Contextualized Word Embeddings Research Articles

A Novel Knowledge-augmented Model Customization Approach for Arabic Offensive Language Detection

Multiple attempts to develop systems for detecting online Arabic offensive language have been explored in previous studies. However, most of these attempts do not consider the variation of Arabic dialects, cultures, and offensive phrases. In contrast, this study aims to extract knowledge from multiple offensive language datasets to build a cross-dialect and culture knowledge-based repository. This knowledge-based repository is utilized to develop novel system architecture based on customizing the AraBERT model in a unique method to preserve dialectal knowledge and offensive cultural knowledge within the contextual word embedding of BERT architecture. Performance evaluation procedures consist of statistical evaluation metrics and a behavioral checklist. Results report more effective predictions by the customized model than the uncustomized one, particularly for offensive text. The customization process allows the model to gain more knowledge of informal text in general, and a better understanding of dialectal Arabic.

Personalized Query Expansion with Contextual Word Embeddings

Personalized Query Expansion, the task of expanding queries with additional terms extracted from the user-related vocabulary, is a well-known solution to improve the retrieval performance of a system w.r.t. short queries. Recent approaches rely on word embeddings to select expansion terms from user-related texts. Although promising results have been delivered with former word embedding techniques, we argue that these methods are not suited for contextual word embeddings, which produce a unique vector representation for each term occurrence. In this article, we propose a Personalized Query Expansion method designed to solve the issues arising from the use of contextual word embeddings with the current Personalized Query Expansion approaches based on word embeddings. Specifically, we employ a clustering-based procedure to identify the terms that better represent the user interests and to improve the diversity of those selected for expansion, achieving improvements of up to 4% w.r.t. the best-performing baseline in terms of MAP@100. Moreover, our approach outperforms previous ones in terms of efficiency, allowing us to achieve sub-millisecond expansion times even in data-rich scenarios. Finally, we introduce a novel metric to evaluate the expansion terms’ diversity and empirically show the unsuitability of previous approaches based on word embeddings when employed along with contextual word embeddings, which cause the selection of semantically overlapping expansion terms.

Shahmukhi named entity recognition by using contextualized word embeddings

Named Entity Recognition (NER) is an imperative Natural Language Processing (NLP) task which intents to identify and classify predefined named entities in a given span of text. For many Western and Asian languages, NER is a systematically premeditated and established task, however, a little work has been done for Shahmukhi. This paper presents Shahmukhi NER with four key contributions. First, a Bi-directional Long-Short Term Memory (BiLSTM) network based NER model has been developed by incorporating various features including character and word embeddings and Part of Speech (POS) tagging. Second, transfer learning has been employed by training context-free Word2Vec and contextualized Embeddings from Language Models (ELMo) word representations. The word representations have been trained using a Shahmukhi corpus of 14.9 million words. Third, we prepared a cleaner version of an existing Shahmukhi NER corpus by performing Unicode normalization and tokenization errands. The corpus has been deduplicated and results are reported on an unseen evaluation set which produced valid results. Fourth, we have studied the impact of two annotation schemes; Inside-Outside (IO) and Inside-Outside-Beginning (IOB) for Shahmukhi. Transfer learning was quite helpful to enhance the performance of NER models especially ELMo embeddings significantly improved the results by prompting contextualized embedding vectors. This is the first study to use character embeddings, POS tagging and transfer learning for Shahmukhi named entity recognition. The IO scheme based model achieved an accuracy of 98.60% with an f-score of 83.75. The IOB scheme based model performed with an accuracy of 98.43% and an f-score of 75.55. These scores are quite promising for an under-resourced morphologically-rich language.

Natural language processing (NLP) aided qualitative method in health research

Qualitative data analysis is produced frequently in healthcare settings, which is a time-consuming and skilled analytic task. The use of qualitative research findings in clinical settings takes years, which is sometimes obsolete knowledge as the health context is dynamic. Artificial Intelligence (AI)-based qualitative data analysis might present with rapid analysis of text-based data in real-time, thereby empowering qualitative researchers to expedite their analysis and facilitate timely use of the research findings. We tested an AI-based method to complement the manual analysis of text-based data from the verbatim transcripts of seven mall managers’ interviews. First, we prepared text data into a machine-calculable format and employed BERT model to extract sentence-level features in our case. Second, we implement TF-IDF-based keywords mining techniques to extract the main candidate themes from the interview transcripts to support text-based analysis, including: 1) primary cluster detection algorithm, and 2) keyword extraction algorithm. The extracted core themes provide qualitative researchers with a more comprehensive overview of the qualitative data. Most of the sentences clustered in meaningful short topics or sentences carrying independent and clear information. The extracted topics and clustered sentences reduced qualitative researchers’ workload by condensing and identifying meaningful concepts and naming them. This method combining contextualized word embeddings, unsupervised clustering, and keyword extraction techniques can significantly reduce the overall workload and time consumed in qualitative research using conventional methods.

