Pretrained Transformer-based Language Model Research Articles

A transformer-based approach to Nigerian Pidgin text generation

AbstractThis paper describes the development of a transformer-based text generation model for Nigerian Pidgin also known as Naijá, a popular language in West Africa. Despite its wide use, Nigerian Pidgin remains under-resourced, particularly in areas related to text generation and natural language processing. These difficulties are primarily due to technological constraints rather than the language’s fundamental attributes. There is currently a demand for Nigerian Pidgin-specific solutions because it is used in everyday communication and has a unique linguistic blend. This paper aims to close this gap by exploring the application of state-of-the-art transformer technology to develop a text generation model for Nigerian Pidgin. This work uses the public Afriberta-corpus dataset to optimize the Generative Pre-trained Transformer (GPT-2) model across a sizeable dataset. The performance evaluators, BLEU and Perplexity metrics provide a detailed breakdown of the model’s text quality and predictive accuracy. Despite the difficulties caused by a limited amount of training data, preliminary evaluations show that the model can generate coherent Nigerian Pidgin text. The performance evaluation yielded perplexity scores of 43.56 for variable target reference length and 43.26 for fixed text length. BLEU scores of 0.15 for fixed max length and 0.56 for variable reference target length. This highlights the quality of generated text and the significant improvement when the generated text length is aligned with the reference target. Our work was benchmarked against African American Vernacular (AAVE) revealing that BLEU scores for AAVE are significantly lower than those for Standard American English, with BLEU given as 0.26. Our Nigerian Pidgin model, with a BLEU score of 0.56, shows a better performance. However, both results suggest that both dialects are challenging for language models. Leveraging the pre-trained transformer-based language model and evaluation metrics, we showcase the model’s capacity for coherent Nigerian Pidgin text generation. For future research, the research work can serve as a good foundation for advancement and progress in the Nigerian Pidgin language generation and other low-resource languages.

GPT-Based Model for Concise Summaries of Patient Health Information.

Although Patient Health Records (PHRs) are vital tools for patients, enabling them to access and manage health information, it remains challenging for doctors and patients to gather a swift overview of a patient's health status based on the extensive information included in the PHR. Our study introduces a generative pre-trained transformer-based language model to summarize health information documented in previously developed PHRs efficiently. By fine-tuning the model, we achieved results comparable to those of other studies in this domain, despite utilizing a smaller dataset. This data-to-text application represents a novel method that can be expected to promote enhanced information management in the medical field.

TASP: Topic-based abstractive summarization of Facebook text posts

Summarizing trending topics in large collections of Facebook posts is particularly relevant to profile social user activities and interests. However, automatically generating these summaries poses significant challenges due to the high heterogeneity of the input data, the limited fluency of extractive summaries, and the absence of abstractive summarization methods capable of handling multiple posts simultaneously. Existing abstractive models are either not suited to handle large post collections or disregard topic-level text relations. In this work, we present TASP, a novel tool for trending topic detection and summarization from English-written Facebook posts. It trains abstractive summarization models on multi-post collections by leveraging a shortlist of authoritative posts published by renowned newspapers. At inference time, TASP first creates clusters of semantically similar social posts, each one representing a distinct topic, using pre-trained transformer-based language models. Then, it generates abstractive summaries of the clusters for which authoritative information is missing. To the best of our knowledge, TASP is the first available tool suited to abstractive multi-post summarization. We test our approach on a large-scale dataset of real Facebook posts. The results show (1) The higher effectiveness of transformer-based approaches in generating topic-specific post clusters compared to traditional methods. (2) The importance of attending long pieces of text in multi-post abstractive summary generation.

