion of recognized segments’ relations into attributed graphs, for the preservation of the curves’ structural characteristics. The graph relations are also are expressed into natural language (NL) text sentences, enriching the document’s text and facilitating their conversion into Stochastic Petri-net (SPN) graphs, which depict the internal functionality represented in the chart image. Extensive evaluation results demonstrate the accuracy of Kyrtos’ recognition and analysis methods by measuring the structural similarity between input chart curves and the approximations generated by Kyrtos for charts with multiple functions.

Automatically Generated Texts for Fauna and Flora from Structured Data Based on a Controlled Vocabulary

Information systems like Xper3 and Fulgoromorpha Lists On the Web (FLOW) play a crucial role in managing and using biological data. These platforms store extensive collections of normalized data and structured taxonomic descriptions. By using controlled terminologies, they standardize the vocabulary, significantly enhancing the processes of identification, description, and comparison of various taxa. The massive assemblages of data hosted in these repositories could be reused to generate texts in natural languages automatically. The most immediate goal is to produce, from these information systems, more accessible and user-friendly displays in the form of taxon summary pages. This automated production of textual outputs is a great addition that can be continuously updated as the databases evolve. Can structured data be reused to provide better species pages and to ensure updating if the data evolves? Will AI assist in this process, or will specific computing be needed? To address these questions, multiple possible approaches have been identified. The most basic level of this process involves converting a single line from a taxon-by-character matrix into text that resembles natural language, similar to the descriptions found in botanical or zoological guides. The primary goal is to move beyond the rigid format of matrix lines or lists of characteristics (such as characters and states) of a species, and instead generate a coherent, easy-to-read paragraph intended for human eyes. To achieve this objective, a first solution is given by Descrxp, a tool currently developed in conjunction with Xper. The user specifies the desired key outline for the output, and Descrxp fills this canvas using the database contents. While this approach is highly reliable and adaptable to various contexts, its drawback is being labor-intensive, and requiring significant human input and oversight (Fig. 1). The second solution uses AI. Data on well-known taxa, and on more obscure taxa have been tested with ChatGPT 3.5 or 4.0. ChatGPT succeeds well at generating natural language descriptions for well-known taxa (Fig. 2), providing text that appears accurate and coherent due to the extensive information available online for these groups. However, when it comes to more obscure taxa, less represented on the internet, the results can be more inconsistent and unpredictable (Fig. 3). This variability can be explained by the quality of the data the model has been trained on, which is significantly richer for well-known species. Consequently, for lesser-known groups, the generated descriptions may lack the precision and reliability seen with more familiar taxa. However, if more careful attention is given to the output of chatGPT (Fig. 4), it is possible to notice that the AI filled part of the description with inaccurate characters. Due to its inability to recognize specific scientific terminology, and to process hierarchical information, chatGPT ends up producing an over-generalized and redundant description. To go further, it would be possible to use a comparison between taxa to produce a text that highlights the remarkable features of a taxon. Moreover, it could be interesting to try and synthesize several lines of the initial matrix to extract generalized descriptions of groups of taxa, such as the description of a genus based on its included species. Using data from FLOW, Fishbase, and Xper3, the French ACDC (Counterfactual Learning for Controlled Data-to-text) project is creating a variety of text outputs. These range from "fill-in-the-blank" canvas to fully AI-generated content, utilizing data matrices derived from these pilot databases.