Specialised pre-trained language models are becoming more frequent in Natural language Processing (NLP) since they can potentially outperform models trained on generic texts. BioBERT (Sanh et al., Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv: 1910.01108, 2019) and BioClinicalBERT (Alsentzer et al., Publicly available clinical bert embeddings. In Proceedings of the 2nd Clinical Natural Language Processing Workshop, pp. 72–78, 2019) are two examples of such models that have shown promise in medical NLP tasks. Many of these models are overparametrised and resource-intensive, but thanks to techniques like knowledge distillation, it is possible to create smaller versions that perform almost as well as their larger counterparts. In this work, we specifically focus on development of compact language models for processing clinical texts (i.e. progress notes, discharge summaries, etc). We developed a number of efficient lightweight clinical transformers using knowledge distillation and continual learning, with the number of parameters ranging from $15$ million to $65$ million. These models performed comparably to larger models such as BioBERT and ClinicalBioBERT and significantly outperformed other compact models trained on general or biomedical data. Our extensive evaluation was done across several standard datasets and covered a wide range of clinical text-mining tasks, including natural language inference, relation extraction, named entity recognition and sequence classification. To our knowledge, this is the first comprehensive study specifically focused on creating efficient and compact transformers for clinical NLP tasks. The models and code used in this study can be found on our Huggingface profile at https://huggingface.co/nlpie and Github page at https://github.com/nlpie-research/Lightweight-Clinical-Transformers, respectively, promoting reproducibility of our results.

Table-to-text generation aims to generate descriptions for structured data (i.e., tables) and has been applied in many fields like question-answering systems and search engines. Current approaches mostly use neural language models to learn alignment between output and input based on the attention mechanisms, which are still flawed by the gradual weakening of attention when processing long texts and the inability to utilize the records’ structural information. To solve these problems, we propose a novel generative model SAN-T2T, which consists of a field-content selective encoder and a descriptive decoder, connected with a selective attention network. In the encoding phase, the table’s structure is integrated into its field representation, and a content selector with self-aligned gates is applied to take advantage of the fact that different records can determine each other’s importance. In the decoding phase, the content selector’s semantic information enhances the alignment between description and records, and a featured copy mechanism is applied to solve the rare word problem. Experiments on WikiBio and WeatherGov datasets show that SAN-T2T outperforms the baselines by a large margin, and the content selector indeed improves the model’s performance.

Deep neural networks as an end-to-end approach lack robustness from an application point of view, as it is very difficult to fix an obvious problem without retraining the model, for example, when a model consistently predicts positive when seeing the word “terrible.” Meanwhile, it is less stressed that the commonly used attention mechanism is likely to “over-fit” by being overly sparse, so that some key positions in the input sequence could be overlooked by the network. To address these problems, we proposed a lexicon-enhanced attention LSTM model in 2019, named ATLX. In this paper, we describe extended experiments and analysis of the ATLX model. And, we also try to further improve the aspect-based sentiment analysis system by combining a vector-based sentiment domain adaptation method.

House price prediction is an important problem that could benefit home buyers and sellers. Traditional models for house price prediction use numerical attributes such as the number of rooms but disregard the house description text. The recent developments in text processing suggest these can be valuable attributes, which motivated us to use house descriptions. This paper focuses on the house asking/advertising price and studies the impact of using house description texts to predict the final house price. To achieve this, we collected a large and diverse set of attributes on house postings, including the house advertising price. Then, we compare the performance of three scenarios: using only the house description, only numeric attributes, or both. We processed the description text through three word embedding techniques: TF-IDF, Word2Vec, and BERT. Four regression algorithms are trained using only textual data, non-textual data, or both. Our results show that by using exclusively the description data with Word2Vec and a Deep Learning model, we can achieve good performance. However, the best overall performance is obtained when using both textual and non-textual features. An $R^2$ of 0.7904 is achieved by the deep learning model using only description data on the testing data. This clearly indicates that using the house description text alone is a strong predictor for the house price. However, when observing the RMSE on the test data, the best model was gradient boosting using both numeric and description data. Overall, we observe that combining the textual and non-textual features improves the learned model and provides performance benefits when compared against using only one of the feature types. We also provide a freely available application for house price prediction, which is solely based on a house text description and uses our final developed model with Word2Vec and Deep Learning to predict the house price.

