Disambiguation of acronyms in clinical narratives with large language models.

OBJECTIVE: To assess the performance of large language models (LLMs) for zero-shot disambiguation of acronyms in clinical narratives. MATERIALS AND METHODS: Clinical narratives in English, German, and Portuguese were applied for testing the performance of four LLMs: GPT-3.5, GPT-4, Llama-2-7b-chat, and Llama-2-70b-chat. For English, the anonymized Clinical Abbreviation Sense Inventory (CASI, University of Minnesota) was used. For German and Portuguese, at least 500 text spans were processed. The output of LLM models, prompted with contextual information, was analyzed to compare their acronym disambiguation capability, grouped by document-level metadata, the source language, and the LLM. RESULTS: On CASI, GPT-3.5 achieved 0.91 in accuracy. GPT-4 outperformed GPT-3.5 across all datasets, reaching 0.98 in accuracy for CASI, 0.86 and 0.65 for two German datasets, and 0.88 for Portuguese. Llama models only reached 0.73 for CASI and failed severely for German and Portuguese. Across LLMs, performance decreased from English to German and Portuguese processing languages. There was no evidence that additional document-level metadata had a significant effect. CONCLUSION: For English clinical narratives, acronym resolution by GPT-4 can be recommended to improve readability of clinical text by patients and professionals. For German and Portuguese, better models are needed. Llama models, which are particularly interesting for processing sensitive content on premise, cannot yet be recommended for acronym resolution.

Medical eponyms versus acronyms: what medical terminology is most beneficial to learn? A question of goals.

Appropriate use of medical terminology is one of the core conditions for successful communication in monolingual and multilingual healthcare communities. The modern scientific language is based on the descriptive terminology. However, it is often the case that the advantages of descriptive terminology are at odds with the ability to express complex concepts in just a few words. To solve this practicality problem it is customary to coin abbreviations and acronyms preferred to traditional eponyms. Today eponyms are considered ambiguous and non-descriptive, linked to the terminology of the past. The overview of this study demonstrates that the current habit of using acronyms can increase the scientific descriptive capacity compared to eponyms. On the other hand, acronyms remain ambiguous and more ephemeral than eponyms. Furthermore, eponyms are not as descriptive as acronyms, but they still carry important information for a medical student. If you truly believe in the importance of Medical Humanities in the medical curriculum, two aspects cannot be overlooked. First, eponyms bring students closer in an almost subliminal way to the history of medicine and the non-strictly technical-scientific field of medicine. Second, medicine is a complex science applied to humans and must strive to keep the patient at the center of its interests. Patients and their families preferably ask us to use eponyms. Which terminology to choose for medical students? The teachers have the last word.

Definition, acronyms, nomenclature, and classification of angioedema (DANCE): AAAAI, ACAAI, ACARE, and APAAACI DANCE consensus.

BACKGROUND: Angioedema (AE) manifests with intermittent, localized, self-limiting swelling of the subcutaneous and/or submucosal tissue. AE is heterogeneous, can be hereditary or acquired, may occur only once or be recurrent, may exhibit wheals or not, and may be due to mast cell mediators, bradykinin, or other mechanisms. Several different taxonomic systems are currently used, making it difficult to compare the results of studies, develop multicenter collaboration, and harmonize AE treatment. OBJECTIVE: We developed a consensus on the definition, acronyms, nomenclature, and classification of AE (DANCE). METHODS: The initiative involved 91 experts from 35 countries and was endorsed by 53 scientific and medical societies, and patient organizations. A consensus was reached by online discussion and voting using the Delphi process over a period of 16 months (June 2021 to November 2022). RESULTS: The DANCE initiative resulted in an international consensus on the definition, classification, and terminology of AE. The new consensus classification features 5 types and endotypes of AE and a harmonized vocabulary of abbreviations/acronyms. CONCLUSION: The DANCE classification complements current clinical guidelines and expert consensus recommendations on the diagnostic assessment and treatment of AE. DANCE does not replace current clinical guidelines, and expert consensus algorithms and should not be misconstrued in a way that affects reimbursement of medicines prescribed by physicians using sound clinical judgment. We anticipate that this new AE taxonomy and nomenclature will harmonize and facilitate AE research and clinical studies, thereby improving patient care.

[Handoffs in the intensive care unit]. / Übergaben auf der Intensivstation.