Fairness-Aware Structured Pruning in Transformers

The increasing size of large language models (LLMs) has introduced challenges in their training and inference. Removing model components is perceived as a solution to tackle the large model sizes, however, existing pruning methods solely focus on performance, without considering an essential aspect for the responsible use of LLMs: model fairness. It is crucial to address the fairness of LLMs towards diverse groups, such as women, Black people, LGBTQ+, Jewish communities, among others, as they are being deployed and available to a wide audience. In this work, first, we investigate how attention heads impact fairness and performance in pre-trained transformer-based language models. We then propose a novel method to prune the attention heads that negatively impact fairness while retaining the heads critical for performance, i.e. language modeling capabilities. Our approach is practical in terms of time and resources, as it does not require fine-tuning the final pruned, and fairer, model. Our findings demonstrate a reduction in gender bias by 19%, 19.5%, 39.5%, 34.7%, 23%, and 8% for DistilGPT-2, GPT-2, GPT-Neo of two different sizes, GPT-J, and Llama 2 models, respectively, in comparison to the biased model, with only a slight decrease in performance. WARNING: This work uses language that is offensive in nature.

The Defeat of the Winograd Schema Challenge (Abstract Reprint)

The Winograd Schema Challenge—a set of twin sentences involving pronoun reference disambiguation that seem to require the use of commonsense knowledge—was proposed by Hector Levesque in 2011. By 2019, a number of AI systems, based on large pre-trained transformer-based language models and fine-tuned on these kinds of problems, achieved better than 90% accuracy. In this paper, we review the history of the Winograd Schema Challenge and discuss the lasting contributions of the flurry of research that has taken place on the WSC in the last decade. We discuss the significance of various datasets developed for WSC, and the research community's deeper understanding of the role of surrogate tasks in assessing the intelligence of an AI system.

Large-Scale Pretraining Improves Sample Efficiency of Active Learning-Based Virtual Screening.

Virtual screening of large compound libraries to identify potential hit candidates is one of the earliest steps in drug discovery. As the size of commercially available compound collections grows exponentially to the scale of billions, active learning and Bayesian optimization have recently been proven as effective methods of narrowing down the search space. An essential component of those methods is a surrogate machine learning model that predicts the desired properties of compounds. An accurate model can achieve high sample efficiency by finding hits with only a fraction of the entire library being virtually screened. In this study, we examined the performance of a pretrained transformer-based language model and graph neural network in a Bayesian optimization active learning framework. The best pretrained model identifies 58.97% of the top-50,000 compounds after screening only 0.6% of an ultralarge library containing 99.5 million compounds, improving 8% over the previous state-of-the-art baseline. Through extensive benchmarks, we show that the superior performance of pretrained models persists in both structure-based and ligand-based drug discovery. Pretrained models can serve as a boost to the accuracy and sample efficiency of active learning-based virtual screening.

Answer Agnostic Question Generation in Bangla Language

Question generation (QG) from a given context paragraph is a demanding task in natural language processing for its practical applications and prospects in various fields. Several studies have been conducted on QG in high-resource languages like English, however, very few have been done on resource-poor languages like Arabic and Bangla. In this work, we propose a finetuning method for QG that uses pre-trained transformer-based language models to generate questions from a given context paragraph in Bangla. Our approach is based on the idea that a transformer-based language model can be used to learn the relationships between words and phrases in a context paragraph which allows the models to generate questions that are both relevant and grammatically correct. We finetuned three different transformer models: (1) BanglaT5, (2) mT5-base, (3) BanglaGPT2, and demonstrated their capabilities using two different data formatting techniques: (1) AQL—All Question Per Line, (2) OQL—One Question Per Line, making it a total of six different variations of QG models. For each of these variants, six different decoding algorithms: (1) Greedy search, (2) Beam search, (3) Random Sampling, (4) Top K sampling, (5) Top- p Sampling, 6) a combination of Top K and Top-p Sampling were used to generate questions from the test dataset. For evaluation of the quality of questions generated using different models and decoding techniques, we also fine-tuned another transformer model BanglaBert on two custom datasets of our own and created two question classifier (QC) models that check the relevancy and Grammatical correctness of the questions generated by our QG models. The QC models showed test accuracy of 88.54% and 95.76% in the case of correctness and relevancy checks, respectively. Our results show that among all the variants of the QG, the mT5 OQL approach and beam decoding algorithm outperformed all the other ones in terms of relevancy (77%) and correctness (96%) with 36.60 Bleu_4, 48.98 METEOR, and 63.38 ROUGE-L scores.

