A GPT-2 Language Model for Biomedical Texts in Portuguese

Nowadays, a large portion of clinical data only exists in free text. The wide adoption of Electronic Health Records (EHRs) has enabled the increases in accessing to clinical documents, which provide challenges and opportunities for clinical Natural Language Processing (NLP) researchers. Given free-text clinical notes as input, an ideal system for clinical text understanding should have the ability to support clinical decisions. At corpus level, the system should recommend similar notes based on disease or patient types, and provide medication recommendation, or any other type of recommendations, based on patients' symptoms and other similar medical cases. At document level, it should return a list of important clinical concepts. Moreover, the system should be able to make diagnostic inferences over clinical concepts and output diagnosis. Unfortunately, current work has not systematically studied this system. This study focuses on developing and applying methods/techniques in different aspects of the system for clinical text understanding, at both corpus and document level. We deal with two major research questions: First, we explore the question of How to model the underlying relationships from clinical notes at corpus level? Documents clustering methods can group clinical notes into meaningful clusters, which can assist physicians and patients to understand medical conditions and diseases from clinical notes. We use Nonnegative Matrix Factorization (NMF) and Multi-view NMF to cluster clinical notes based on extracted medical concepts. The clustering results display latent patterns existed among clinical notes. Our method provides a feasible way to visualize a corpus of clinical documents. Based on extracted concepts, we further build a symptom-medication (Symp-Med) graph to model the Symp-Med relations in clinical notes corpus. We develop two Symp-Med matching algorithms to predict and recommend medications for patients based on their symptoms. Second, we want to solve the question of How to integrate structured knowledge with unstructured text to improve results for Clinical NLP tasks? On the one hand, the unstructured clinical text contains lots of information about medical conditions. On the other hand, structured Knowledge Bases (KBs) are frequently used for supporting clinical NLP tasks. We propose graph-regularized word embedding models to integrate knowledge from both KBs and free text. We evaluate our models on standard datasets and biomedical NLP tasks, and results showed encouraging improvements on both datasets. We further apply the graph-regularized word embedding models and present a novel approach to automatically infer the most probable diagnosis from a given clinical narrative.

Background: Syntactic analysis, or parsing, is a key task in natural language processing and a required component for many text mining approaches. In recent years, Universal Dependencies (UD) has emerged as the leading formalism for dependency parsing. While a number of recent tasks centering on UD have substantially advanced the state of the art in multilingual parsing, there has been only little study of parsing texts from specialized domains such as biomedicine.Methods: We explore the application of state-of-the-art neural dependency parsing methods to biomedical text using the recently introduced CRAFT-SA shared task dataset. The CRAFT-SA task broadly follows the UD representation and recent UD task conventions, allowing us to fine-tune the UD-compatible Turku Neural Parser and UDify neural parsers to the task. We further evaluate the effect of transfer learning using a broad selection of BERT models, including several models pre-trained specifically for biomedical text processing.Results: We find that recently introduced neural parsing technology is capable of generating highly accurate analyses of biomedical text, substantially improving on the best performance reported in the original CRAFT-SA shared task. We also find that initialization using a deep transfer learning model pre-trained on in-domain texts is key to maximizing the performance of the parsing methods.

