Contextual Language Model Research Articles

Automatic International Classification of Diseases Coding System: Deep Contextualized Language Model With Rule-Based Approaches.

BackgroundThe tenth revision of the International Classification of Diseases (ICD-10) is widely used for epidemiological research and health management. The clinical modification (CM) and procedure coding system (PCS) of ICD-10 were developed to describe more clinical details with increasing diagnosis and procedure codes and applied in disease-related groups for reimbursement. The expansion of codes made the coding time-consuming and less accurate. The state-of-the-art model using deep contextual word embeddings was used for automatic multilabel text classification of ICD-10. In addition to input discharge diagnoses (DD), the performance can be improved by appropriate preprocessing methods for the text from other document types, such as medical history, comorbidity and complication, surgical method, and special examination.ObjectiveThis study aims to establish a contextual language model with rule-based preprocessing methods to develop the model for ICD-10 multilabel classification.MethodsWe retrieved electronic health records from a medical center. We first compared different word embedding methods. Second, we compared the preprocessing methods using the best-performing embeddings. We compared biomedical bidirectional encoder representations from transformers (BioBERT), clinical generalized autoregressive pretraining for language understanding (Clinical XLNet), label tree-based attention-aware deep model for high-performance extreme multilabel text classification (AttentionXLM), and word-to-vector (Word2Vec) to predict ICD-10-CM. To compare different preprocessing methods for ICD-10-CM, we included DD, medical history, and comorbidity and complication as inputs. We compared the performance of ICD-10-CM prediction using different preprocesses, including definition training, external cause code removal, number conversion, and combination code filtering. For the ICD-10 PCS, the model was trained using different combinations of DD, surgical method, and key words of special examination. The micro F1 score and the micro area under the receiver operating characteristic curve were used to compare the model’s performance with that of different preprocessing methods.ResultsBioBERT had an F1 score of 0.701 and outperformed other models such as Clinical XLNet, AttentionXLM, and Word2Vec. For the ICD-10-CM, the model had an F1 score that significantly increased from 0.749 (95% CI 0.744-0.753) to 0.769 (95% CI 0.764-0.773) with the ICD-10 definition training, external cause code removal, number conversion, and combination code filter. For the ICD-10-PCS, the model had an F1 score that significantly increased from 0.670 (95% CI 0.663-0.678) to 0.726 (95% CI 0.719-0.732) with a combination of discharge diagnoses, surgical methods, and key words of special examination. With our preprocessing methods, the model had the highest area under the receiver operating characteristic curve of 0.853 (95% CI 0.849-0.855) and 0.831 (95% CI 0.827-0.834) for ICD-10-CM and ICD-10-PCS, respectively.ConclusionsThe performance of our model with the pretrained contextualized language model and rule-based preprocessing method is better than that of the state-of-the-art model for ICD-10-CM or ICD-10-PCS. This study highlights the importance of rule-based preprocessing methods based on coder coding rules.

Geoscience language models and their intrinsic evaluation

Geoscientists use observations and descriptions of the rock record to study the origins and history of our planet, which has resulted in a vast volume of scientific literature. Recent progress in natural language processing (NLP) has the potential to parse through and extract knowledge from unstructured text, but there has, so far, been only limited work on the concepts and vocabularies that are specific to geoscience. Herein we harvest and process public geoscientific reports (i.e., Canadian federal and provincial geological survey publications databases) and a subset of open access and peer-reviewed publications to train new, geoscience-specific language models to address that knowledge gap. Language model performance is validated using a series of new geoscience-specific NLP tasks (i.e., analogies, clustering, relatedness, and nearest neighbour analysis) that were developed as part of the current study. The raw and processed national geological survey corpora, language models, and evaluation criteria are all made public for the first time. We demonstrate that non-contextual (i.e., Global Vectors for Word Representation, GloVe) and contextual (i.e., Bidirectional Encoder Representations from Transformers, BERT) language models updated using the geoscientific corpora outperform the generic versions of these models for each of the evaluation criteria. Principal component analysis further demonstrates that word embeddings trained on geoscientific text capture meaningful semantic relationships, including rock classifications, mineral properties and compositions, and the geochemical behaviour of elements. Semantic relationships that emerge from the vector space have the potential to unlock latent knowledge within unstructured text, and perhaps more importantly, also highlight the potential for other downstream geoscience-focused NLP tasks (e.g., keyword prediction, document similarity, recommender systems, rock and mineral classification).