Semantic-Driven Topic Modeling Using Transformer-Based Embeddings and Clustering Algorithms

Topic modeling is a powerful technique to discover hidden topics and patterns within a collection of documents without prior knowledge. Traditional topic modeling and clustering-based techniques encounter challenges in capturing contextual semantic information. This study introduces an innovative end-to-end semantic-driven topic modeling technique for the topic extraction process, utilizing advanced word and document embeddings combined with a powerful clustering algorithm. This semantic-driven approach represents a significant advancement in topic modeling methodologies. It leverages contextual semantic information to extract coherent and meaningful topics. Specifically, our model generates document embeddings using pre-trained transformer-based language models, reduces the dimensions of the embeddings, clusters the embeddings based on semantic similarity, and generates coherent topics for each cluster. Compared to ChatGPT and traditional topic modeling algorithms, our model provides more coherent and meaningful topics.

A Study of Vietnamese Sentiment Classification with Ensemble Pre-Trained Language Models

Sentiment Analysis (SA) has attracted increasing research attention in recent years. Most existing works tackle the SA task by fine-tuning single pre-trained language models combined with specific layers. Despite their effectiveness, the previous studies overlooked the utilization of feature representations from various contextual language models. Ensemble learning techniques have garnered increasing attention within the field of Natural Language Processing (NLP). However, there is still room for improvement in ensemble models for the SA task, particularly in the aspect-level SA task. Furthermore, the utilization of heterogeneous ensembles, which combine various pre-trained transformer-based language models, may prove beneficial in enhancing overall performance by incorporating diverse linguistic representations. This paper introduces two ensemble models that leverage soft voting and feature fusion techniques by combining individual pre-trained transformer-based language models for the SA task. The latest transformer-based models, including PhoBERT, XLM, XLM-Align, InfoXLM, and viBERT_FPT, are employed to integrate knowledge and representations using feature fusion and a soft voting strategy. We conducted extensive experiments on various Vietnamese benchmark datasets, encompassing sentence-level, document-level, and aspect-level SA. The experimental results demonstrate that our approaches outperform most existing methods, achieving new state-of-the-art results with F1-weighted scores of 94.03%, 95.65%, 75.36%, and 76.23% on the UIT_VSFC, Aivivn, UIT_ABSA for the restaurant domain, and UIT_ViSFD datasets, respectively.

Comparison of pretrained transformer-based models for influenza and COVID-19 detection using social media text data in Saskatchewan, Canada.

