The lack of resources for languages in the Americas has proven to be a problem for the creation of digital systems such as machine translation, search engines, chat bots, and more. The scarceness of digital resources for a language causes a higher impact on populations where the language is spoken by millions of people. We introduce the first official large combined corpus for deep learning of an indigenous South American low-resource language spoken by millions called Quechua. Specifically, our curated corpus is created from text gathered from the southern region of Peru where a dialect of Quechua is spoken that has not traditionally been used for digital systems as a target dialect in the past. In order to make our work repeatable by others, we also offer a public, pre-trained, BERT model called QuBERT which is the largest linguistic model ever trained for any Quechua type, not just the southern region dialect. We furthermore test our corpus and its corresponding BERT model on two major tasks: (1) named-entity recognition (NER) and (2) part-of-speech (POS) tagging by using state-of-the-art techniques where we achieve results comparable to other work on higher-resource languages. In this article, we describe the methodology, challenges, and results from the creation of QuBERT which is on par with other state-of-the-art multilingual models for natural language processing achieving between 71 and 74% F1 score on NER and 84–87% on POS tasks.

The distant supervision (DS) paradigm has been widely used for relation extraction (RE) to alleviate the need for expensive annotations. However, it suffers from noisy labels, which leads to worse performance than models trained on human-annotated data, even when trained using hundreds of times more data. We present a systematic study on the use of natural language inference (NLI) to improve distantly supervised document-level RE. We apply NLI in three scenarios: (i) as a filter for denoising DS labels, (ii) as a filter for model prediction, and (iii) as a standalone RE model. Our results show that NLI filtering consistently improves performance, reducing the performance gap with a model trained on human-annotated data by 2.3 F1.

Large pre-trained models are usually fine-tuned on downstream task data, and tested on unseen data. When the train and test data come from different domains, the model is likely to struggle, as it is not adapted to the test domain. We propose a new approach for domain adaptation (DA), using neuron-level interventions: We modify the representation of each test example in specific neurons, resulting in a counterfactual example from the source domain, which the model is more familiar with. The modified example is then fed back into the model. While most other DA methods are applied during training time, ours is applied during inference only, making it more efficient and applicable. Our experiments show that our method improves performance on unseen domains.

Training a tagger for Named Entity Recognition (NER) requires a substantial amount of labeled data in the task domain. Manual labeling is a tedious and complicated task. Semisupervised learning methods can reduce the quantity of labeled data necessary to train a model. However, these methods require large quantities of unlabeled data, which remains an issue in many cases.We address this problem by generating unlabeled data. Large language models have proven to be powerful tools for text generation. We use their generative capacity to produce new sentences and variations of the sentences of our available data. This generation method, combined with a semi-supervised method, is evaluated on CoNLL and I2B2. We prepare both of these corpora to simulate a low resource setting. We obtain significant improvements for semisupervised learning with synthetic data against supervised learning on natural data.

Text segmentation and extraction from unstructured documents can provide business researchers with a wealth of new information on firms and their behaviors. However, the most valuable text is often difficult to extract consistently due to substantial variations in how content can appear from document to document. Thus, the most successful way to extract this content has been through costly crowdsourcing and training of manual workers. We propose the Assisted Neural Text Segmentation (ANTS) framework to identify pertinent text in unstructured documents from a small set of labeled examples. ANTS leverages deep learning and transfer learning architectures to empower researchers to identify relevant text with minimal manual coding. Using a real world sample of accounting documents, we identify targeted sections 96% of the time using only 5 training examples.

Deep learning methods have enabled taskoriented semantic parsing of increasingly complex utterances. However, a single model is still typically trained and deployed for each task separately, requiring labeled training data for each, which makes it challenging to support new tasks, even within a single business vertical (e.g., food-ordering or travel booking). In this paper we describe Cross-TOP (Cross-Schema Task-Oriented Parsing), a zero-shot method for complex semantic parsing in a given vertical. By leveraging the fact that user requests from the same vertical share lexical and semantic similarities, a single cross-schema parser is trained to service an arbitrary number of tasks, seen or unseen, within a vertical. We show that Cross-TOP can achieve high accuracy on a previously unseen task without requiring any additional training data, thereby providing a scalable way to bootstrap semantic parsers for new tasks. As part of this work we release the FoodOrdering dataset, a task-oriented parsing dataset in the food-ordering vertical, with utterances and annotations derived from five schemas, each from a different restaurant menu.