The advances in deep learning have led to great achievements in many Natural Language Processing (NLP) tasks. With the nature of language, i.e., sequential data, most NLP tasks can be framed into the sequence learning framework, such as text generation. As one of the most important foundations for modern NLP techniques, autoregressive generation models have achieved dominant performance in a great deal of NLP tasks. Therefore, this thesis emphasizes improving the autoregressive generation model for different NLP tasks. While many tasks can naturally fit into the sequence learning framework, some of them, e.g., building discourse parsing tree, require sophisticated designs to fit into neural models. Therefore, this thesis firstly emphasizes a novel unified framework for discourse parsing, which builds a discourse tree in a top-down depth-first manner, and it frames the task as an autoregressive generation task with the goal of each step being the prediction of the node position given a piece of text. The proposed approach is proven effective with extensive empirical experiments. In addition, I extend the above framework by proposing a hierarchical decoder, which leverages the information from parents and siblings of the nodes that are currently processed. The proposed decoder utilizes the nature of the tree structure and further improves the experiment performance on both discourse parsing and dependency parsing tasks. On the other hand, the de facto strategies, i.e., cross entropy loss and teacher forcing, for training the autoregressive generation models have been shown problematic in certain aspects. For example, cross entropy loss, which is one of the widely leveraged training objective functions, often leads to text degeneration in text generation, and teacher forcing suffers from the exposure bias problem, where there exists a mismatch between the training and testing setup. For text degeneration, I introduce a class of diminishing attentions, which enforces the submodularity of the coverage calculated by cross attention in the sequence-to-sequence model. The proposed diminishing attentions achieve notable improvement on several neural text generation tasks, including text summarization, machine translation, and image paragraph generation. Further, I propose a novel training objective, ScaleGrad, to replace cross entropy, which significantly reduces the degeneration problem in different text generation tasks. In fact, ScaleGrad can be extended to problems beyond text degeneration. It provides wide flexibility to inject different inductive biases into the text generation model by directly modifying the gradient information in the output layer. Next, for the exposure bias problem, this thesis introduces a novel type of scheduled sampling based on training accuracy, which requires only minimal hyper-parameter tuning compared to existing scheduled sampling methods. Additionally, a novel imitation loss is proposed to further enforce the model’s generative behavior to match the teacher-forced behavior. Moreover, this thesis demonstrates that reducing exposure bias can improve the robustness of language models against repetition and toxic errors.

POS tagging is a basic Natural Language Processing (NLP) task that tags the words in an input text according to its grammatical values. Although POS Tagging is a fundamental application for very resourced languages, such as Limbu, is still unknown due to only few tagged datasets and linguistic resources. This research project uses deep learning techniques, transfer learning, and the BiLSTM-CRF model to develop an accurate POS-tagging system for the Limbu language. Using annotated and unannotated language data, we progress in achieving a small yet informative dataset of Limbu text. Skilled multilingual tutoring was modified to enhance success on low-resource language tests. The model as propose attains 90% accuracy, which is very much better than traditional rule-based and machine learning methods for Limbu POS tagging. The results indicate that deep learning methods can address linguistic issues facing low-resource languages even with limited data. In turn, this study provides a cornerstone for follow up NLP-based applications of Limbu and similar low-resource languages, demonstrating how deep learning can fill the gap where data is scarce.

BackgroundDisease name recognition is a fundamental task in clinical natural language processing, enabling the extraction of critical patient information from electronic health records. While recent advances in large language models (LLMs) have shown promise, most evaluations have focused on English, and little is known about their robustness in low-resource languages such as Japanese. In particular, whether these models can perform reliably on previously unseen in-hospital data, which differs from training data in writing styles and clinical contexts, has not been thoroughly investigated.ObjectiveThis study evaluated the robustness of fine-tuned LLMs for disease name recognition in Japanese clinical notes, with a particular focus on their performance on in-hospital data that was not included during training.MethodsWe used two corpora for this study: (1) a publicly available set of Japanese case reports denoted as CR, and (2) a newly constructed corpus of progress notes, denoted as PN, written by ten physicians to capture stylistic variations of in-hospital clinical notes. To reflect real-world deployment scenarios, we first fine-tuned models on CR. Specifically, we compared a LLM and a baseline-masked language model (MLM). These models were then evaluated under two conditions: (1) on CR, representing the in-domain (ID) setting with the same document type, similar to training, and (2) on PN, representing the out-of-domain (OOD) setting with a different document type. Robustness was assessed by calculating the performance gap (ie, the performance drop from in-domain to out-of-domain settings).ResultsThe LLM demonstrated greater robustness, with a smaller performance gap in F1-scores (ID–OOD = −8.6) compared to the MLM baseline performance (ID–OOD = −13.9). This indicated more stable performance across ID and OOD settings, highlighting the effectiveness of fine-tuned LLMs for reliable use in diverse clinical settings.ConclusionsFine-tuned LLMs demonstrate superior robustness for disease name recognition in Japanese clinical notes, with a smaller performance gap. These findings highlight the potential of LLMs as reliable tools for clinical natural language processing in low-resource language settings and support their deployment in real-world health care applications, where diversity in documentation is inevitable.