The use of social media data provides an opportunity to complement traditional influenza and COVID-19 surveillance methods for the detection and control of outbreaks and informing public health interventions. The first aim of this study is to investigate the degree to which Twitter users disclose health experiences related to influenza and COVID-19 that could be indicative of recent plausible influenza cases or symptomatic COVID-19 infections. Second, we seek to use the Twitter datasets to train and evaluate the classification performance of Bidirectional Encoder Representations from Transformers (BERT) and variant language models in the context of influenza and COVID-19 infection detection. We constructed two Twitter datasets using a keyword-based filtering approach on English-language tweets collected from December 2016 to December 2022 in Saskatchewan, Canada. The influenza-related dataset comprised tweets filtered with influenza-related keywords from December 13, 2016, to March 17, 2018, while the COVID-19 dataset comprised tweets filtered with COVID-19 symptom-related keywords from January 1, 2020, to June 22, 2021. The Twitter datasets were cleaned, and each tweet was annotated by at least two annotators as to whether it suggested recent plausible influenza cases or symptomatic COVID-19 cases. We then assessed the classification performance of pre-trained transformer-based language models, including BERT-base, BERT-large, RoBERTa-base, RoBERT-large, BERTweet-base, BERTweet-covid-base, BERTweet-large, and COVID-Twitter-BERT (CT-BERT) models, on each dataset. To address the notable class imbalance, we experimented with both oversampling and undersampling methods. The influenza dataset had 1129 out of 6444 (17.5%) tweets annotated as suggesting recent plausible influenza cases. The COVID-19 dataset had 924 out of 11939 (7.7%) tweets annotated as inferring recent plausible COVID-19 cases. When compared against other language models on the COVID-19 dataset, CT-BERT performed the best, supporting the highest scores for recall (94.8%), F1(94.4%), and accuracy (94.6%). For the influenza dataset, BERTweet models exhibited better performance. Our results also showed that applying data balancing techniques such as oversampling or undersampling method did not lead to improved model performance. Utilizing domain-specific language models for monitoring users' health experiences related to influenza and COVID-19 on social media shows improved classification performance and has the potential to supplement real-time disease surveillance.

A Multimodal Ensemble Machine Learning Approach to COVID-19 Misinformation Detection in Twitter

The emergence of social media platforms has unquestionably altered the manner in which people ingest information, with tweets now functioning as the primary source for news and other types of content. However, the proliferation of false news on these platforms has become a major concern, as it poses a severe threat to both individuals and society as a whole. Consequently, it is crucial to develop efficient methods for detecting false news in tweets. This study presents a novel hybrid approach that integrates the textual content of tweets with auxiliary features to detect false news. Our approach uses a pre-trained transformer-based language model, COVID-twitter-BERT to encode the text content of tweets into a dense representation that captures their meaning. The auxiliary features, such as sentiment score, credibility score, engagement score, average retweet count, average favourite count, and average followers of followers, are fed into a stacking classifier-based model to predict the trustworthiness score of the tweet. By combining the predictions of both models, we demonstrate that our approach outperforms baseline methods, emphasising the significance of utilising both text content and auxiliary features for Twitter false news detection. Our research considerably advances the field of detecting false news by demonstrating the effectiveness of integrating transformer-based language models and machine learning models for this task. Our findings provide valuable insights for improving the detection of false news on social media.

Arabic abstractive text summarization using RNN-based and transformer-based architectures

Recently, the Transformer model architecture and the pre-trained Transformer-based language models have shown impressive performance when used in solving both natural language understanding and text generation tasks. Nevertheless, there is little research done on using these models for text generation in Arabic. This research aims at leveraging and comparing the performance of different model architectures, including RNN-based and Transformer-based ones, and different pre-trained language models, including mBERT, AraBERT, AraGPT2, and AraT5 for Arabic abstractive summarization. We first built an Arabic summarization dataset of 84,764 high-quality text-summary pairs. To use mBERT and AraBERT in the context of text summarization, we employed a BERT2BERT-based encoder-decoder model where we initialized both the encoder and decoder with the respective model weights. The proposed models have been tested using ROUGE metrics and manual human evaluation. We also compared their performance on out-of-domain data. Our pre-trained Transformer-based models give a large improvement in performance with ∼79% less data. We found that AraT5 scores ∼3 ROUGE higher than a BERT2BERT-based model that is initialized with AraBERT, indicating that an encoder-decoder pre-trained Transformer is more suitable for summarizing Arabic text. Also, both of these two models perform better than AraGPT2 by a clear margin, which we found to produce summaries with high readability but with relatively lesser quality. On the other hand, we found that both AraT5 and AraGPT2 are better at summarizing out-of-domain text. We released our models and dataset publicly11https://huggingface.co/malmarjeh,.22https://data.mendeley.com/datasets/7kr75c9h24/1

One Clinician Is All You Need-Cardiac Magnetic Resonance Imaging Measurement Extraction: Deep Learning Algorithm Development.