BACKGROUND: Effective handoffs in the intensive care unit (ICU) are key to patient safety. PURPOSE: This article aims to raise awareness of the significance of structured and thorough handoffs and highlights possible challenges as well as means for improvement. MATERIALS AND METHODS: Based on the available literature, the evidence regarding handoffs in ICUs is summarized and suggestions for practical implementation are derived. RESULTS: The quality of handoffs has an impact on patient safety. At the same time, communication in the intensive care setting is particularly challenging due to the complexity of cases, a disruptive work environment, and a multitude of inter- and intraprofessional interactions. Hierarchical team structures, deficiencies in feedback and error-management culture, (technical) language barriers in communication, as well as substantial physical and psychological stress may negatively influence the effectiveness of handoffs. Sets of interventions such as the implementation of checklists, mnemonics, and communication workshops contribute to a more structured and thorough handoff process and have the potential to significantly improve patient safety. CONCLUSION: Effective handoffs are the cornerstone of high-quality and safe patient care but face particular challenges in ICUs. Interventional measures such as structuring handoff concepts and periodic communication trainings can help to improve handoffs and thus increase patient safety.

The impact of COVID-19 vaccination prior to SARS-CoV-2 infection on prevalence of long COVID among a population-based probability sample of Michiganders, 2020-2022.

PURPOSE: To estimate the association between COVID-19 vaccination status at the time of COVID-19 onset and long COVID prevalence. METHODS: We used data from the Michigan COVID-19 Recovery Surveillance Study, a population-based probability sample of adults with COVID-19 (n = 4695). We considered 30-day and 90-day long COVID (illness duration ≥30 or ≥90 days, respectively), using Poisson regression to estimate prevalence ratios (PRs) comparing vaccinated (completed an initial series ≥14 days before COVID-19 onset) to unvaccinated individuals (received 0 doses before COVID-19 onset), accounting for differences in age, sex, race and ethnicity, education, employment, health insurance, and rurality/urbanicity. The full unvaccinated comparison group was further divided into historic and concurrent comparison groups based on timing of COVID-19 onset relative to vaccine availability. We used inverse probability of treatment weights to account for sociodemographic differences between groups. RESULTS: Compared to the full unvaccinated comparison group, the adjusted prevalence of 30-day and 90-day long COVID were lower among vaccinated individuals [PR30-day= 0.57(95%CI:0.49,0.66); PR90-day= 0.42(95%CI:0.34,0.53)]. Estimates were consistent across comparison groups (full, historic, and concurrent). CONCLUSIONS: Long COVID prevalence was 40-60% lower among adults vaccinated (vs. unvaccinated) prior to their COVID-19 onset. COVID-19 vaccination may be an important tool to reduce the burden of long COVID.

Are abbreviations and acronyms to describe hernia repair techniques overused and helpful? A proposal for rationalisation.

Surgeons use abbreviations and acronyms frequently to describe surgical techniques. Recent advances and innovations in repair of abdominal wall hernias, have given rise to a plenitude of novel acronyms. For each small deviation of an existing technique authors have proposed a novel acronym. Since an acronym is most of times not self-explaining it is often hard to understand literature, lectures, symposia programs and discussions in social media. Regularly, we discover different acronyms used for the same procedure and sometimes the same or similar acronyms are used for different techniques. A clear and non-ambivalent description of surgical techniques in the literature is most valuable to summarize scientific evidence in systematic reviews and meta-analyses. We would like to propose a more rational use of abbreviations to describe hernia repair techniques based on the type of access, type of hernia, mesh position, type of mesh used and type of mesh fixation.

Cell cycle control by the insulin-like growth factor signal: at the crossroad between cell growth and mitotic regulation.

In proliferating cells and tissues a number of checkpoints (G1/S and G2/M) preceding cell division (M-phase) require the signal provided by growth factors present in serum. IGFs (I and II) have been demonstrated to constitute key intrinsic components of the peptidic active fraction of mammalian serum. In vivo genetic ablation studies have shown that the cellular signal triggered by the IGFs through their cellular receptors represents a non-replaceable requirement for cell growth and cell cycle progression. Retroactive and current evaluation of published literature sheds light on the intracellular circuitry activated by these factors providing us with a better picture of the pleiotropic mechanistic actions by which IGFs regulate both cell size and mitogenesis under developmental growth as well as in malignant proliferation. The present work aims to summarize the cumulative knowledge learned from the IGF ligands/receptors and their intracellular signaling transducers towards control of cell size and cell-cycle with particular focus to their actionable circuits in human cancer. Furthermore, we bring novel perspectives on key functional discriminants of the IGF growth-mitogenic pathway allowing re-evaluation on some of its signal components based upon established evidences.