To Augment or Not to Augment? A Comparative Study on Text Augmentation Techniques for Low-Resource NLP

Abstract Data-hungry deep neural networks have established themselves as the de facto standard for many NLP tasks, including the traditional sequence tagging ones. Despite their state-of-the-art performance on high-resource languages, they still fall behind their statistical counterparts in low-resource scenarios. One methodology to counterattack this problem is text augmentation, that is, generating new synthetic training data points from existing data. Although NLP has recently witnessed several new textual augmentation techniques, the field still lacks a systematic performance analysis on a diverse set of languages and sequence tagging tasks. To fill this gap, we investigate three categories of text augmentation methodologies that perform changes on the syntax (e.g., cropping sub-sentences), token (e.g., random word insertion), and character (e.g., character swapping) levels. We systematically compare the methods on part-of-speech tagging, dependency parsing, and semantic role labeling for a diverse set of language families using various models, including the architectures that rely on pretrained multilingual contextualized language models such as mBERT. Augmentation most significantly improves dependency parsing, followed by part-of-speech tagging and semantic role labeling. We find the experimented techniques to be effective on morphologically rich languages in general rather than analytic languages such as Vietnamese. Our results suggest that the augmentation techniques can further improve over strong baselines based on mBERT, especially for dependency parsing. We identify the character-level methods as the most consistent performers, while synonym replacement and syntactic augmenters provide inconsistent improvements. Finally, we discuss that the results most heavily depend on the task, language pair (e.g., syntactic-level techniques mostly benefit higher-level tasks and morphologically richer languages), and model type (e.g., token-level augmentation provides significant improvements for BPE, while character-level ones give generally higher scores for char and mBERT based models).

ABNIRML: Analyzing the Behavior of Neural IR Models

Abstract Pretrained contextualized language models such as BERT and T5 have established a new state-of-the-art for ad-hoc search. However, it is not yet well understood why these methods are so effective, what makes some variants more effective than others, and what pitfalls they may have. We present a new comprehensive framework for Analyzing the Behavior of Neural IR ModeLs (ABNIRML), which includes new types of diagnostic probes that allow us to test several characteristics—such as writing styles, factuality, sensitivity to paraphrasing and word order—that are not addressed by previous techniques. To demonstrate the value of the framework, we conduct an extensive empirical study that yields insights into the factors that contribute to the neural model’s gains, and identify potential unintended biases the models exhibit. Some of our results confirm conventional wisdom, for example, that recent neural ranking models rely less on exact term overlap with the query, and instead leverage richer linguistic information, evidenced by their higher sensitivity to word and sentence order. Other results are more surprising, such as that some models (e.g., T5 and ColBERT) are biased towards factually correct (rather than simply relevant) texts. Further, some characteristics vary even for the same base language model, and other characteristics can appear due to random variations during model training.1

Evaluation of weakly-supervised methods for aspect extraction

Aspect-based sentiment analysis (ABSA) may provide more detailed information than general sentiment analysis. It aims to extract aspects from reviews and predict their polarities. In this paper, we focus on aspect extraction sub-task. We propose three weakly-supervised systems based on contextual language models and topic modeling. We evaluate and compare our systems on SemEval-2016 restaurant french benchmark. The experimental results reveal that our systems is quite competitive in aspect extraction from user reviews. We obtain 60.65% as F1 score with our best system. The latter outperforms the existing supervised ones. We deduct that weakly-supervised systems are efficient in terms of performance, time and human effort.