Trie-based rule processing for clinical NLP: A use-case study of n-trie, making the ConText algorithm more efficient and scalable.

Pretrained language models augmented with in-domain corpora show impressive results in biomedicine and clinical Natural Language Processing (NLP) tasks in English. However, there has been minimal work in low-resource languages. Although some pioneering works have shown promising results, many scenarios still need to be explored to engineer effective pretrained language models in biomedicine for low-resource settings. This study introduces the BioBERTurk family and four pretrained models in Turkish for biomedicine. To evaluate the models, we also introduced a labeled dataset to classify radiology reports of head CT examinations. Two parts of the reports, impressions and findings, are evaluated separately to observe the performance of models on longer and less informative text. We compared the models with the Turkish BERT (BERTurk) pretrained with general domain text, multilingual BERT (mBERT), and LSTM+attention-based baseline models. The first model initialized from BERTurk and then further pretrained with biomedical corpus performs statistically better than BERTurk, multilingual BERT, and baseline for both datasets. The second model continues to pretrain the BERTurk model by using only radiology Ph.D. theses to test the effect of task-related text. This model slightly outperformed all models on the impression dataset and showed that using only radiology-related data for continual pre-training could be effective. The third model continues to pretrain by adding radiology theses to the biomedical corpus but does not show a statistically meaningful difference for both datasets. The final model combines radiology and biomedicine corpora with the corpus of BERTurk and pretrains a BERT model from scratch. This model is the worst-performing model of the BioBERT family, even worse than BERTurk and multilingual BERT.

Word embeddings have become the predominant representation scheme on a token-level for various clinical natural language processing (NLP) tasks. More recently, character-level neural language models, exploiting recurrent neural networks, have again received attention, because they achieved similar performance against various NLP benchmarks. We investigated to what extent character-based language models can be applied to the clinical domain and whether they are able to capture reasonable lexical semantics using this maximally fine-grained representation scheme. We trained a long short-term memory network on an excerpt from a table of de-identified 50-character long problem list entries in German, each of which assigned to an ICD-10 code. We modelled the task as a time series of one-hot encoded single character inputs. After the training phase we accessed the top 10 most similar character-induced word embeddings related to a clinical concept via a nearest neighbour search and evaluated the expected interconnected semantics. Results showed that traceable semantics were captured on a syntactic level above single characters, addressing the idiosyncratic nature of clinical language. The results support recent work on general language modelling that raised the question whether token-based representation schemes are still necessary for specific NLP tasks.

Evaluation of clinical named entity recognition methods for Serbian electronic health records

Long Short-Term Memory (LSTM) networks are unique to exercise data in its memory cell with long-term memory as Natural Language Processing (NLP) tasks have inklings of intensive time and computational power due to their complex structures like magnitude language model Transformer required to pre-train and learn billions of data performing different NLP tasks. In this paper, a dynamic chaotic model is proposed for the objective of transforming neurons states in network with neural dynamic characteristics by restructuring LSTM as Chaotic Neural Oscillatory-Long-Short Term Memory (CNO-LSTM), where neurons in LSTM memory cells are weighed in substitutes by oscillatory neurons to speed up computational training of language model and improve text classification accuracy for real-world applications. From the implementation perspective, five popular datasets of general text classification including binary, multi classification and multi-label classification are used to compare with mainstream baseline models on NLP tasks. Results showed that the performance of CNO-LSTM, a simplified model structure and oscillatory neurons state in exercising different types of text classification tasks are above baseline models in terms of evaluation index such as Accuracy, Precision, Recall and F1. The main contributions are time reduction and improved accuracy. It achieved approximately 46.76% of the highest reduction training time and 2.55% accuracy compared with vanilla LSTM model. Further, it achieved approximately 35.86% in time reduction compared with attention model without oscillatory indicating that the model restructure has reduced GPU dependency to improve training accuracy.