Toward standard abbreviations and acronyms for use in articles on aortic disease.

Objectives: Academic medical literature is fraught with complex article-specific acronyms and abbreviations that can impair communication and make reading arduous. Our goal is to ease frustration with bespoke, inconsistent, and variable sets of abbreviations that currently exist for common aorta-related terminology (eg, anatomy, imaging, disease, and therapy). We hope to ease reading and improve communication in the aortic sphere of cardiovascular literature. Methods: We reviewed a total of 205 published references related to aortic disease, including a systematic review of aorta-related articles in the Journal of Thoracic and Cardiovascular Surgery from the years 2020 and 2021. The array of variable definitions, abbreviations, and acronyms encountered in different papers that refer to the same terminology was striking, revealing that there were few standardized abbreviations in the aortic literature. We cataloged these terms, their associated abbreviations, and their frequency of use, and compiled a list of proposed standard abbreviations for commonly used terms that could be implemented uniformly in articles written about aortic diseases. Results: We present suggested acronyms and abbreviations for common terminology related to the aorta. It is anticipated that this standard list will evolve over time as the literature and technology of the field grows and develops. Conclusions: A proposed standard set of acronyms and abbreviations for aorta-related terminology is provided that, if found useful, could be implemented broadly in the aortic literature.

The impact of physical conditions on the incidence of major depressive disorder in Chinese university students: Results from a longitudinal study.

BACKGROUND: Major depressive disorder (MDD) is prevalent, and highly comorbid with physical illnesses. Few longitudinal studies have investigated the relationship between physical health conditions and MDD. The objectives of this study were to investigate the comorbid relationship between physical conditions and MDD, and the association between physical conditions and the 2-year risk of MDD. METHODS: A study was conducted in first-year Chinese university students (n = 8,079) over two and half years, using a longitudinal design. An adapted version of the Composite International Diagnostic Interview (CIDI - 3.0) was used to assess for MDD. The presence of physician diagnosed physical conditions was assessed using ten self-report questions. Cross-sectional and longitudinal associations between self-reported physical conditions and MDD were estimated, adjusting for possible confounders. RESULTS: The most frequently reported physical conditions were migraines, chronic rhinitis, and gastritis. We found that migraines, gastritis, and stomach ulcers were associated with a significantly higher lifetime prevalence of MDD than those without any physical health conditions. In those without a lifetime MDD, migraines, gastritis and stomach ulcers were also found to be significant predictors for 2-year risk of new onset MDD. LIMITATIONS: Recall and selection biases are possible when using self-reporting measures. Additionally, the COVID-19 outbreak impacted the response rate at the second follow-up assessment. Lastly, the severity of the physical conditions was not measured. CONCLUSIONS: Physical conditions and MDD are highly prevalent and comorbid in university students. Migraines, gastritis and stomach ulcers are associated with the risk of developing MDD. Future studies should further investigate how this information can be used to prevent MDD.

Prevalence of markers of atrial cardiomyopathy in embolic stroke of undetermined source: A systematic review.

BACKGROUND: Emerging evidence suggests the potential role of atrial cardiomyopathy (AC) as a direct thromboembolic determinant in embolic stroke of undetermined source (ESUS). OBJECTIVE: We aimed to quantify the prevalence of potential AC markers among ESUS, non-cardioembolic (NCE) and cardioembolic (CE) stroke patients. METHODS: PubMed and the Cochrane Central Register of Controlled Trials (CENTRAL) were searched for publications from inception to October 2021, with duplicate data extraction and risk of bias assessment. The Newcastle-Ottawa assessment scale was used to evaluate study quality. RESULTS: Among 398 screened studies, 11 observational studies with 2009 ESUS patients (mean age 66.5 years) fulfilled the inclusion criteria. Of electrocardiographic markers, increased P-wave terminal force in lead V1 was more prevalent in ESUS vs NCE (OR=2.26, 95%CI: 1.40-3.66). Of imaging markers, left atrial volume index (LAVI) and left atrial diameter (LAd) were higher in ESUS vs NCE (OR=1.04, 95%CI: 1.02-1.06 and OR=3.41, 95%CI: 1.35-8.61 respectively). Non-chicken wing morphology of the left atrial appendage was more frequent in ESUS compared to NCE patients in the majority of studies. Of serum biomarkers, the prevalence of NT-proBNP >250 pg/ml did not differ among ESUS vs NCE (OR=0.73, 95%CI: 0.39 -1.35). CONCLUSIONS: Electrocardiographic, echocardiographic markers and advanced imaging modalities able to assess the morphologic characteristics of left atrial appendage and left atrial function may be important tools to discriminate AC among ESUS vs NCE stroke patients. Prospective studies exploring the association of potential AC markers with ESUS occurrence are warranted to validate their clinical utility.