Adaptable Closed-Domain Question Answering Using Contextualized CNN-Attention Models and Question Expansion

In closed-domain Question Answering (QA), the goal is to retrieve answers to questions within a specific domain. The main challenge of closed-domain QA is to develop a model that only requires small datasets for training since large-scale corpora may not be available. One approach is a flexible QA model that can adapt to different closed domains and train on their corpora. In this paper, we present a novel versatile reading comprehension style approach for closed-domain QA (called CA-AcdQA). The approach is based on pre-trained contextualized language models, Convolutional Neural Network (CNN), and a self-attention mechanism. The model captures the relevance between the question and context sentences at different levels of granularity by exploring the dependencies between the features extracted by the CNN. Moreover, we include candidate answer identification and question expansion techniques for context reduction and rewriting ambiguous questions. The model can be tuned to different domains with a small training dataset for sentence-level QA. The approach is tested on four publicly-available closed-domain QA datasets: Tesla (person), California (region), EU-law (system), and COVID-QA (biomedical) against nine other QA approaches. Results show that the ALBERT model variant outperforms all approaches on all datasets with a significant increase in Exact Match and F1 score. Furthermore, for the Covid-19 QA in which the text is complicated and specialized, the model is improved considerably with additional biomedical training resources (an F1 increase of 15.9 over the next highest baseline).

Deep contextual multi-task feature fusion for enhanced concept, negation and speculation detection from clinical notes

Effective clinical decision support calls for precise detection of clinical entities such as diseases/disorders and associated assertions such as negation and speculation from clinical text. Contemporary approaches have relied on domain-specific Bidirectional Encoder Representations from Transformers (BERT) language models to achieve robust performance in clinical concept recognition. However, due to annotation scarcity, these approaches face a challenge in assertion detection. This study proposes a novel end-to-end neural model utilizing contextual features derived from a BERT ensemble, syntactic features derived from constituency parse trees, and multi-task learning for the enhanced detection of concept and assertion entities. Two clinical note benchmark datasets (n2c2 2010, n2c2 2012) were used for validating the proposed approach. Apart from achieving state-of-the-art performance in concept recognition (n2c2 2012), the proposed model significantly enhanced clinical note negation (+2.35 F1, McNemar’s test) and speculation (+5.26 F1) detection as compared to standalone transformer-based models. Assertion generalization improved by +2.23 F1, further reinforcing the effectiveness of the proposed strategy. Additionally, this study offers a generic methodology that integrates feature fusion, contextual language model ensembling, and multi-task learning to utilize transformer-based language models effectively.

A contextual multi-task neural approach to medication and adverse events identification from clinical text

Effective wide-scale pharmacovigilance calls for accurate named entity recognition (NER) of medication entities such as drugs, dosages, reasons, and adverse drug events (ADE) from clinical text. The scarcity of adverse event annotations and underlying semantic ambiguities make accurate scope identification challenging. The current research explores integrating contextualized language models and multi-task learning from diverse clinical NER datasets to mitigate this challenge. We propose a novel multi-task adaptation method to refine the embeddings generated by the Bidirectional Encoder Representations from Transformers (BERT) language model to improve inter-task knowledge sharing. We integrated the adapted BERT model into a unique hierarchical multi-task neural network comprised of the medication and auxiliary clinical NER tasks. We validated the model using two different versions of BERT on diverse well-studied clinical tasks: Medication and ADE (n2c2 2018/n2c2 2009), Clinical Concepts (n2c2 2010/n2c2 2012), Disorders (ShAReCLEF 2013). Overall medication extraction performance enhanced by up to +1.19 F1 (n2c2 2018) while generalization enhanced by +5.38 F1 (n2c2 2009) as compared to standalone BERT baselines. ADE recognition enhanced significantly (McNemar’s test), out-performing prior baselines. Similar benefits were observed on the auxiliary clinical and disorder tasks. We demonstrate that combining multi-dataset BERT adaptation and multi-task learning out-performs prior medication extraction methods without requiring additional features, newer training data, or ensembling. Taken together, the study contributes an initial case study towards integrating diverse clinical datasets in an end-to-end NER model for clinical decision support.