BackgroundControlled vocabularies are fundamental resources for information extraction from clinical texts using natural language processing (NLP). Standard language resources available in the healthcare domain such as the UMLS metathesaurus or SNOMED CT are widely used for this purpose, but with limitations such as lexical ambiguity of clinical terms. However, most of them are unambiguous within text limited to a given clinical specialty. This is one rationale besides others to classify clinical text by the clinical specialty to which they belong.ResultsThis paper addresses this limitation by proposing and applying a method that automatically extracts Spanish medical terms classified and weighted per sub-domain, using Spanish MEDLINE titles and abstracts as input. The hypothesis is biomedical NLP tasks benefit from collections of domain terms that are specific to clinical subdomains. We use PubMed queries that generate sub-domain specific corpora from Spanish titles and abstracts, from which token n-grams are collected and metrics of relevance, discriminatory power, and broadness per sub-domain are computed. The generated term set, called Spanish core vocabulary about clinical specialties (SCOVACLIS), was made available to the scientific community and used in a text classification problem obtaining improvements of 6 percentage points in the F-measure compared to the baseline using Multilayer Perceptron, thus demonstrating the hypothesis that a specialized term set improves NLP tasks.ConclusionThe creation and validation of SCOVACLIS support the hypothesis that specific term sets reduce the level of ambiguity when compared to a specialty-independent and broad-scope vocabulary.

Text extraction is an important part of natural language processing (NLP) tasks. Most NLP tasks like text classification, machine translation, text-to-speech, text-based language identification, text summarization, and named-entity recognition involve the use of textual data. Such data is limited for low-resourced languages making it difficult to experiment advanced NLP techniques on these languages. This paper presents a Python-based toolkit for text analysis and text extraction from different types of images, documents, and audio files. The toolkit is built as a library that has functions that can be imported and utilized for text extraction.

Cross-lingual transfer, where a high-resource transfer language is used to improve the accuracy of a low-resource task language, is now an invaluable tool for improving performance of natural language processing (NLP) on low-resource languages. However, given a particular task language, it is not clear which language to transfer from, and the standard strategy is to select languages based on ad hoc criteria, usually the intuition of the experimenter. Since a large number of features contribute to the success of cross-lingual transfer (including phylogenetic similarity, typological properties, lexical overlap, or size of available data), even the most enlightened experimenter rarely considers all these factors for the particular task at hand. In this paper, we consider this task of automatically selecting optimal transfer languages as a ranking problem, and build models that consider the aforementioned features to perform this prediction. In experiments on representative NLP tasks, we demonstrate that our model predicts good transfer languages much better than ad hoc baselines considering single features in isolation, and glean insights on what features are most informative for each different NLP tasks, which may inform future ad hoc selection even without use of our method. Code, data, and pre-trained models are available at https://github.com/neulab/langrank

Chinese clinical named entity recognition via multi-head self-attention based BiLSTM-CRF

Dependency parsing is a fundmental task in natural language processing(NLP). Burmese belongs to a low resource language with a special language structure, therefore, it exists the problems with extremely lacking of high quality data for Burmese dependency parsing and the inaccurate representation of semantic. We propose a Burmese dependency parsing model based on transfer learning, our method generate partially accurate Burmese dependency parsing data by constructing the relationship of English-Burmese. The embedding of Burmese represented by syllables and words to obtain accurate bilingual word vectors representation of English-Burmese. To verify the effectiveness of our method, during the training process, we fuse the dependency parsing data of Burmese and English, which transfer the dependency arc and POS tagging of English to Burmese. The experimental results show that our proposed method has a UAS value of 44.10% and a LAS value of 30.01% on Burmese datadset.