While standard Estonian is not a low-resourced language, the different dialects of the language are under-resourced from the point of view of NLP, given that there are no vast hand normalized resources available for training a machine learning model to normalize dialectal Estonian to standard Estonian. In this paper, we crawl a small corpus of parallel dialectal Estonian - standard Estonian sentences. In addition, we take a savvy approach of generating more synthetic training data for the normalization task by using an existing dialect generator model built for Finnish to “dialectalize” standard Estonian sentences from the Universal Dependencies tree banks. Our BERT based normalization model achieves a word error rate that is 26.49 points lower when using both the synthetic data and Estonian data in comparison to training the model with only the available Estonian data. Our results suggest that synthetic data generated by a model trained on a more resourced related language can indeed boost the results for a less resourced language.

Back translation is one of the most widely used methods for improving the performance of neural machine translation systems. Recent research has sought to enhance the effectiveness of this method by increasing the ‘diversity’ of the generated translations. We argue that the definitions and metrics used to quantify ‘diversity’ in previous work have been insufficient. This work puts forward a more nuanced framework for understanding diversity in training data, splitting it into lexical diversity and syntactic diversity. We present novel metrics for measuring these different aspects of diversity and carry out empirical analysis into the effect of these types of diversity on final neural machine translation model performance for low-resource English↔Turkish and mid-resource English↔Icelandic. Our findings show that generating back translation using nucleus sampling results in higher final model performance, and that this method of generation has high levels of both lexical and syntactic diversity. We also find evidence that lexical diversity is more important than syntactic for back translation performance.

Automatic Speech Recognition (ASR) systems typically produce unpunctuated transcripts that have poor readability. In addition, building a punctuation restoration system is challenging for low-resource languages, especially for domain-specific applications. In this paper, we propose a Spanish punctuation restoration system designed for a real-time customer support transcription service. To address the data sparsity of Spanish transcripts in the customer support domain, we introduce two transferlearning-based strategies: 1) domain adaptation using out-of-domain Spanish text data; 2) crosslingual transfer learning leveraging in-domain English transcript data. Our experiment results show that these strategies improve the accuracy of the Spanish punctuation restoration system.

Multilingual language models such as mBERT have seen impressive cross-lingual transfer to a variety of languages, but many languages remain excluded from these models. In this paper, we analyse the effect of pre-training with monolingual data for a low-resource language that is not included in mBERT – Maltese – with a range of pre-training set ups. We conduct evaluations with the newly pre-trained models on three morphosyntactic tasks – dependency parsing, part-of-speech tagging, and named-entity recognition – and one semantic classification task – sentiment analysis. We also present a newly created corpus for Maltese, and determine the effect that the pre-training data size and domain have on the downstream performance. Our results show that using a mixture of pre-training domains is often superior to using Wikipedia text only. We also find that a fraction of this corpus is enough to make significant leaps in performance over Wikipedia-trained models. We pre-train and compare two models on the new corpus: a monolingual BERT model trained from scratch (BERTu), and a further pretrained multilingual BERT (mBERTu). The models achieve state-of-the-art performance on these tasks, despite the new corpus being considerably smaller than typically used corpora for high-resourced languages. On average, BERTu outperforms or performs competitively with mBERTu, and the largest gains are observed for higher-level tasks.

Supervised event extraction models require a substantial amount of training data to perform well. However, event annotation requires a lot of human effort and costs much time, which limits the application of existing supervised approaches to new event types. In order to reduce manual labor and shorten the time to build an event extraction system for an arbitrary event ontology, we present a new framework to train such systems much more efficiently without large annotations. Our event trigger labeling model uses a weak supervision approach, which only requires a set of keywords, a small number of examples and an unlabeled corpus, on which our approach automatically collects weakly supervised annotations. Our argument role labeling component performs zero-shot learning, which only requires the names of the argument roles of new event types. The source codes of our event trigger detection1 and event argument extraction2 models are publicly available for research purposes. We also release a dockerized system connecting the two models into an unified event extraction pipeline.

Pretrained language models have shown success in various areas of natural language processing, including reading comprehension tasks. However, when applying machine learning methods to new domains, labeled data may not always be available. To address this, we use supervised pretraining on source-domain data to reduce sample complexity on domainspecific downstream tasks. We evaluate zeroshot performance on domain-specific reading comprehension tasks by combining task transfer with domain adaptation to fine-tune a pretrained model with no labelled data from the target task. Our approach outperforms DomainAdaptive Pretraining on downstream domainspecific reading comprehension tasks in 3 out of 4 domains.