Multilabel Classification for Keyword Determination of Scientific Articles

In writing scientific articles, there are provisions regarding the structure or parts of writing that must be fulfilled. One part of the scientific article that must be included is keywords. The process of determining keywords manually can cause discrepancies with the specific themes discussed in the article. Thus, causing readers to be unable to reach the scientific article. The process of determining the keywords of scientific articles is determined automatically by the classification method. The classification process is carried out by determining the set of keywords owned by each scientific article data based on the abstract and title. Therefore, the classification process applied is multi-feature and multi-label. Classification is done by applying the Contextualized Word Embedding Method. The implementation of Contextualized Word Embedding Method is done by applying BERT Model. By applying the BERT Model, it is expected to provide good performance in determining the keywords of scientific articles. The evaluation results by applying the BERT Model to the case of multi-label classification on abstract data for keyword determination resulted in a loss value of Training Data is 0.514, loss value of Validation Data is 0.511, and an accuracy value of 0.71, a precision value of 0.71, a recall value of 0.71, an error value of 0.29 and f-1 score of 0.83. Based on the results of the evaluation, it shows that the BERT Classification Model can carry out a classification process to determine a set of keywords from each abstract data in scientific articles.

The Contribution of Selected Linguistic Markers for Unsupervised Arabic Verb Sense Disambiguation

Word sense disambiguation (WSD) is the task of automatically determining the meaning of a polysemous word in a specific context. Word sense induction is the unsupervised clustering of word usages in a different context to distinguish senses and perform unsupervised WSD. Most studies consider function words as stop words and delete them in the pre-processing step. However, function words can encode meaningful information that can help to improve the performance of WSD approaches. We propose in this work a novel approach to solve Arabic verb sense disambiguation that is based on a preposition-based classification that is used in an automatic word sense induction step to build sense inventories to disambiguate Arabic verbs. However, in the wake of the success of neural language models, recent works obtained encouraging results using BERT pre-trained models for English-language WSD approaches. Hence, we use contextualized word embeddings for an unsupervised Arabic WSD that is based on linguistic markers and uses sentence-BERT Transformer pre-trained models, which yields encouraging results that outperform other existing unsupervised neural AWSD approaches.

ADscreen: A speech processing-based screening system for automatic identification of patients with Alzheimer's disease and related dementia

Alzheimer's disease and related dementias (ADRD) present a looming public health crisis, affecting roughly 5 million people and 11 % of older adults in the United States. Despite nationwide efforts for timely diagnosis of patients with ADRD, >50 % of them are not diagnosed and unaware of their disease. To address this challenge, we developed ADscreen, an innovative speech-processing based ADRD screening algorithm for the protective identification of patients with ADRD. ADscreen consists of five major components: (i) noise reduction for reducing background noises from the audio-recorded patient speech, (ii) modeling the patient's ability in phonetic motor planning using acoustic parameters of the patient's voice, (iii) modeling the patient's ability in semantic and syntactic levels of language organization using linguistic parameters of the patient speech, (iv) extracting vocal and semantic psycholinguistic cues from the patient speech, and (v) building and evaluating the screening algorithm. To identify important speech parameters (features) associated with ADRD, we used the Joint Mutual Information Maximization (JMIM), an effective feature selection method for high dimensional, small sample size datasets. Modeling the relationship between speech parameters and the outcome variable (presence/absence of ADRD) was conducted using three different machine learning (ML) architectures with the capability of joining informative acoustic and linguistic with contextual word embedding vectors obtained from the DistilBERT (Bidirectional Encoder Representations from Transformers). We evaluated the performance of the ADscreen on an audio-recorded patients' speech (verbal description) for the Cookie-Theft picture description task, which is publicly available in the dementia databank. The joint fusion of acoustic and linguistic parameters with contextual word embedding vectors of DistilBERT achieved F1-score = 84.64 (standard deviation [std] = ±3.58) and AUC-ROC = 92.53 (std = ±3.34) for training dataset, and F1-score = 89.55 and AUC-ROC = 93.89 for the test dataset. In summary, ADscreen has a strong potential to be integrated with clinical workflow to address the need for an ADRD screening tool so that patients with cognitive impairment can receive appropriate and timely care.