Does the magic of BERT apply to medical code assignment? A quantitative study

Unsupervised pretraining is an integral part of many natural language processing systems, and transfer learning with language models has achieved remarkable results in downstream tasks. In the clinical application of medical code assignment, diagnosis and procedure codes are inferred from lengthy clinical notes such as hospital discharge summaries. However, it is not clear if pretrained models are useful for medical code prediction without further architecture engineering. This paper conducts a comprehensive quantitative analysis of various contextualized language models' performances, pretrained in different domains, for medical code assignment from clinical notes. We propose a hierarchical fine-tuning architecture to capture interactions between distant words and adopt label-wise attention to exploit label information. Contrary to current trends, we demonstrate that a carefully trained classical CNN outperforms attention-based models on a MIMIC-III subset with frequent codes. Our empirical findings suggest directions for building robust medical code assignment models.

Automated quality assessment of cognitive behavioral therapy sessions through highly contextualized language representations.

During a psychotherapy session, the counselor typically adopts techniques which are codified along specific dimensions (e.g., 'displays warmth and confidence', or 'attempts to set up collaboration') to facilitate the evaluation of the session. Those constructs, traditionally scored by trained human raters, reflect the complex nature of psychotherapy and highly depend on the context of the interaction. Recent advances in deep contextualized language models offer an avenue for accurate in-domain linguistic representations which can lead to robust recognition and scoring of such psychotherapy-relevant behavioral constructs, and support quality assurance and supervision. In this work, we propose a BERT-based model for automatic behavioral scoring of a specific type of psychotherapy, called Cognitive Behavioral Therapy (CBT), where prior work is limited to frequency-based language features and/or short text excerpts which do not capture the unique elements involved in a spontaneous long conversational interaction. The model focuses on the classification of therapy sessions with respect to the overall score achieved on the widely-used Cognitive Therapy Rating Scale (CTRS), but is trained in a multi-task manner in order to achieve higher interpretability. BERT-based representations are further augmented with available therapy metadata, providing relevant non-linguistic context and leading to consistent performance improvements. We train and evaluate our models on a set of 1,118 real-world therapy sessions, recorded and automatically transcribed. Our best model achieves an F1 score equal to 72.61% on the binary classification task of low vs. high total CTRS.

Continual knowledge infusion into pre-trained biomedical language models.

Biomedical language models produce meaningful concept representations that are useful for a variety of biomedical natural language processing (bioNLP) applications such as named entity recognition, relationship extraction and question answering. Recent research trends have shown that the contextualized language models (e.g. BioBERT, BioELMo) possess tremendous representational power and are able to achieve impressive accuracy gains. However, these models are still unable to learn high-quality representations for concepts with low context information (i.e. rare words). Infusing the complementary information from knowledge-bases (KBs) is likely to be helpful when the corpus-specific information is insufficient to learn robust representations. Moreover, as the biomedical domain contains numerous KBs, it is imperative to develop approaches that can integrate the KBs in a continual fashion. We propose a new representation learning approach that progressively fuses the semantic information from multiple KBs into the pretrained biomedical language models. Since most of the KBs in the biomedical domain are expressed as parent-child hierarchies, we choose to model the hierarchical KBs and propose a new knowledge modeling strategy that encodes their topological properties at a granular level. Moreover, the proposed continual learning technique efficiently updates the concepts representations to accommodate the new knowledge while preserving the memory efficiency of contextualized language models. Altogether, the proposed approach generates knowledge-powered embeddings with high fidelity and learning efficiency. Extensive experiments conducted on bioNLP tasks validate the efficacy of the proposed approach and demonstrates its capability in generating robust concept representations.