Randomized prospective studies represent the gold standard for experimental design. In this paper, we present a randomized prospective study to validate the benefits of combining rule-based and data-driven natural language understanding methods in a virtual patient dialogue system. The system uses a rule-based pattern matching approach together with a machine learning (ML) approach in the form of a text-based convolutional neural network, combining the two methods with a simple logistic regression model to choose between their predictions for each dialogue turn. In an earlier, retrospective study, the hybrid system yielded a nearly 50% error reduction on our initial data, in part due to the differential performance between the two methods as a function of label frequency. Given these gains, and considering that our hybrid approach is unique among virtual patient systems, we compare the hybrid system to the rule-based system by itself in a randomized prospective study. We evaluate 110 unique medical student subjects interacting with the system over 5,296 conversation turns, to verify whether similar gains are observed in a deployed system. This prospective study broadly confirms the findings from the earlier one but also highlights important deficits in our training data. The hybrid approach still improves over either rule-based or ML approaches individually, even handling unseen classes with some success. However, we observe that live subjects ask more out-of-scope questions than expected. To better handle such questions, we investigate several modifications to the system combination component. These show significant overall accuracy improvements and modest F1 improvements on out-of-scope queries in an offline evaluation. We provide further analysis to characterize the difficulty of the out-of-scope problem that we have identified, as well as to suggest future improvements over the baseline we establish here.

In this paper, we propose an enhanced version of the vanilla transformer for data-to-text generation and then use it as the generator of a conditional generative adversarial model to improve the semantic quality and diversity of output sentences. Specifically, by adding a diagonal mask matrix to the attention scores of the encoder and using the history of the attention weights in the decoder, this enhanced version of the vanilla transformer prevents semantic defects in the output text. Also, using this enhanced transformer along with a triplet network, respectively, as the generator and discriminator of conditional generative adversarial network, diversity and semantic quality of sentences are guaranteed. To prove the effectiveness of the proposed model, called conditional generative adversarial with enhanced transformer (CGA-ET), we performed experiments on three different datasets and observed that our proposed model is able to achieve better results than the baselines models in terms of BLEU, METEOR, NIST, ROUGE-L, CIDEr, BERTScore, and SER automatic evaluation metrics as well as human evaluation.

We propose a multilingual data-driven method for generating reading comprehension questions using dependency trees. Our method provides a strong, deterministic and inexpensive-to-train baseline for less-resourced languages. While a language-specific corpus is still required, its size is nowhere near those required by modern neural question generation (QG) architectures. Our method surpasses QG baselines previously reported in the literature in terms of automatic evaluation metrics and shows a good performance in terms of human evaluation.

Recent work on text simplification has focused on the use of control tokens to further the state-of-the-art. However, it is not easy to further improve without an in-depth comprehension of the mechanisms underlying control tokens. One unexplored factor is the tokenization strategy, which we also explore. In this paper, we (1) reimplemented AudienCe-CEntric Sentence Simplification, (2) explored the effects and interactions of varying control tokens, (3) tested the influences of different tokenization strategies, (4) demonstrated how separate control tokens affect performance and (5) proposed new methods to predict the value of control tokens. We show variations of performance in the four control tokens separately. We also uncover how the design of control tokens could influence performance and give some suggestions for designing control tokens. We show the newly proposed method with higher performance in both SARI (a common scoring metric in text simplificaiton) and BERTScore (a score derived from the BERT language model) and potential in real applications.

This paper proposes a sequence-to-sequence model for data-to-text generation, called DM-NLG, to generate a natural language text from structured nonlinguistic input. Specifically, by adding a dynamic memory module to the attention-based sequence-to-sequence model, it can store the information that leads to generate previous output words and use it to generate the next word. In this way, the decoder part of the model is aware of all previous decisions, and as a result, the generation of duplicate words or incomplete semantic concepts is prevented. To improve the generated sentences quality by the DM-NLG decoder, a postprocessing step is performed using the pretrained language models. To prove the effectiveness of the DM-NLG model, we performed experiments on five different datasets and observed that our proposed model is able to reduce the slot error rate rate by 50% and improve the BLEU by 10%, compared to the state-of-the-art models.

We examine the binary classification of sentiment views for verbal multiword expressions (MWEs). Sentiment views denote the perspective of the holder of some opinion. We distinguish between MWEs conveying the view of the speaker of the utterance (e.g., in “The company reinvented the wheel” the holder is the implicit speaker who criticizes the company for creating something already existing) and MWEs conveying the view of explicit entities participating in an opinion event (e.g., in “Peter threw in the towel” the holder is Peter having given up something). The task has so far been examined on unigram opinion words. Since many features found effective for unigrams are not usable for MWEs, we propose novel ones taking into account the internal structure of MWEs, a unigram sentiment-view lexicon and various information from Wiktionary. We also examine distributional methods and show that the corpus on which a representation is induced has a notable impact on the classification. We perform an extrinsic evaluation in the task of opinion holder extraction and show that the learnt knowledge also improves a state-of-the-art classifier trained on BERT. Sentiment-view classification is typically framed as a task in which only little labeled training data are available. As in the case of unigrams, we show that for MWEs a feature-based approach beats state-of-the-art generic methods.