A benchmark for evaluating Arabic contextualized word embedding models

Word embeddings, which represent words as numerical vectors in a high-dimensional space, are contextualized by generating a unique vector representation for each sense of a word based on the surrounding words and sentence structure. They are typically generated using such deep learning models as BERT and trained on large amounts of text data and using self-supervised learning techniques. Resulting embeddings are highly effective at capturing the nuances of language, and have been shown to significantly improve the performance of numerous NLP tasks. Word embeddings represent textual records of human thinking, with all the mental relations that we utilize to produce the succession of sentences that make up texts and discourses. Consequently, the distributed representation of words within embeddings ought to capture the reasoning relations that hold texts together. This paper makes its contribution to the field by proposing a benchmark for the assessment of contextualized word embeddings that probes into their capability for true contextualization by inspecting how well they capture resemblance, contrariety, comparability, identity, relations in time and space, causation, analogy, and sense disambiguation. The proposed metrics adopt a triangulation approach, so they use (1) Hume’s reasoning relations, (2) standard analogy, and (3) sense disambiguation. The benchmark has been evaluated against 22 Arabic contextualized embeddings and has proven to be capable of quantifying their differential performance in terms of these reasoning relations. Results of evaluation of the target embeddings revealed that they do take context into account and that they do reasonably well in sense disambiguation but have weakness in their identification of converseness, synonymy, complementarity, and analogy. Results also show that size of an embedding has diminishing returns because the highly frequent language patterns swamp low frequency patterns. Furthermore, the suggest that future research endeavors should not be concerned with the quantity of data as much as its quality, and that it should focus more on the representativeness of data, and on model architecture, design, and training.

ADEPT: A DEbiasing PrompT Framework

Several works have proven that finetuning is an applicable approach for debiasing contextualized word embeddings. Similarly, discrete prompts with semantic meanings have shown to be effective in debiasing tasks. With unfixed mathematical representation at the token level, continuous prompts usually surpass discrete ones at providing a pre-trained language model (PLM) with additional task-specific information. Despite this, relatively few efforts have been made to debias PLMs by prompt tuning with continuous prompts compared to its discrete counterpart. Furthermore, for most debiasing methods that alter a PLM's original parameters, a major problem is the need to not only decrease the bias in the PLM but also to ensure that the PLM does not lose its representation ability. Finetuning methods typically have a hard time maintaining this balance, as they tend to violently remove meanings of attribute words (like the words developing our concepts of "male" and "female" for gender), which also leads to an unstable and unpredictable training process. In this paper, we propose ADEPT, a method to debias PLMs using prompt tuning while maintaining the delicate balance between removing biases and ensuring representation ability. To achieve this, we propose a new training criterion inspired by manifold learning and equip it with an explicit debiasing term to optimize prompt tuning. In addition, we conduct several experiments with regard to the reliability, quality, and quantity of a previously proposed attribute training corpus in order to obtain a clearer prototype of a certain attribute, which indicates the attribute's position and relative distances to other words on the manifold. We evaluate ADEPT on several widely acknowledged debiasing benchmarks and downstream tasks, and find that it achieves competitive results while maintaining (and in some cases even improving) the PLM's representation ability. We further visualize words' correlation before and after debiasing a PLM, and give some possible explanations for the visible effects.

Methods of Annotating and Identifying Metaphors in the Field of Natural Language Processing

Metaphors are an integral and important part of human communication and greatly impact the way our thinking is formed and how we understand the world. The theory of the conceptual metaphor has shifted the focus of research from words to thinking, and also influenced research of the linguistic metaphor, which deals with the issue of how metaphors are expressed in language or speech. With the development of natural language processing over the past few decades, new methods and approaches to metaphor identification have been developed. The aim of the paper is to map the methods of annotating and identifying metaphors in the field of natural language processing and to give a systematic overview of how relevant linguistic theories and natural language processing intersect. The paper provides an outline of cognitive linguistic metaphor theory and an overview of relevant methods of annotating linguistic and conceptual metaphors as well as publicly available datasets. Identification methods are presented chronologically, from early approaches and hand-coded knowledge to statistical methods of machine learning and contemporary methods of using neural networks and contextual word embeddings.

