Pre-trained Masked Language Model Research Articles

Clinical information extraction for lower-resource languages and domains with few-shot learning using pretrained language models and prompting

Abstract A vast amount of clinical data are still stored in unstructured text. Automatic extraction of medical information from these data poses several challenges: high costs of clinical expertise, restricted computational resources, strict privacy regulations, and limited interpretability of model predictions. Recent domain adaptation and prompting methods using lightweight masked language models showed promising results with minimal training data and allow for application of well-established interpretability methods. We are first to present a systematic evaluation of advanced domain-adaptation and prompting methods in a lower-resource medical domain task, performing multi-class section classification on German doctor’s letters. We evaluate a variety of models, model sizes (further-pre)training and task settings, and conduct extensive class-wise evaluations supported by Shapley values to validate the quality of small-scale training data and to ensure interpretability of model predictions. We show that in few-shot learning scenarios, a lightweight, domain-adapted pretrained language model, prompted with just 20 shots per section class, outperforms a traditional classification model, by increasing accuracy from $48.6\%$ to $79.1\%$ . By using Shapley values for model selection and training data optimization, we could further increase accuracy up to $84.3\%$ . Our analyses reveal that pretraining of masked language models on general-language data is important to support successful domain-transfer to medical language, so that further-pretraining of general-language models on domain-specific documents can outperform models pretrained on domain-specific data only. Our evaluations show that applying prompting based on general-language pretrained masked language models combined with further-pretraining on medical-domain data achieves significant improvements in accuracy beyond traditional models with minimal training data. Further performance improvements and interpretability of results can be achieved, using interpretability methods such as Shapley values. Our findings highlight the feasibility of deploying powerful machine learning methods in clinical settings and can serve as a process-oriented guideline for lower-resource languages and domains such as clinical information extraction projects.

Scoring Multi-hop Question Decomposition Using Masked Language Models

Question answering (QA) is a sub-field of Natural Language Processing (NLP) that focuses on developing systems capable of answering natural language queries. Within this domain, multi-hop question answering represents an advanced QA task that requires gathering and reasoning over multiple pieces of information from diverse sources or passages. To handle the complexity of multi-hop questions, question decomposition has been proven to be a valuable approach. This technique involves breaking down complex questions into simpler sub-questions, reducing the complexity of the problem. However, it’s worth noting that existing question decomposition methods often rely on training data, which may not always be readily available for low-resource languages or specialized domains. To address this issue, we propose a novel approach that utilizes pre-trained masked language models to score decomposition candidates in a zero-shot manner. The method involves generating decomposition candidates, scoring them using a pseudo-log likelihood estimation, and ranking them based on their scores. To evaluate the efficacy of the decomposition process, we conducted experiments on two datasets annotated on decomposition in two different languages, Arabic and English. Subsequently, we integrated our approach into a complete QA system and conducted a reading comprehension performance evaluation on the HotpotQA dataset. The obtained results emphasize that while the system exhibited a small drop in performance, it still maintained a significant advance compared to the baseline model. The proposed approach highlights the efficiency of the language model scoring technique in complex reasoning tasks such as multi-hop question decomposition.

Enriching Urdu NER with BERT Embedding, Data Augmentation, and Hybrid Encoder-CNN Architecture

Named Entity Recognition (NER) is an indispensable component of Natural Language Processing (NLP), which aims to identify and classify entities within text data. While Deep Learning (DL) models have excelled in NER for well-resourced languages such as English, Spanish, and Chinese, they face significant hurdles when dealing with low-resource languages such as Urdu. These challenges stem from the intricate linguistic characteristics of Urdu, including morphological diversity, a context-dependent lexicon, and the scarcity of training data. This study addresses these issues by focusing on Urdu Named Entity Recognition (U-NER) and introducing three key contributions. First, various pre-trained embedding methods are employed, encompassing Word2vec (W2V), GloVe, FastText, Bidirectional Encoder Representations from Transformers (BERT), and Embeddings from language models (ELMo). In particular, fine-tuning is performed on BERT BASE and ELMo using Urdu Wikipedia and news articles. Second, a novel generative Data Augmentation (DA) technique replaces Named Entities (NEs) with mask tokens, employing pre-trained masked language models to predict masked tokens, effectively expanding the training dataset. Finally, the study introduces a novel hybrid model combining a Transformer Encoder with a Convolutional Neural Network (CNN) to capture the intricate morphology of Urdu. These modules enable the model to handle polysemy, extract short- and long-range dependencies, and enhance learning capacity. Empirical experiments demonstrate that the proposed model, incorporating BERT embeddings and an innovative DA approach, attains the highest F1-score of 93.99%, highlighting its efficacy for the U-NER task.