GPT-3 is a large-scale natural language model developed by OpenAI that can perform many different tasks, including topic classification. Although researchers claim that it requires only a small number of in-context examples to learn a task, in practice GPT-3 requires these training examples to be either of exceptional quality or a higher quantity than easily created by hand. To address this issue, this study teaches GPT-3 to classify whether a question is related to data science by augmenting a small training set with additional examples generated by GPT-3 itself. This study compares two augmented classifiers: the Classification Endpoint with an increased training set size and the Completion Endpoint with an augmented prompt optimized using a genetic algorithm. We find that data augmentation significantly increases the accuracy of both classifiers, and that the embedding-based Classification Endpoint achieves the best accuracy of about 76%, compared to human accuracy of 85%. In this way, giving large language models like GPT-3 the ability to propose their own training examples can improve short text classification performance.

Current research in yes/no question answering (QA) focuses on transfer learning techniques and transformer-based models. Models trained on large corpora are fine-tuned on tasks similar to yes/no QA, and then the captured knowledge is transferred for solving the yes/no QA task. Most previous studies use existing similar tasks, such as natural language inference or extractive QA, for the fine-tuning step. This paper follows a different perspective, hypothesizing that an artificial yes/no task can transfer useful knowledge for improving the performance of yes/no QA. We introduce three such tasks for this purpose, by adapting three corresponding existing tasks: candidate answer validation, sentiment classification, and lexical simplification. Furthermore, we experimented with three different variations of the BERT model (BERT base, RoBERTa, and ALBERT). The results show that our hypothesis holds true for all artificial tasks, despite the small size of the corresponding datasets that are used for the fine-tuning process, the differences between these tasks, the decisions that we made to adapt the original ones, and the tasks’ simplicity. This gives an alternative perspective on how to deal with the yes/no QA problem, that is more creative, and at the same time more flexible, as it can exploit multiple other existing tasks and corresponding datasets to improve yes/no QA models.

Latent semantic analysis (LSA) and correspondence analysis (CA) are two techniques that use a singular value decomposition for dimensionality reduction. LSA has been extensively used to obtain low-dimensional representations that capture relationships among documents and terms. In this article, we present a theoretical analysis and comparison of the two techniques in the context of document-term matrices. We show that CA has some attractive properties as compared to LSA, for instance that effects of margins, that is, sums of row elements and column elements, arising from differing document lengths and term frequencies are effectively eliminated so that the CA solution is optimally suited to focus on relationships among documents and terms. A unifying framework is proposed that includes both CA and LSA as special cases. We empirically compare CA to various LSA-based methods on text categorization in English and authorship attribution on historical Dutch texts and find that CA performs significantly better. We also apply CA to a long-standing question regarding the authorship of the Dutch national anthem Wilhelmus and provide further support that it can be attributed to the author Datheen, among several contenders.

This article discusses the development of an automated plot extraction system for narrative texts. Acknowledging the distinction between plot, as an object of study with its own rich history and literature, and features of a text that may be automatically extractable, we begin by characterizing a text’s scatter plot of entities. This visualization of a text reveals entity density patterns characterizing the particular telling of the story under investigation and leads to effective scene partitioning. We then introduce the concept of narrative flow, a graph representation of the narrative ordering of scenes (the syuzhet) that includes how entities move through scenes from the text, and investigate the degree to which narrative flow can be automatically extracted given a glossary of plot-important objects, actors, and locations. Our subsequent analysis then explores the correlation between subjective notions of plot and the information extracted through these visualizations. In particular, we discuss narrative structures commonly found within the graphs and make comparisons with ground truth narrative flow graphs, showing mixed results highlighting the difficulty of plot extraction. However, the visual artifacts and common structural relationships seen in the graphs provide insight into narrative and its underlying plot.

Recently, neural abstractive text summarization (NATS) models based on sequence-to-sequence architecture have drawn a lot of attention. Real-world texts that need to be summarized range from short news with dozens of words to long reports with thousands of words. However, most existing NATS models are not good at summarizing long documents, due to the inherent limitations of their underlying neural architectures. In this paper, we focus on the task of long document summarization (LDS). Based on the inherent section structures of source documents, we divide an abstractive LDS problem into several smaller-sized problems. In this circumstance, how to provide a less-biased target summary as the supervision for each section is vital for the model’s performance. As a preliminary, we formally describe the section-to-summary-sentence (S2SS) alignment for LDS. Based on this, we propose a novel NATS framework for the LDS task. Our framework is built based on the theory of unbalanced optimal transport (UOT), and it is named as UOTSumm. It jointly learns three targets in a unified training objective, including the optimal S2SS alignment, a section-level NATS summarizer, and the number of aligned summary sentences for each section. In this way, UOTSumm directly learns the text alignment from summarization data, without resorting to any biased tool such as ROUGE. UOTSumm can be easily adapted to most existing NATS models. And we implement two versions of UOTSumm, with and without the pretrain-finetune technique. We evaluate UOTSumm on three publicly available LDS benchmarks: PubMed, arXiv, and GovReport. UOTSumm obviously outperforms its counterparts that use ROUGE for the text alignment. When combined with UOTSumm, the performance of two vanilla NATS models improves by a large margin. Besides, UOTSumm achieves better or comparable performance when compared with some recent strong baselines.