Curriculum learning strategies in prior multitask learning approaches arrange datasets in a difficulty hierarchy either based on human perception or by exhaustively searching the optimal arrangement. However, human perception of difficulty may not always correlate well with machine interpretation leading to poor performance and exhaustive search is computationally expensive. Addressing these concerns, we propose two classes of techniques to arrange training instances into a learning curriculum based on difficulty scores computed via model-based approaches. The two classes i.e Dataset-level and Instance-level differ in granularity of arrangement. Through comprehensive experiments with 12 datasets, we show that instance-level and dataset-level techniques result in strong representations as they lead to an average performance improvement of 4.17% and 3.15% over their respective baselines. Furthermore, we find that most of this improvement comes from correctly answering the difficult instances, implying a greater efficacy of our techniques on difficult tasks

Pretrained language models represent the state of the art in NLP, but the successful construction of such models often requires large amounts of data and computational resources. Thus, the paucity of data for low-resource languages impedes the development of robust NLP capabilities for these languages. There has been some recent success in pretraining encoderonly models solely on a combination of lowresource African languages, exemplified by AfriBERTa. In this work, we extend the approach of “small data” pretraining to encoder– decoder models. We introduce AfriTeVa, a family of sequence-to-sequence models derived from T5 that are pretrained on 10 African languages from scratch. With a pretraining corpus of only around 1GB, we show that it is possible to achieve competitive downstream effectiveness for machine translation and text classification, compared to larger models trained on much more data. All the code and model checkpoints described in this work are publicly available at https://github.com/castorini/afriteva.

This paper presents our study in exploring the task of named entity recognition (NER) in a low resource setting, focusing on few-shot learning on the Sumerian NER task. The Sumerian language is deemed as an extremely low-resource language due to that (1) it is a long dead language, (2) highly skilled language experts are extremely scarce. NER on Sumerian text is important in that it helps identify the actors and entities active in a given period of time from the collections of tens of thousands of texts in building socio-economic networks of the archives of interest. As a text classification task, NER tends to become challenging when the amount of annotated data is limited or the model is required to handle new classes. The Sumerian NER is no exception. In this work, we propose to use two few-shot learning systems, ProtoBERT and NNShot, to the Sumerian NER task. Our experiments show that the ProtoBERT NER generally outperforms both the NNShot NER and the fully supervised BERT NER in low resource settings on the predictions of rare classes. In particular, F1-score of ProtoBERT on unseen entity types on our test set has achieved 89.6% that is significantly better than the F1-score of 84.3% of the BERT NER.

Recent advances in the field of deep learning have led to a growing interest in the development of NLP approaches for low-resource and endangered languages. Nevertheless, relatively little research, related to NLP, has been conducted on indigenous languages. These languages are considered to be filled with complexities and challenges that make their study incredibly difficult in the NLP and AI fields. This paper focuses on the morphological segmentation of indigenous languages, an extremely challenging task because of polysynthesis, dialectal variations with rich morpho-phonemics, misspellings and resource-limited scenario issues. The proposed approach, towards a morphological segmentation of Innu-Aimun, an extremely low-resource indigenous language of Canada, is based on deep learning. Experiments and evaluations have shown promising results, compared to state-of-the-art rule-based and unsupervised approaches.

Crowdsourcing platforms are often used to collect datasets for training machine learning models, despite higher levels of inaccurate labeling compared to expert labeling. There are two common strategies to manage the impact of such noise: The first involves aggregating redundant annotations, but comes at the expense of labeling substantially fewer examples. Secondly, prior works have also considered using the entire annotation budget to label as many examples as possible and subsequently apply denoising algorithms to implicitly clean the dataset. We find a middle ground and propose an approach which reserves a fraction of annotations to explicitly clean up highly probable error samples to optimize the annotation process. In particular, we allocate a large portion of the labeling budget to form an initial dataset used to train a model. This model is then used to identify specific examples that appear most likely to be incorrect, which we spend the remaining budget to relabel. Experiments across three model variations and four natural language processing tasks show our approach outperforms or matches both label aggregation and advanced denoising methods designed to handle noisy labels when allocated the same finite annotation budget.

Knowledge Graphs (KGs) are directed labeled graphs representing entities and the relationships between them. Most prior work focuses on supervised or semi-supervised approaches which require large amounts of annotated data. While unsupervised approaches do not need labeled training data, most existing methods either generate too many redundant relations or require manual mapping of the extracted relations to a known schema. To address these limitations, we propose an unsupervised method for KG generation that requires neither labeled data nor manual mapping to the predefined relation schema. Instead, our method leverages sentence-level semantic similarity for automatically generating relations between pairs of entities. Our proposed method outperforms two baseline systems when evaluated over four datasets.

Our collective attention span is shortened by the flood of online information. With FarFetched, we address the need for automated claim validation based on the aggregated evidence derived from multiple online news sources. We introduce an entity-centric reasoning framework in which latent connections between events, actions, or statements are revealed via entity mentions and represented in a graph database. Using entity linking and semantic similarity, we offer a way for collecting and combining information from diverse sources in order to generate evidence relevant to the user’s claim. Then, we leverage textual entailment recognition to quantitatively determine whether this assertion is credible, based on the created evidence. Our approach tries to fill the gap in automated claim validation for less-resourced languages and is showcased on the Greek language, complemented by the training of relevant semantic textual similarity (STS) and natural language inference (NLI) models that are evaluated on translated versions of common benchmarks.