The common sense behind clinical trial names: An empirical study of trial acronyms.

The use of acronyms to name clinical trials, some of which might be manipulative or even coercive, is increasingly popular yet controversial. We aimed to evaluate whether trial acronyms are associated with appealing linguistic cues born of marketing psychology using trade names of perfumes. The proportion of trials (730 clinical trials) titled with an acronym was 61%. Among acronym-named trials, 70% have matching trade names of perfumes, i.e., - alluring names. Industry-sponsored trials were more likely to use acronyms (OR 1.61; 95% CI 1.15-2.26; p = 0.006) and alluring acronyms (OR 2.58; 95% CI 1.61-4.12; p < 0.001). During the period from 2000 to 2020, the proportion of alluring trials increased both for industry and academic funding, from 50% to 77% and from 0% to 57% , respectively. Also, trials with alluring acronyms were cited more often (relative rate of citation, 1.37; 95% CI 1.13-1.66; p = 0.001). The growing use of acronyms coincides with a noticeable increase in manipulative names. Through overt or subliminal enticement, inspirational acronyms that downplay the risks or raise expectations to patients with life-threatening illnesses, may exert undue influence. The observed relationship between manipulative acronyms and sponsorship by the pharmaceutical industry enhances this concern.

Choroidal thickness changes in healthcare professionals wearing surgical masks or FFP2 masks: Pilot study.

PURPOSE: To evaluate the choroidal thickness (CT) with enhanced depth-imaging optical coherence tomography (EDI-OCT) in healthcare professionals using surgical masks or FFP2 (N95) masks. METHODS: We included the 120 eyes of 120 healthy volunteers who were using a surgical mask (Group 1) or FFP2 mask (Group 2) in the study. Spectral domain (SD) OCT was used to measure CT. EDI-OCT was used to measure subfoveal and perifoveal CT. Points 1500 µm nasal (CN1500) and temporal (CT1500) to the foveal center were used to measure perifoveal CT. Oxygen saturation and heart rate were measured with a pulse oximeter. All measurements were performed at 8:30, before wearing the mask, and at 12:30, when the mask was removed for the lunch break. RESULTS: Of a total of 120 subjects, Group 1 consisted of 60 subjects (mean age 38.50±8.60 (range 24-44) years) and Group 2 also consisted of 60 subjects (mean age 36.60±6.53 (range 26-45) years). Although not statistically significant, CT was seen to have increased at 3 measurement points in Group 1 after using the mask for 4 h: subfoveal CT (CSF) (p = 0.545), CT1500 (p = 0.080), and CN1500 (p = 0.251)). In Group 2, the increase in CSF (p = 0.001) was statistically significant while the increases in CN1500 and CT1500 were not (p = 0.162 and p = 0.058, respectively) after using the mask for 4 h. CONCLUSION: We found CT to increase after 4 h of mask use, and this increase was more marked in Group 2. The increase in subfoveal CT in particular was statistically significant in Group 2.

Augmented words to improve a deep learning-based Indonesian syllabification.

Recent deep learning-based syllabification models generally give low error rates for high-resource languages with big datasets but sometimes produce high error rates for the low-resource ones. In this paper, two procedures: massive data augmentation and validation, are proposed to improve a deep learning-based syllabification, using a combination of bidirectional long short-term memory (BiLSTM), convolutional neural networks (CNN), and conditional random fields (CRF) for a low-resource Indonesian language. The massive data augmentation comprises four methods: transposing nuclei, swapping consonant-graphemes, flipping onsets, and creating acronyms. Meanwhile, the validation is implemented using a phonotactic-based scheme. A preliminary investigation on 50k Indonesian words informs that those augmentation methods significantly enlarge the dataset size by 12.8M valid words based on the phonotactic rules. An examination is then performed using 5-fold cross-validation. It reports that the augmentation methods significantly improve the BiLSTM-CNN-CRF model for 50k formal words and 100k named-entities datasets. A detailed investigation informs that augmenting the training set can reduce the word error rate (WER) coming from the long formal words and named entities.