BackgroundCardiac magnetic resonance imaging (CMR) is a powerful diagnostic modality that provides detailed quantitative assessment of cardiac anatomy and function. Automated extraction of CMR measurements from clinical reports that are typically stored as unstructured text in electronic health record systems would facilitate their use in research. Existing machine learning approaches either rely on large quantities of expert annotation or require the development of engineered rules that are time-consuming and are specific to the setting in which they were developed.ObjectiveWe hypothesize that the use of pretrained transformer-based language models may enable label-efficient numerical extraction from clinical text without the need for heuristics or large quantities of expert annotations. Here, we fine-tuned pretrained transformer-based language models on a small quantity of CMR annotations to extract 21 CMR measurements. We assessed the effect of clinical pretraining to reduce labeling needs and explored alternative representations of numerical inputs to improve performance.MethodsOur study sample comprised 99,252 patients that received longitudinal cardiology care in a multi-institutional health care system. There were 12,720 available CMR reports from 9280 patients. We adapted PRAnCER (Platform Enabling Rapid Annotation for Clinical Entity Recognition), an annotation tool for clinical text, to collect annotations from a study clinician on 370 reports. We experimented with 5 different representations of numerical quantities and several model weight initializations. We evaluated extraction performance using macroaveraged F1-scores across the measurements of interest. We applied the best-performing model to extract measurements from the remaining CMR reports in the study sample and evaluated established associations between selected extracted measures with clinical outcomes to demonstrate validity.ResultsAll combinations of weight initializations and numerical representations obtained excellent performance on the gold-standard test set, suggesting that transformer models fine-tuned on a small set of annotations can effectively extract numerical quantities. Our results further indicate that custom numerical representations did not appear to have a significant impact on extraction performance. The best-performing model achieved a macroaveraged F1-score of 0.957 across the evaluated CMR measurements (range 0.92 for the lowest-performing measure of left atrial anterior-posterior dimension to 1.0 for the highest-performing measures of left ventricular end systolic volume index and left ventricular end systolic diameter). Application of the best-performing model to the study cohort yielded 136,407 measurements from all available reports in the study sample. We observed expected associations between extracted left ventricular mass index, left ventricular ejection fraction, and right ventricular ejection fraction with clinical outcomes like atrial fibrillation, heart failure, and mortality.ConclusionsThis study demonstrated that a domain-agnostic pretrained transformer model is able to effectively extract quantitative clinical measurements from diagnostic reports with a relatively small number of gold-standard annotations. The proposed workflow may serve as a roadmap for other quantitative entity extraction.

Leveraging weak supervision to perform named entity recognition in electronic health records progress notes to identify the ophthalmology exam

ObjectiveTo develop deep learning models to recognize ophthalmic examination components from clinical notes in electronic health records (EHR) using a weak supervision approach. MethodsA corpus of 39,099 ophthalmology notes weakly labeled for 24 examination entities was assembled from the EHR of one academic center. Four pre-trained transformer-based language models (DistilBert, BioBert, BlueBert, and ClinicalBert) were fine-tuned to this named entity recognition task and compared to a baseline regular expression model. Models were evaluated on the weakly labeled test dataset, a human-labeled sample of that set, and a human-labeled independent dataset. ResultsOn the weakly labeled test set, all transformer-based models had recall > 0.93, with precision varying from 0.815 to 0.843. The baseline model had lower recall (0.769) and precision (0.682). On the human-annotated sample, the baseline model had high recall (0.962, 95 % CI 0.955–0.067) with variable precision across entities (0.081–0.999). Bert models had recall ranging from 0.771 to 0.831, and precision >=0.973. On the independent dataset, precision was 0.926 and recall 0.458 for BlueBert. The baseline model had better recall (0.708, 95 % CI 0.674–0.738) but worse precision (0.399, 95 % CI -0.352–0.451). ConclusionWe developed the first deep learning system to recognize eye examination components from clinical notes, leveraging a novel opportunity for weak supervision. Transformer-based models had high precision on human-annotated labels, whereas the baseline model had poor precision but higher recall. This system may be used to improve cohort and feature identification using free-text notes.Our weakly supervised approach may help amass large datasets of domain-specific entities from EHRs in many fields.