Natural Language Processing (NLP) tasks in non-dominant and low-resource languages have not experienced significant progress. Although pre-trained BERT models are available, GPU-dependency, large memory requirement, and data scarcity often limit their applicability. As a solution, this paper proposes a fusion chain architecture comprised of one or more layers of CNN, LSTM, and BiLSTM and identifies precise configuration and chain length. The study shows that a simpler, CPU-trainable non-BERT fusion CNN + BiLSTM + CNN is sufficient to surpass the textual classification performance of the BERT-related models in resource-limited languages and environments. The fusion architecture competitively approaches the state-of-the-art accuracy in several Bengali NLP tasks and a six-class emotion detection task for a newly developed Bengali dataset. Interestingly, the performance of the identified fusion model, for instance, CNN + BiLSTM + CNN, also holds for other lowresource languages and environments. Efficacy study shows that the CNN + BiLSTM + CNN model outperforms BERT implementation for Vietnamese languages and performs almost equally in English NLP tasks experiencing artificial data scarcity. For the GLUE benchmark and other datasets such as Emotion, IMDB, and Intent classification, the CNN + BiLSTM + CNN model often surpasses or competes with BERT-base, TinyBERT, DistilBERT, and mBERT. Besides, a position-sensitive selfattention layer role further improves the fusion models’ performance in the Bengali emotion classification. The models are also compressible to as low as ≈ 5× smaller through pruning and retraining, making them more viable for resource-constrained environments. Together, this study may help NLP practitioners and serve as a blueprint for NLP model choices in textual classification for low-resource languages and environments.

Aspect Term Extraction (ATE) is the task of identifying the word(s) in a review text toward which the author express an opinion. A major challenges for ATE involve data scarcity that hinder the training of deep sequence taggers to identify rare targets. To overcome these issues, we propose a novel method to better exploit the available labeled data for ATE by computing effective complement sentences to augment the input data and facilitate the aspect term prediction. In particular, we introduce a multistep training procedure that first obtains optimal complement representations and sentences for training data with respect to a deep ATE model. Afterward, we fine-tune the generative language model GPT-2 to allow complement sentence generation at test data. The REINFORCE algorithm is employed to incorporate different expected properties into the reward function to perform the fine-tuning. We perform extensive experiments on the benchmark datasets to demonstrate the benefits of the proposed method that achieve the state-of-the-art performance on different datasets.

The lack of resources for languages in the Americas has proven to be a problem for the creation of digital systems such as machine translation, search engines, chat bots, and more. The scarceness of digital resources for a language causes a higher impact on populations where the language is spoken by millions of people. We introduce the first official large combined corpus for deep learning of an indigenous South American low-resource language spoken by millions called Quechua. Specifically, our curated corpus is created from text gathered from the southern region of Peru where a dialect of Quechua is spoken that has not traditionally been used for digital systems as a target dialect in the past. In order to make our work repeatable by others, we also offer a public, pre-trained, BERT model called QuBERT which is the largest linguistic model ever trained for any Quechua type, not just the southern region dialect. We furthermore test our corpus and its corresponding BERT model on two major tasks: (1) named-entity recognition (NER) and (2) part-of-speech (POS) tagging by using state-of-the-art techniques where we achieve results comparable to other work on higher-resource languages. In this article, we describe the methodology, challenges, and results from the creation of QuBERT which is on on par with other state-of-the-art multilingual models for natural language processing achieving between 71 and 74% F1 score on NER and 84–87% on POS tasks.

Highly accurate machine translation systems are very important in societies and countries where multilinguality is very common, and where English often does not suffice. The Indian subcontinent (or South Asia) is such a region, with all the Indic languages currently being under-represented in the NLP ecosystem. It is essential to thoroughly explore various techniques to improve the performance of such lowresource languages at least using the data available in open-source, which itself is something not very explored in the Indic ecosystem. In our work, we perform a study with a focus on improving the performance of very-low-resource South Asian languages, especially of countries in addition to India. Specifically, we propose how unified models can be built that can exploit the data from comparatively resource-rich languages of the same region. We propose strategies to unify different types of unexplored scripts, especially Perso–Arabic scripts and Indic scripts to build multilingual models for all the South Asian languages despite the script barrier. We also study how augmentation techniques like back-translation can be made useof to build unified models just using openly available raw data, to understand what levels of improvements can be expected for these Indic languages.