Serous retinal detachment as a presenting sign of acute lymphoblastic leukemia: A case report and literature review.

PURPOSE: To describe a unique case of unilateral serous retinal detachment as the presenting sign of B-cell acute lymphoblastic leukemia (ALL). OBSERVATIONS: A 74 year old woman presented with right eye blurry vision and was found to have an underlying serous retinal detachment, along with cotton wool spots, inner retinal hemorrhages, and retinal pigment epithelial changes throughout her bilateral fundi. Fluorescein angiography demonstrated bilateral vasculitis and ultrasonography revealed asymmetric thickening and enhancement of the affected eyes' choroid. This prompted a systemic lab workup and results were suspicious for an underlying hematologic malignancy. The patient was admitted to the hospital for bone marrow biopsy confirming B-cell ALL, underwent intensive intravenous and intrathecal chemotherapy, and was discharged one month later. Follow up appointment in the ophthalmology clinic demonstrated functional and anatomic improvement in the serous retinal detachment and choroidal thickening suggestive of infiltration in her right eye. CONCLUSIONS: SRDs are an uncommon ocular manifestation of leukemia, and even less common as a presenting sign of the disease. A comprehensive literature review demonstrated 11 other cases reported worldwide. We present the first such case with additional findings of leukemic retinopathy, optic nerve and choroidal infiltration, and vasculitis, as well as a complete library of ophthalmic imaging from the patient's initial presentation. IMPORTANCE: A new diagnosis of serous retinal detachment(s) without any obvious cause should raise suspicion for leukemia and prompt further workup. Early recognition of this hematologic malignancy is crucial for prompt initiation of life-saving therapy.

Health technology assessment through the six sigma approach in abdominoplasty: Scalpel vs electrosurgery.

Abdominoplasty is a surgical procedure conducted to reduce excess abdominal skin and fat and improve body contouring. Despite being commonly performed, it is associated with a risk of complications such as infection, seroma, haematoma and wound dehiscence. To reduce the incidence of complications, different methods are used to create the abdominal flap, i.e., incision with a scalpel or electrosurgery. In this study, health technology assessment (HTA) using the Six Sigma methodology was conducted to compare these incision techniques in patients undergoing abdominoplasty. Two consecutively enroled groups of patients (33 in the scalpel group and 35 in the electrosurgery group) who underwent surgery at a single institution, the University of Campania "Luigi Vanvitelli", were analysed using the drain output as the main outcome for comparison of the incision techniques. While no difference was found regarding haematoma or seroma formation (no cases in either group), the main results also indicate a greater drain output (p-value<0.001) and a greater incidence of dehiscence (p-value=0.056) in patients whose incisions were made through electrosurgery. The combination of HTA and the Six Sigma methodology was useful to prove the possible advantages of creating skin incisions with a scalpel in full abdominoplasty, particularly a significant reduction in the total drain output and a reduction in wound healing problems, namely, wound dehiscence, when compared with electrosurgery, despite considering two limited and heterogeneous groups.

Harmonization of acronyms for volatile organic compound metabolites using a standardized naming system.

Increased interest in volatile organic compound (VOC) exposure has led to an increased need for consistent, systematic, and informative naming of VOC metabolites. As analytical methods have expanded to include many metabolites in a single assay, the number of acronyms in use for a single metabolite has expanded in an unplanned and inconsistent manner due to a lack of guidance or group consensus. Even though the measurement of VOC metabolites is a well-established means to investigate exposure to VOCs, a formal attempt to harmonize acronyms amongst investigators has not been published. The aim of this work is to establish a system of acronym naming that provides consistency in current acronym usage and a foundation for creating acronyms for future VOC metabolites.

Bosniak Classification version 2019: validation and comparison to original classification in pathologically confirmed cystic masses.