Query bot for retrieving patients’ clinical history: A COVID-19 use-case

ObjectiveWith increasing patient complexity whose data are stored in fragmented health information systems, automated and time-efficient ways of gathering important information from the patients' medical history are needed for effective clinical decision making. Using COVID-19 as a case study, we developed a query-bot information retrieval system with user-feedback to allow clinicians to ask natural questions to retrieve data from patient notes. Materials and methodsWe applied clinicalBERT, a pre-trained contextual language model, to our dataset of patient notes to obtain sentence embeddings, using K-Means to reduce computation time for real-time interaction. Rocchio algorithm was then employed to incorporate user-feedback and improve retrieval performance. ResultsIn an iterative feedback loop experiment, MAP for final iteration was 0.93/0.94 as compared to initial MAP of 0.66/0.52 for generic and 1./1. compared to 0.79/0.83 for COVID-19 specific queries confirming that contextual model handles the ambiguity in natural language queries and feedback helps to improve retrieval performance. User-in-loop experiment also outperformed the automated pseudo relevance feedback method. Moreover, the null hypothesis which assumes identical precision between initial retrieval and relevance feedback was rejected with high statistical significance (p ≪ 0.05). Compared to Word2Vec, TF-IDF and bioBERT models, clinicalBERT works optimally considering the balance between response precision and user-feedback. DiscussionOur model works well for generic as well as COVID-19 specific queries. However, some generic queries are not answered as well as others because clustering reduces query performance and vague relations between queries and sentences are considered non-relevant. We also tested our model for queries with the same meaning but different expressions and demonstrated that these query variations yielded similar performance after incorporation of user-feedback. ConclusionIn conclusion, we develop an NLP-based query-bot that handles synonyms and natural language ambiguity in order to retrieve relevant information from the patient chart. User-feedback is critical to improve model performance.

How well do pre-trained contextual language representations recommend labels for GitHub issues?

Motivation:Open-source organizations use issues to collect user feedback, software bugs, and feature requests in GitHub. Many issues do not have labels, which makes labeling time-consuming work for the maintainers. Recently, some researchers used deep learning to improve the performance of automated tagging for software objects. However, these researches use static pre-trained word vectors that cannot represent the semantics of the same word in different contexts. Pre-trained contextual language representations have been shown to achieve outstanding performance on lots of NLP tasks. Description:In this paper, we study whether the pre-trained contextual language models are really better than other previous language models in the label recommendation for the GitHub labels scenario. We try to give some suggestions in fine-tuning pre-trained contextual language representation models. First, we compared four deep learning models, in which three of them use traditional pre-trained word embedding. Furthermore, we compare the performances when using different corpora for pre-training. Results:The experimental results show that: (1) When using large training data, the performance of BERT model is better than other deep learning language models such as Bi-LSTM, CNN and RCNN. While with a small size training data, CNN performs better than BERT. (2) Further pre-training on domain-specific data can indeed improve the performance of models. Conclusions:When recommending labels for issues in GitHub, using pre-trained contextual language representations is better if the training dataset is large enough. Moreover, we discuss the experimental results and provide some implications to improve label recommendation performance for GitHub issues.

Qualifying Certainty in Radiology Reports through Deep Learning-Based Natural Language Processing.

Communication gaps exist between radiologists and referring physicians in conveying diagnostic certainty. We aimed to explore deep learning-based bidirectional contextual language models for automatically assessing diagnostic certainty expressed in the radiology reports to facilitate the precision of communication. We randomly sampled 594 head MR imaging reports from an academic medical center. We asked 3 board-certified radiologists to read sentences from the Impression section and assign each sentence 1 of the 4 certainty categories: "Non-Definitive," "Definitive-Mild," "Definitive-Strong," "Other." Using the annotated 2352 sentences, we developed and validated a natural language-processing system based on the start-of-the-art bidirectional encoder representations from transformers (BERT), which can capture contextual uncertainty semantics beyond the lexicon level. Finally, we evaluated 3 BERT variant models and reported standard metrics including sensitivity, specificity, and area under the curve. A κ score of 0.74 was achieved for interannotator agreement on uncertainty interpretations among 3 radiologists. For the 3 BERT variant models, the biomedical variant (BioBERT) achieved the best macro-average area under the curve of 0.931 (compared with 0.928 for the BERT-base and 0.925 for the clinical variant [ClinicalBERT]) on the validation data. All 3 models yielded high macro-average specificity (93.13%-93.65%), while the BERT-base obtained the highest macro-average sensitivity of 79.46% (compared with 79.08% for BioBERT and 78.52% for ClinicalBERT). The BioBERT model showed great generalizability on the heldout test data with a macro-average sensitivity of 77.29%, specificity of 92.89%, and area under the curve of 0.93. A deep transfer learning model can be developed to reliably assess the level of uncertainty communicated in a radiology report.