Text style transfer (TST) aims at automatically changing a text’s stylistic features, such as formality, sentiment, authorial style, humor, and complexity, while still trying to preserve its content. Although the scientific community has investigated TST since the 1980s, it has recently regained attention by adopting deep unsupervised strategies to address the challenge of training without parallel data. In this manuscript, we investigate how relying on sequence-to-sequence pretraining models affects the performance of TST when the pretraining step leverages pairs of paraphrase data. Furthermore, we propose a new technique to enhance the sequence-to-sequence model by distilling knowledge from masked language models. We evaluate our proposals on three unsupervised style transfer tasks with widely used benchmarks: author imitation, formality transfer, and polarity swap. The evaluation relies on quantitative and qualitative analyses and comparisons with the results of state-of-the-art models. For the author imitation and the formality transfer task, we show that using the proposed techniques improves all measured metrics and leads to state-of-the-art (SOTA) results in content preservation and an overall score in the author imitation domain. In the formality transfer domain, we paired with the SOTA method in the style control metric. Regarding the polarity swap domain, we show that the knowledge distillation component improves all measured metrics. The paraphrase pretraining increases content preservation at the expense of harming style control. Based on the results reached in these domains, we also discuss in the manuscript if the tasks we address have the same nature and should be equally treated as TST tasks.

Word order is one of the most important grammatical devices and the basis for language understanding. However, as one of the most popular NLP architectures, Transformer does not explicitly encode word order. A solution to this problem is to incorporate position information by means of position encoding/embedding (PE). Although a variety of methods of incorporating position information have been proposed, the NLP community is still in want of detailed statistical researches on position information in real-life language. In order to understand the influence of position information on the correlation between words in more detail, we investigated the factors that affect the frequency of words and word sequences in large corpora. Our results show that absolute position, relative position, being at one of the two ends of a sentence and sentence length all significantly affect the frequency of words and word sequences. Besides, we observed that the frequency distribution of word sequences over relative position carries valuable grammatical information. Our study suggests that in order to accurately capture word–word correlations, it is not enough to focus merely on absolute and relative position. Transformers should have access to more types of position-related information which may require improvements to the current architecture.

Solving the Arithmetic Word Problems (AWPs) using AI techniques has attracted much attention in recent years. We feel that the current AWP solvers are under-utilizing the relevant domain knowledge. We present a knowledge- and learning-based system that effectively solves AWPs of a specific type—those that involve transfer of objects from one agent to another (Transfer Cases (TC)). We represent the knowledge relevant to these problems as TC Ontology. The sentences in TC-AWPs contain information of essentially four types: before-transfer, transfer, after-transfer, and query. Our system (KLAUS-Tr) uses statistical classifier to recognize the types of sentences. The sentence types guide the information extraction process used to identify the agents, quantities, units, types of objects, and the direction of transfer from the AWP text. The extracted information is represented as an RDF graph that utilizes the TC Ontology terminology. To solve the given AWP, we utilize semantic web rule language (SWRL) rules that capture the knowledge about how object transfer affects the RDF graph of the AWP. Using the TC ontology, we also analyze if the given problem is consistent or otherwise. The different ways in which TC-AWPs can be inconsistent are encoded as SWRL rules. Thus, KLAUS-Tr can identify if the given AWP is invalid and accordingly notify the user. Since the existing datasets do not have inconsistent AWPs, we create AWPs of this type and augment the datasets. We have implemented KLAUS-Tr and tested it on TC-type AWPs drawn from the All-Arith and other datasets. We find that TC-AWPs constitute about 40% of the AWPs in a typical dataset like All-Arith. Our system achieves an impressive accuracy of 92%, thus improving the state-of-the-art significantly. We plan to extend the system to handle AWPs that contain multiple transfers of objects and also offer explanations of the solutions.

We introduce an extensive dataset for multilingual probing of morphological information in language models (247 tasks across 42 languages from 10 families), each consisting of a sentence with a target word and a morphological tag as the desired label, derived from the Universal Dependencies treebanks. We find that pre-trained Transformer models (mBERT and XLM-RoBERTa) learn features that attain strong performance across these tasks. We then apply two methods to locate, for each probing task, where the disambiguating information resides in the input. The first is a new perturbation method that “masks” various parts of context; the second is the classical method of Shapley values. The most intriguing finding that emerges is a strong tendency for the preceding context to hold more information relevant to the prediction than the following context.