Empowering machine learning models with contextual knowledge for enhancing the detection of eating disorders in social media posts

Social networks have become information dissemination channels, where announcements are posted frequently; they also serve as frameworks for debates in various areas (e.g., scientific, political, and social). In particular, in the health area, social networks represent a channel to communicate and disseminate novel treatments’ success; they also allow ordinary people to express their concerns about a disease or disorder. The Artificial Intelligence (AI) community has developed analytical methods to uncover and predict patterns from posts that enable it to explain news about a particular topic, e.g., mental disorders expressed as eating disorders or depression. Albeit potentially rich while expressing an idea or concern, posts are presented as short texts, preventing, thus, AI models from accurately encoding these posts’ contextual knowledge. We propose a hybrid approach where knowledge encoded in community-maintained knowledge graphs (e.g., Wikidata) is combined with deep learning to categorize social media posts using existing classification models. The proposed approach resorts to state-of-the-art named entity recognizers and linkers (e.g., Falcon 2.0) to extract entities in short posts and link them to concepts in knowledge graphs. Then, knowledge graph embeddings (KGEs) are utilized to compute latent representations of the extracted entities, which result in vector representations of the posts that encode these entities’ contextual knowledge extracted from the knowledge graphs. These KGEs are combined with contextualized word embeddings (e.g., BERT) to generate a context-based representation of the posts that empower prediction models. We apply our proposed approach in the health domain to detect whether a publication is related to an eating disorder (e.g., anorexia or bulimia) and uncover concepts within the discourse that could help healthcare providers diagnose this type of mental disorder. We evaluate our approach on a dataset of 2,000 tweets about eating disorders. Our experimental results suggest that combining contextual knowledge encoded in word embeddings with the one built from knowledge graphs increases the reliability of the predictive models. The ambition is that the proposed method can support health domain experts in discovering patterns that may forecast a mental disorder, enhancing early detection and more precise diagnosis towards personalized medicine.

A Comparative Analysis of Word Embedding and Deep Learning for Arabic Sentiment Classification

Sentiment analysis on social media platforms (i.e., Twitter or Facebook) has become an important tool to learn about users’ opinions and preferences. However, the accuracy of sentiment analysis is disrupted by the challenges of natural language processing (NLP). Recently, deep learning models have proved superior performance over statistical- and lexical-based approaches in NLP-related tasks. Word embedding is an important layer of deep learning models to generate input features. Many word embedding models have been presented for text representation of both classic and context-based word embeddings. In this paper, we present a comparative analysis to evaluate both classic and contextualized word embeddings for sentiment analysis. The four most frequently used word embedding techniques were used in their trained and pre-trained versions. The selected embedding represents classical and contextualized techniques. Classical word embedding includes algorithms such as GloVe, Word2vec, and FastText. By contrast, ARBERT is used as a contextualized embedding model. Since word embedding is more typically employed as the input layer in deep networks, we used deep learning architectures BiLSTM and CNN for sentiment classification. To achieve these goals, the experiments were applied to a series of benchmark datasets: HARD, Khooli, AJGT, ArSAS, and ASTD. Finally, a comparative analysis was conducted on the results obtained for the experimented models. Our outcomes indicate that, generally, generated embedding by one technique achieves higher performance than its pretrained version for the same technique by around 0.28 to 1.8% accuracy, 0.33 to 2.17% precision, and 0.44 to 2% recall. Moreover, the contextualized transformer-based embedding model BERT achieved the highest performance in its pretrained and trained versions. Additionally, the results indicate that BiLSTM outperforms CNN by approximately 2% in 3 datasets, HARD, Khooli, and ArSAS, while CNN achieved around 2% higher performance in the smaller datasets, AJGT and ASTD.