Enhancing argumentation component classification using contextual language model

Arguments facilitate humans to deliver their ideas. The outcome of the discussion heavily relies on the validity of the argument. If an argument is well-composed, it is more effective to grasp the core idea behind the argument. To grade the argument, machines can be utilized by decomposing into semantic label components. In natural language processing, multiple language models are available to perform this task. It is divided into context-free and contextual models. The majority of previous studies used hand-crafted features to perform argument component classification, while state of the art language models utilize machine learning. The majority of these language models ignore the context in an argument. This research paper aims to analyze whether by including the context in the classification process may improve the accuracy of the language model which will enhance the argumentation mining process as well. The same document corpus is fed into several language models. Word2Vec and GLoVe represent the context free models, while BERT and ELMo as context sensitive language models. Accuracy and time from each model are then compared to determine the importance of context. The result shows that contextual language models are proven to be able to boost classification accuracy by approximately 20%. However, time comes as a cost where contextual models require longer training and prediction time. The benefit from the increase in accuracy outweighs the burden of time. Thus, as a contextual task, argumentation mining is suggested to use contextual model where context must be included to achieve promising results.

Investigating Gender Bias in BERT

In this work, we analyze the gender bias induced by BERT in downstream tasks. We also propose solutions to reduce gender bias. Contextual language models (CLMs) have pushed the NLP benchmarks to a new height. It has become a new norm to utilize CLM-provided word embeddings in downstream tasks such as text classification. However, unless addressed, CLMs are prone to learn intrinsic gender bias in the dataset. As a result, predictions of downstream NLP models can vary noticeably by varying gender words, such as replacing “he” to “she”, or even gender-neutral words. In this paper, we focus our analysis on a popular CLM, i.e., $$\text {BERT}$$ . We analyze the gender bias it induces in five downstream tasks related to emotion and sentiment intensity prediction. For each task, we train a simple regressor utilizing $$\text {BERT}$$ ’s word embeddings. We then evaluate the gender bias in regressors using an equity evaluation corpus. Ideally and from the specific design, the models should discard gender informative features from the input. However, the results show a significant dependence of the system’s predictions on gender-particular words and phrases. We claim that such biases can be reduced by removing gender-specific features from word embedding. Hence, for each layer in BERT, we identify directions that primarily encode gender information. The space formed by such directions is referred to as the gender subspace in the semantic space of word embeddings. We propose an algorithm that finds fine-grained gender directions, i.e., one primary direction for each BERT layer. This obviates the need of realizing gender subspace in multiple dimensions and prevents other crucial information from being omitted. Experiments show that removing embedding components in gender directions achieves great success in reducing BERT-induced bias in the downstream tasks. The investigation reveals significant gender bias a contextualized language model ( i.e., $$\text {BERT}$$ ) induces in downstream tasks. The proposed solution seems promising in reducing such biases.

Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences

In the field of artificial intelligence, a combination of scale in data and model capacity enabled by unsupervised learning has led to major advances in representation learning and statistical generation. In the life sciences, the anticipated growth of sequencing promises unprecedented data on natural sequence diversity. Protein language modeling at the scale of evolution is a logical step toward predictive and generative artificial intelligence for biology. To this end, we use unsupervised learning to train a deep contextual language model on 86 billion amino acids across 250 million protein sequences spanning evolutionary diversity. The resulting model contains information about biological properties in its representations. The representations are learned from sequence data alone. The learned representation space has a multiscale organization reflecting structure from the level of biochemical properties of amino acids to remote homology of proteins. Information about secondary and tertiary structure is encoded in the representations and can be identified by linear projections. Representation learning produces features that generalize across a range of applications, enabling state-of-the-art supervised prediction of mutational effect and secondary structure and improving state-of-the-art features for long-range contact prediction.