Few-shot biomedical named entity recognition via knowledge-guided instance generation and prompt contrastive learning.

Few-shot learning (FSL) that can effectively perform named entity recognition in low-resource scenarios has raised growing attention, but it has not been widely studied yet in the biomedical field. In contrast to high-resource domains, biomedical named entity recognition (BioNER) often encounters limited human-labeled data in real-world scenarios, leading to poor generalization performance when training only a few labeled instances. Recent approaches either leverage cross-domain high-resource data or fine-tune the pre-trained masked language model using limited labeled samples to generate new synthetic data, which is easily stuck in domain shift problems or yields low-quality synthetic data. Therefore, in this paper, we study a more realistic scenario, i.e., few-shot learning for BioNER. Leveraging the domain knowledge graph, we propose knowledge-guided instance generation for few-shot BioNER, which generates diverse and novel entities based on similar semantic relations of neighbor nodes. In addition, by introducing question prompt, we cast BioNER as question answering (QA) task and propose prompt contrastive learning to improve the robustness of the model by measuring the mutual information (MI) between query-answer pairs. Extensive experiments conducted on various few-shot settings show that the proposed framework achieves superior performance. Particularly, in a low-resource scenario with only 20 samples, our approach substantially outperforms recent state-of-the-art (SoTA) models on four benchmark datasets, achieving an average improvement of up to 7.1% F1. Our source code and data are available at https://github.com/cpmss521/KGPC. Supplementary data are available at Bioinformatics online.

Speciesist language and nonhuman animal bias in English Masked Language Models

Warning: This paper contains examples of offensive language, including insulting or objectifying expressions.Various existing studies have analyzed what social biases are inherited by NLP models. These biases may directly or indirectly harm people, therefore previous studies have focused only on human attributes. However, until recently no research on social biases in NLP regarding nonhumans existed. In this paper,11Anonymous previous version of this paper is accessible at https://openreview.net/forum?id=dfqMpjZOgv4, we have also published a paper on this topic in Japanese (Takeshita et al., 2021). we analyze biases to nonhuman animals, i.e. speciesist bias, inherent in English Masked Language Models such as BERT. We analyzed speciesist bias against 46 animal names using template-based and corpus-extracted sentences containing speciesist (or non-speciesist) language. We found that pre-trained masked language models tend to associate harmful words with nonhuman animals and have a bias toward using speciesist language for some nonhuman animal names. Our code for reproducing the experiments will be made available on GitHub.22https://github.com/Language-Media-Lab/speciesist-language.

InfoLM: A New Metric to Evaluate Summarization & Data2Text Generation

Assessing the quality of natural language generation (NLG) systems through human annotation is very expensive. Additionally, human annotation campaigns are time-consuming and include non-reusable human labour. In practice, researchers rely on automatic metrics as a proxy of quality. In the last decade, many string-based metrics (e.g., BLEU or ROUGE) have been introduced. However, such metrics usually rely on exact matches and thus, do not robustly handle synonyms. In this paper, we introduce InfoLM a family of untrained metrics that can be viewed as a string-based metric that addresses the aforementioned flaws thanks to a pre-trained masked language model. This family of metrics also makes use of information measures allowing the possibility to adapt InfoLM to different evaluation criteria. Using direct assessment, we demonstrate that InfoLM achieves statistically significant improvement and two figure correlation gains in many configurations compared to existing metrics on both summarization and data2text generation tasks.