A comparative analysis of text representation, classification and clustering methods over real project proposals

PurposeWhen a large number of project proposals are evaluated to allocate available funds, grouping them based on their similarities is beneficial. Current approaches to group proposals are primarily based on manual matching of similar topics, discipline areas and keywords declared by project applicants. When the number of proposals increases, this task becomes complex and requires excessive time. This paper aims to demonstrate how to effectively use the rich information in the titles and abstracts of Turkish project proposals to group them automatically.Design/methodology/approachThis study proposes a model that effectively groups Turkish project proposals by combining word embedding, clustering and classification techniques. The proposed model uses FastText, BERT and term frequency/inverse document frequency (TF/IDF) word-embedding techniques to extract terms from the titles and abstracts of project proposals in Turkish. The extracted terms were grouped using both the clustering and classification techniques. Natural groups contained within the corpus were discovered using k-means, k-means++, k-medoids and agglomerative clustering algorithms. Additionally, this study employs classification approaches to predict the target class for each document in the corpus. To classify project proposals, various classifiers, including k-nearest neighbors (KNN), support vector machines (SVM), artificial neural networks (ANN), classification and regression trees (CART) and random forest (RF), are used. Empirical experiments were conducted to validate the effectiveness of the proposed method by using real data from the Istanbul Development Agency.FindingsThe results show that the generated word embeddings can effectively represent proposal texts as vectors, and can be used as inputs for clustering or classification algorithms. Using clustering algorithms, the document corpus is divided into five groups. In addition, the results demonstrate that the proposals can easily be categorized into predefined categories using classification algorithms. SVM-Linear achieved the highest prediction accuracy (89.2%) with the FastText word embedding method. A comparison of manual grouping with automatic classification and clustering results revealed that both classification and clustering techniques have a high success rate.Research limitations/implicationsThe proposed model automatically benefits from the rich information in project proposals and significantly reduces numerous time-consuming tasks that managers must perform manually. Thus, it eliminates the drawbacks of the current manual methods and yields significantly more accurate results. In the future, additional experiments should be conducted to validate the proposed method using data from other funding organizations.Originality/valueThis study presents the application of word embedding methods to effectively use the rich information in the titles and abstracts of Turkish project proposals. Existing research studies focus on the automatic grouping of proposals; traditional frequency-based word embedding methods are used for feature extraction methods to represent project proposals. Unlike previous research, this study employs two outperforming neural network-based textual feature extraction techniques to obtain terms representing the proposals: BERT as a contextual word embedding method and FastText as a static word embedding method. Moreover, to the best of our knowledge, there has been no research conducted on the grouping of project proposals in Turkish.

Data-driven dependency parsing of Vedic Sanskrit

This paper describes the first data-driven parser for Vedic Sanskrit, an ancient Indo-Aryan language in which a corpus of important religious and philosophical texts has been composed. We report and critically discuss experiments with the input feature representations, paying special attention to the performance of contextualized word embeddings and to the influence of morpho-syntactic representations on the parsing quality. In addition, we provide an in-depth discussion of the parsing errors that covers structural traits of the predicted trees as well as linguistic and extra-textual influence factors. In its optimal configuration, the proposed model achieves 87.61 unlabeled and 81.84 labeled attachment score on a held-out set of test sentences, demonstrating good performance for an under-resourced language.

Examining the effect of whitening on static and contextualized word embeddings

Static word embeddings (SWE) and contextualized word embeddings (CWE) are the foundation of modern natural language processing. However, these embeddings suffer from spatial bias in the form of anisotropy, which has been demonstrated to reduce their performance. A method to alleviate the anisotropy is the “whitening” transformation. Whitening is a standard method in signal processing and other areas, however, its effect on SWE and CWE is not well understood. In this study, we conduct an experiment to elucidate the effect of whitening on SWE and CWE. The results indicate that whitening predominantly removes the word frequency bias in SWE, and biases other than the word frequency bias in CWE.

Semantic Role Labeling for Amharic Text Using Multiple Embeddings and Deep Neural Network

Amharic is morphologically complex and under-resourced language, posing difficulties in the development of natural language processing applications. This paper presents the development of semantic role labeler for Amharic text using end-to-end deep neural network architecture. The system implicitly captures morphological, semantic and contextual features of a word at different levels of the architecture, and incorporates the syntactic structure of an input sentence. The proposed neural network architecture has four core layers from bottom to top, namely non-contextual word embedding, contextual word embedding, fully connected and sequence decoding layers. The non-contextual word embedding layer is formed from the concatenation of character-based, word-based and sentence-based word embeddings. This layer captures the morphological and semantic features of a given word by making use of BiLSTM recurrent neural network. At the contextual word embedding layer, a context sensitive embedding of a word is generated by applying a new LSTM layer on the top of the non-contextual concatenated word embedding layer. A fully connected network layer is added on top of contextual word embedding layer to supplement it by extracting dependencies among training samples in the corpus. At the sequence decoding layer, a sequence of semantic role labels is predicted using a linear-chain conditional random field algorithm by capturing the dependency among semantic role labels. In addition to the four core layers, the architecture has dropout layers to prevent overfitting problem. The proposed system achieves 94.96% accuracy and 81.2% F1 score when it is tested using test data.