ELASPIC2 (EL2): Combining Contextualized Language Models and Graph Neural Networks to Predict Effects of Mutations

The ELASPIC web server allows users to evaluate the effect of mutations on protein folding and protein-protein interaction on a proteome-wide scale. It uses homology models of proteins and protein-protein interactions, which have been precalculated for several proteomes, and machine learning models, which integrate structural information with sequence conservation scores, in order to make its predictions. Since the original publication of the ELASPIC web server, several advances have motivated a revisiting of the problem of mutation effect prediction. First, progress in neural network architectures and self-supervised pre-trained has resulted in models which provide more informative embeddings of protein sequence and structure than those used by the original version of ELASPIC. Second, the amount of training data has increased several-fold, largely driven by advances in deep mutation scanning and other multiplexed assays of variant effect. Here, we describe two machine learning models which leverage the recent advances in order to achieve superior accuracy in predicting the effect of mutation on protein folding and protein-protein interaction. The models incorporate features generated using pre-trained transformer- and graph convolution-based neural networks, and are trained to optimize a ranking objective function, which permits the use of heterogeneous training data. The outputs from the new models have been incorporated into the ELASPIC web server, available at http://elaspic.kimlab.org.

Classical Arabic Named Entity Recognition Using Variant Deep Neural Network Architectures and BERT

Recurrent Neural Networks (RNNs) and transformers are deep learning models that have achieved remarkable success in several Natural Language Processing (NLP) tasks since they do not rely on handcrafted features nor enormous knowledge resources. Named Entity Recognition (NER) is an essential NLP task that is used in many applications such as information retrieval, question answering, and machine translation. NER aims to locate, extract, and classify named entities into predefined categories such as person, organization and location. Arabic NER is considered a challenging task because of the complexity and the unique characteristics of Arabic. Most of the previous research on deep learning based-Arabic NER focused on Modern Standard Arabic and Dialectal Arabic, which are different variations from Classical Arabic. In this paper, we investigate deep learning-based Classical Arabic NER using different deep neural network architectures and a BERT based contextual language model that is trained on general domain Arabic text. We propose two RNN-based models by fine-tunning the pretrained BERT language model to recognize and classify named entities from Classical Arabic. The pre-trained BERT contextual language model representations were used as input features to a BGRU/BLSTM model and were fine-tuned using a Classical Arabic NER dataset. In addition, we explore variant architectures of the proposed BERT-BGRU/BLSTM-CRF models. Experimentations showed that the BERT-BGRU-CRF model outperformed the other models by achieving an F-measure of 94.76% on the CANERCorpus. To the best of our knowledge, this is the first work that aims to recognize named entities in Classical Arabic using deep learning.

Unified Transformer Multi-Task Learning for Intent Classification With Entity Recognition

Intent classification (IC) and Named Entity Recognition (NER) are arguably the two main components needed to build a Natural Language Understanding (NLU) engine, which is a main component of conversational agents. The IC and NER components are closely intertwined and the entities are often connected to the underlying intent. Current research has primarily focused to model IC and NER as two separate units, which results in error propagation, and thus, sub-optimal performance. In this paper, we propose a simple yet effective novel framework for NLU where the parameters of the IC and the NER models are jointly trained in a consolidated parameter space. Text semantic representations are obtained from popular pre-trained contextual language models, which are fine-tuned for our task, and these parameters are propagated to other deep neural layers in our framework leading to a faithful unified modelling of the IC and NER parameters. The overall framework results in a faithful parameter sharing when the training is underway, leading to a more coherent learning. Experiments on two public datasets, ATIS and SNIPS, show that our model outperforms other methods by a noticeable margin. On the SNIPS dataset, we obtain a 1.42% improvement in NER in terms of the F1 score, and 1% improvement in intent accuracy score. On ATIS, we achieve 1.54% improvement in intent accuracy score. We also present qualitative results to showcase the effectiveness of our model.