The Contribution of Lyrics and Acoustics to Collaborative Understanding of Mood

In this work, we study the association between song lyrics and mood through a data-driven analysis. Our data set consists of nearly one million songs, with song-mood associations derived from user playlists on the Spotify streaming platform. We take advantage of state-of-the-art natural language processing models based on transformers to learn the association between the lyrics and moods. We find that a pretrained transformer-based language model in a zero-shot setting -- i.e., out of the box with no further training on our data -- is powerful for capturing song-mood associations. Moreover, we illustrate that training on song-mood associations results in a highly accurate model that predicts these associations for unseen songs. Furthermore, by comparing the prediction of a model using lyrics with one using acoustic features, we observe that the relative importance of lyrics for mood prediction in comparison with acoustics depends on the specific mood. Finally, we verify if the models are capturing the same information about lyrics and acoustics as humans through an annotation task where we obtain human judgments of mood-song relevance based on lyrics and acoustics.

JointMatcher: Numerically-aware entity matching using pre-trained language models with attention concentration

Entity matching (EM) aims to identify whether two records refer to the same underlying real-world entity. Traditional entity matching methods mainly focus on structured data, where the attribute values are short and atomic. Recently, there has been an increasing demand for matching textual records, such as matching descriptions of products that correspond to long spans of text, which challenges the applications of these methods. Although a few deep learning (DL) solutions have been proposed, these solutions tend to “directly” use the DL techniques and treat the EM as NLP tasks without determining the unique demand for the EM task. Thus, the performance of these DL-based solutions is still far from satisfactory. In this paper, we present JointMatcher, a novel EM method based on the pre-trained Transformer-based language models so that the generated features of the textual records contain the context information. We realize that more attention paid to the similar segments and number-contained segments of the record pair is crucial for accurate matching. To integrate the high-contextualized features with the consideration of paying more attention to the similar segments and the number-contained segments, JointMatcher is equipped with the relevance-aware encoder and the numerically-aware encoder. Extensive experiments using structured and real-world textual datasets demonstrated that JointMatcher outperforms the previous state-of-the-art (SOTA) results without injecting any domain knowledge when small or medium size training sets are used.

Pre-trained transformer-based language models for Sundanese

The Sundanese language has over 32 million speakers worldwide, but the language has reaped little to no benefits from the recent advances in natural language understanding. Like other low-resource languages, the only alternative is to fine-tune existing multilingual models. In this paper, we pre-trained three monolingual Transformer-based language models on Sundanese data. When evaluated on a downstream text classification task, we found that most of our monolingual models outperformed larger multilingual models despite the smaller overall pre-training data. In the subsequent analyses, our models benefited strongly from the Sundanese pre-training corpus size and do not exhibit socially biased behavior. We released our models for other researchers and practitioners to use.

Unsupervised cross-lingual model transfer for named entity recognition with contextualized word representations.

Named entity recognition (NER) is one fundamental task in the natural language processing (NLP) community. Supervised neural network models based on contextualized word representations can achieve highly-competitive performance, which requires a large-scale manually-annotated corpus for training. While for the resource-scarce languages, the construction of such as corpus is always expensive and time-consuming. Thus, unsupervised cross-lingual transfer is one good solution to address the problem. In this work, we investigate the unsupervised cross-lingual NER with model transfer based on contextualized word representations, which greatly advances the cross-lingual NER performance. We study several model transfer settings of the unsupervised cross-lingual NER, including (1) different types of the pretrained transformer-based language models as input, (2) the exploration strategies of the multilingual contextualized word representations, and (3) multi-source adaption. In particular, we propose an adapter-based word representation method combining with parameter generation network (PGN) better to capture the relationship between the source and target languages. We conduct experiments on a benchmark ConLL dataset involving four languages to simulate the cross-lingual setting. Results show that we can obtain highly-competitive performance by cross-lingual model transfer. In particular, our proposed adapter-based PGN model can lead to significant improvements for cross-lingual NER.