OBJECTIVE: To evaluate Bosniak Classification v2019 definitions in pathologically confirmed cystic renal masses. MATERIALS AND METHODS: Seventy-three cystic (≤ 25% solid) masses with histological confirmation (57 malignant, 16 benign) imaged by CT (N = 28) or CT+MRI (N = 56) between 2009 and 2019 were independently evaluated by three blinded radiologists using Bosniak v2019 and original classifications. Discrepancies were resolved by consensus with a fourth blinded radiologist. Overall class and v2019 features were compared to pathology. RESULTS: Inter-observer agreement was slightly improved comparing v2019 to Original Bosniak Classification (kappa = 0.26-0.47 versus 0.24-0.34 respectively). v2019 proportion of IIF and III masses (20.5% [15/73, 95% confidence interval (CI) 12.0-31.6%], 38.6% [28/73, 95% CI 27.2-50.5%]) differed from the original classification (6.8% [5/73, 95% CI 2.3-15.3%], 61.6% [45/73, 95% CI 49.5-72.8%]) with overlapping proportion of malignancy in each class. Mean septa number (7 ± 4 [range 1-10]) was not associated with malignancy (p = 0.89). Mean wall and septa thicknesses were 3 ± 3 (1-14) and 3 ± 2 (1-10) mm and higher in malignancies (p = 0.03 and 0.20 respectively). Areas under the receiver-operator-characteristic curve for wall and septa thickness were 0.66 (95% CI 0.54-0.79) and 0.61 (95% CI 0.45-0.78) with an optimal cut point of ≥ 3 mm (sensitivity 33.3%, specificity 86.7% and sensitivity 53%, specificity 73% respectively). Proportion of malignancy occurring in masses with the v2019 features "irregularity" (76.9% [10/13], 95% CI 46.2-94.9%) and "nodule" (89.7% [26/29], 95% CI 72.7-97.8%) overlapped. Angle of "nodule" (p = 0.27) was not associated with malignancy. CONCLUSION: Bosniak v2019 definitions for wall/septa thickness and protrusions are associated with malignancy. Overall, Bosniak v2019 categorizes a higher proportion of malignant masses in Class IIF with slight improvement in inter-observer agreement. KEY POINTS: • Considering Bosniak v2019 Class IIF cystic masses with many (≥ 4) smooth and thin septa, there was no association between the number of septa and malignancy (p = 0.89) in this study. • Increased cyst wall and septa thickness are associated with malignancy and a lower threshold of ≥ 3 mm maximized overall diagnostic accuracy compared to ≥ 4 mm threshold proposed for Bosniak v2019 Class 3. • An overlapping proportion of malignant masses is noted in Bosniak v2019 Class 3 masses with "irregularity" (76.9% [10/13], 95% CI 46.2-94.9%) compared to Bosniak v2019 Class 4 masses with "nodule" (89.7% [26/29], 95% CI 72.7-97.8%).

Assessment of contact tracing for COVID-19 among people experiencing homelessness, Salt Lake County Health Department, March-May 2020.

PURPOSE: Contact tracing is intended to reduce the spread of coronavirus disease 2019 (COVID-19), but it is difficult to conduct among people who live in congregate settings, including people experiencing homelessness (PEH). This analysis compares person-based contact tracing among two populations in Salt Lake County, Utah, from March-May 2020. METHODS: All laboratory-confirmed positive cases among PEH (n = 169) and documented in Utah's surveillance system were included in this analysis. The general population comparison group (n = 163) were systematically selected from all laboratory-confirmed cases identified during the same period. RESULTS: Ninety-three PEH cases (55%) were interviewed compared to 163 (100%) cases among the general population (P < .0001). PEH were more likely to be lost to follow-up at end of isolation (14.2%) versus the general population (0%; P-value < .0001) and provided fewer contacts per case (0.3) than the general population (4.7) (P-value < .0001). Contacts of PEH were more often unreachable (13.0% vs. 7.1%; P-value < .0001). CONCLUSIONS: These findings suggest that contact tracing among PEH should include a location-based approach, along with a person-based approach when resources allow, due to challenges in identifying, locating, and reaching cases among PEH and their contacts through person-based contact tracing efforts alone.

Acronyms and Opportunities for Improving Deep Nets.

Recently, several studies have reported promising results with BERT-like methods on acronym tasks. In this study, we find an older rule-based program, Ab3P, not only performs better, but error analysis suggests why. There is a well-known spelling convention in acronyms where each letter in the short form (SF) refers to "salient" letters in the long form (LF). The error analysis uses decision trees and logistic regression to show that there is an opportunity for many pre-trained models (BERT, T5, BioBert, BART, ERNIE) to take advantage of this spelling convention.