Measuring Media Bias via Masked Language Modeling

Bias in news reporting can lead to tribalism and division on important issues. Scalable and reliable measurement of such biases is an important first step in addressing them. In this work, based on the intuition that media bias is captured by the tone and word choices in articles, we propose a framework for modeling the relative bias of media outlets through masked token prediction via large-scale pretrained masked language models fine-tuned on articles form news outlets. Through experiments on five diverse and politically polarized topics we show that our framework can capture media bias towards these topics with high reliability. Additionally, our experiments show that our framework is general, in that language models fine-tuned on one topic can be applied to other topics with little drop in performance.

Neural entity linking: A survey of models based on deep learning

This survey presents a comprehensive description of recent neural entity linking (EL) systems developed since 2015 as a result of the “deep learning revolution” in natural language processing. Its goal is to systemize design features of neural entity linking systems and compare their performance to the remarkable classic methods on common benchmarks. This work distills a generic architecture of a neural EL system and discusses its components, such as candidate generation, mention-context encoding, and entity ranking, summarizing prominent methods for each of them. The vast variety of modifications of this general architecture are grouped by several common themes: joint entity mention detection and disambiguation, models for global linking, domain-independent techniques including zero-shot and distant supervision methods, and cross-lingual approaches. Since many neural models take advantage of entity and mention/context embeddings to represent their meaning, this work also overviews prominent entity embedding techniques. Finally, the survey touches on applications of entity linking, focusing on the recently emerged use-case of enhancing deep pre-trained masked language models based on the Transformer architecture.

Towards Semantics-Enhanced Pre-Training: Can Lexicon Definitions Help Learning Sentence Meanings?

Self-supervised pre-training techniques, albeit relying on large amounts of text, have enabled rapid growth in learning language representations for natural language understanding. However, as radically empirical models on sentences, they are subject to the input data distribution, inevitably incorporating data bias and reporting bias, which may lead to inaccurate understanding of sentences. To address this problem, we propose to adopt a human learner's approach: when we cannot make sense of a word in a sentence, we often consult the dictionary for specific meanings; but can the same work for empirical models? In this work, we try to inform the pre-trained masked language models of word meanings for semantics-enhanced pre-training. To achieve a contrastive and holistic view of word meanings, a definition pair of two related words is presented to the masked language model such that the model can better associate a word with its crucial semantic features. Both intrinsic and extrinsic evaluations validate the proposed approach on semantics-orientated tasks, with an almost negligible increase of training data.

Keeping Models Consistent between Pretraining and Translation for Low-Resource Neural Machine Translation

Recently, the pretraining of models has been successfully applied to unsupervised and semi-supervised neural machine translation. A cross-lingual language model uses a pretrained masked language model to initialize the encoder and decoder of the translation model, which greatly improves the translation quality. However, because of a mismatch in the number of layers, the pretrained model can only initialize part of the decoder’s parameters. In this paper, we use a layer-wise coordination transformer and a consistent pretraining translation transformer instead of a vanilla transformer as the translation model. The former has only an encoder, and the latter has an encoder and a decoder, but the encoder and decoder have exactly the same parameters. Both models can guarantee that all parameters in the translation model can be initialized by the pretrained model. Experiments on the Chinese–English and English–German datasets show that compared with the vanilla transformer baseline, our models achieve better performance with fewer parameters when the parallel corpus is small.

Encoder-Decoder Models Can Benefit from Pre-trained Masked Language Models in Grammatical Error Correction

This paper investigates how to effectively incorporate a pre-trained masked language model (MLM), such as BERT, into an encoder-decoder (EncDec) model for grammatical error correction (GEC). The answer to this question is not as straightforward as one might expect because the previous common methods for incorporating a MLM into an EncDec model have potential drawbacks when applied to GEC. For example, the distribution of the inputs to a GEC model can be considerably different (erroneous, clumsy, etc.) from that of the corpora used for pre-training MLMs; however, this issue is not addressed in the previous methods. Our experiments show that our proposed method, where we first fine-tune a MLM with a given GEC corpus and then use the output of the fine-tuned MLM as additional features in the GEC model, maximizes the benefit of the MLM. The best-performing model achieves state-of-the-art performances on the BEA-2019 and CoNLL-2014 benchmarks. Our code is publicly available at: https://github.com/kanekomasahiro/bert-gec.