Enhancing Transformer-based language models with commonsense representations for knowledge-driven machine comprehension

Compared to the traditional machine reading comprehension (MRC) with limitation to the information in a passage, knowledge-driven MRC tasks aim to enable models to answer the question according to text and related commonsense knowledge. Although pre-trained Transformer-based language models (TrLMs) such as BERT and Roberta, have shown powerful performance in MRC, external knowledge such as unspoken commonsense and world knowledge still cannot be used and explained explicitly. In this work, we present three simple yet effective injection methods integrated into the structure of TrLMs to fine-tune downstream knowledge-driven MRC tasks with off-the-shelf commonsense representations. Moreover, we introduce a mask mechanism for a token-level multi-hop relationship searching to filter external knowledge. We have conducted extensive experiments on DREAM and CosmosQA, two prevalent knowledge-driven datasets. Experimental results indicate that the incremental TrLMs have outperformed the baseline systems by 1%-4.1% with a fewer computational cost. Further analysis shows the effectiveness of the proposed methods and the robustness of the incremental model in the case of an incomplete training set.

Identification of Semantically Similar Sentences in Clinical Notes: Iterative Intermediate Training Using Multi-Task Learning.

BackgroundAlthough electronic health records (EHRs) have been widely adopted in health care, effective use of EHR data is often limited because of redundant information in clinical notes introduced by the use of templates and copy-paste during note generation. Thus, it is imperative to develop solutions that can condense information while retaining its value. A step in this direction is measuring the semantic similarity between clinical text snippets. To address this problem, we participated in the 2019 National NLP Clinical Challenges (n2c2)/Open Health Natural Language Processing Consortium (OHNLP) clinical semantic textual similarity (ClinicalSTS) shared task.ObjectiveThis study aims to improve the performance and robustness of semantic textual similarity in the clinical domain by leveraging manually labeled data from related tasks and contextualized embeddings from pretrained transformer-based language models.MethodsThe ClinicalSTS data set consists of 1642 pairs of deidentified clinical text snippets annotated in a continuous scale of 0-5, indicating degrees of semantic similarity. We developed an iterative intermediate training approach using multi-task learning (IIT-MTL), a multi-task training approach that employs iterative data set selection. We applied this process to bidirectional encoder representations from transformers on clinical text mining (ClinicalBERT), a pretrained domain-specific transformer-based language model, and fine-tuned the resulting model on the target ClinicalSTS task. We incrementally ensembled the output from applying IIT-MTL on ClinicalBERT with the output of other language models (bidirectional encoder representations from transformers for biomedical text mining [BioBERT], multi-task deep neural networks [MT-DNN], and robustly optimized BERT approach [RoBERTa]) and handcrafted features using regression-based learning algorithms. On the basis of these experiments, we adopted the top-performing configurations as our official submissions.ResultsOur system ranked first out of 87 submitted systems in the 2019 n2c2/OHNLP ClinicalSTS challenge, achieving state-of-the-art results with a Pearson correlation coefficient of 0.9010. This winning system was an ensembled model leveraging the output of IIT-MTL on ClinicalBERT with BioBERT, MT-DNN, and handcrafted medication features.ConclusionsThis study demonstrates that IIT-MTL is an effective way to leverage annotated data from related tasks to improve performance on a target task with a limited data set. This contribution opens new avenues of exploration for optimized data set selection to generate more robust and universal contextual representations of text in the clinical domain.