The VNNLI - VLSP 2021: Leveraging Contextual Word Embedding for NLI Task on Bilingual Dataset

Natural Language Inference (NLI) is one of the critical tasks in natural language understanding which we take through the VLSP2021-NLI Shared Task competition. VLSP2021-NLI Shared Task is a competition to improve existing methods for NLI tasks, thereby enhancing the efficiency of applications. One of the challenges of the competition is the dataset in both Vietnamese and English. In this article, we report on evaluating the NLI task of the competition. We first implement the 5-fold cross-validation evaluation method. We following leverage model architectures pre-trained on cross-lingual language datasets such as XLM-RoBERTa and RemBERT to create contextual word embeddings for classification. Our final result reaches 90.00% on the test dataset of the organizers.

A Comprehensive Analysis of Transformer-Deep Neural Network Models in Twitter Disaster Detection

Social media platforms such as Twitter are a vital source of information during major events, such as natural disasters. Studies attempting to automatically detect textual communications have mostly focused on machine learning and deep learning algorithms. Recent evidence shows improvement in disaster detection models with the use of contextual word embedding techniques (i.e., transformers) that take the context of a word into consideration, unlike the traditional context-free techniques; however, studies regarding this model are scant. To this end, this paper investigates a selection of ensemble learning models by merging transformers with deep neural network algorithms to assess their performance in detecting informative and non-informative disaster-related Twitter communications. A total of 7613 tweets were used to train and test the models. Results indicate that the ensemble models consistently yield good performance results, with F-score values ranging between 76% and 80%. Simpler transformer variants, such as ELECTRA and Talking-Heads Attention, yielded comparable and superior results compared to the computationally expensive BERT, with F-scores ranging from 80% to 84%, especially when merged with Bi-LSTM. Our findings show that the newer and simpler transformers can be used effectively, with less computational costs, in detecting disaster-related Twitter communications.

Contextual word embeddings for tabular data search and integration

This paper presents a new approach to retrieve and further integrate tabular datasets (collections of rows and columns) using union and join operations. In this work, both processes were carried out using a similarity measure based on contextual word embeddings, which allows finding semantically similar tables and overcome the recall problem of lexical approaches based on string similarity. This work is the first attempt to use contextual word embeddings in the whole pipeline of table search and integration, including for the first time their use in the join operation. A comprehensive analysis of their performance was carried out on both retrieving and integrating tabular datasets, comparing them with context-free models. Column headings and cell values were used as contextual information and their impact on each task was evaluated. The results revealed that contextual models significantly outperform context-free models and a traditional weighting schema in ad hoc table retrieval. In the data integration task, contextual models also improved the results on union operation compared to context-free approaches.

Temporal disambiguation of relative temporal expressions in clinical texts.

Temporal expression recognition and normalization (TERN) is the foundation for all higher-level temporal reasoning tasks in natural language processing, such as timeline extraction, so it must be performed well to limit error propagation. Achieving new heights in state-of-the-art performance for TERN in clinical texts requires knowledge of where current systems struggle. In this work, we summarize the results of a detailed error analysis for three top performing state-of-the-art TERN systems that participated in the 2012 i2b2 Clinical Temporal Relation Challenge, and compare our own home-grown system Chrono to identify specific areas in need of improvement. Performance metrics and an error analysis reveal that all systems have reduced performance in normalization of relative temporal expressions, specifically in disambiguating temporal types and in the identification of the correct anchor time. To address the issue of temporal disambiguation we developed and integrated a module into Chrono that utilizes temporally fine-tuned contextual word embeddings to disambiguate relative temporal expressions. Chrono now achieves state-of-the-art performance for temporal disambiguation of relative temporal expressions in clinical text, and is the only TERN system to output dual annotations into both TimeML and SCATE schemes.