Compared to standard language model (LM) pretraining (i.e., from scratch), Knowledge Distillation (KD) entails an additional forward pass through a teacher model that is typically substantially larger than the target student model. As such, KD in LM pretraining materially slows down throughput of pretraining instances vis-a-vis pretraining from scratch. Scaling laws of LM pretraining suggest that smaller models can close the gap to larger counterparts if trained on more data (i.e., processing more tokens)—and under a fixed computation budget, smaller models are able to process more data than larger models. We thus hypothesize that KD might, in fact, be suboptimal to pretraining from scratch for obtaining smaller LMs, when appropriately accounting for the compute budget. To test this, we compare pretraining from scratch against several KD strategies for masked language modeling (MLM) in a fair experimental setup, with respect to amount of computation as well as pretraining data. Downstream results on GLUE, however, do not confirm our hypothesis: while pretraining from scratch performs comparably to ordinary KD under a fixed computation budget, more sophisticated KD strategies, namely TinyBERT and MiniLM, outperform it by a notable margin. We further find that KD yields larger gains over pretraining from scratch when the data can be repeated under the fixed computation budget.

Natural language processing has seen rapid progress over the past decade. Due to the speed of developments, some practices get established without proper evaluation. Considering one such case and focusing on reading comprehension, we ask our first research question: 1) How does the order of inputs – i.e., question and context – affect model performance? Additionally, given recent advancements in input emphasis, we ask a second research question: 2) Does emphasizing either the question, the context, or both enhance performance? Experimenting with 9 large language models across 3 datasets, we find that presenting the context before the question improves model performance, with an accuracy increase of up to 31%. Furthermore, emphasizing the context yields superior results compared to question emphasis, and in general, emphasizing parts of the input is particularly effective for addressing questions that models lack the parametric knowledge to answer. Experimenting with both prompt-based and attention-based emphasis methods, we additionally find that the best method is surprisingly simple: it only requires concatenating a few tokens to the input and results in an ac- curacy improvement of up to 36%, allowing smaller models to outperform their significantly larger counterparts.

Both NLP researchers and linguists have expressed a desire to use language technologies in language documentation, but most documentary work still proceeds without them, presenting a lost opportunity to hasten the preservation of the world’s endangered languages, such as those spoken in Latin America. In this work, we empirically measure two factors that have previously been identified as explanations of this low utilization: curricular offerings in graduate programs, and rates of interdisciplinary collaboration in publications related to NLP in language documentation. Our findings verify the claim that interdisciplinary training and collaborations are scarce and support the view that interdisciplinary curricular offerings facilitate interdisciplinary collaborations.

In this paper, we present the four systems developed by the Meenzer team from JGU for the AmericasNLP 2024 shared task on the creation of educational materials for Indigenous languages. The task involves accurately applying specific grammatical modifications to given source sentences across three low-resource Indigenous languages: Bribri, Guarani, and Maya. We train two types of model architectures: finetuning a sequence-to-sequence pointer-generator LSTM and finetuning the Mixtral 8x7B model by incorporating in-context examples into the training phase. System 1, an ensemble combining finetuned LSTMs, finetuned Mixtral models, and GPT-4, achieves the best performance on Guarani. Meanwhile, system 4, another ensemble consisting solely of fine-tuned Mixtral models, outperforms all other teams on Maya and secures the second place overall. Additionally, we conduct an ablation study to understand the performance of our system 4.

This paper presents the results of the first shared task about the creation of educational materials for three indigenous languages of the Americas.The task proposes to automatically generate variations of sentences according to linguistic features that could be used for grammar exercises.The languages involved in this task are Bribri, Maya, and Guarani.Seven teams took part in the challenge, submitting a total of 22 systems, obtaining very promising results.

This paper presents the findings of the third iteration of the AmericasNLP Shared Task on Machine Translation. This year’s competition features eleven Indigenous languages found across North, Central, and South America. A total of six teams participate with a total of 157 submissions across all languages and models. Two baselines – the Sheffield and Helsinki systems from 2023 – are provided and represent hard-to-beat starting points for the competition. In addition to the baselines, teams are given access to a new repository of training data which consists of data collected by teams in prior shared tasks. Using ChrF++ as the main competition metric, we see improvements over the baseline for 4 languages: Chatino, Guarani, Quechua, and Rarámuri, with performance increases over the best baseline of 4.2 ChrF++. In this work, we present a summary of the submitted systems, results, and a human evaluation of system outputs for Bribri, which consists of both (1) a rating of meaning and fluency and (2) a qualitative error analysis of outputs from the best submitted system.

Cross-lingual transfer can be achieved through two main approaches: zero-shot transfer or machine translation (MT). While the former has been the dominant approach, both have been shown to be competitive. In this work, we compare the current performance and long-term viability of these methods. We leverage lexical gaps to create a multilingual question answering dataset, which provides a difficult domain for evaluation. Both approaches struggle in this setting, though zero-shot transfer performs better, as current MT outputs are not specific enough for the task. Using oracle translation offers the best performance, showing that this approach can perform well long-term, however current MT quality is a bottleneck. We also conduct an exploratory study to see if humans produce translations sufficient for the task with only general instructions. We find this to be true for the majority of translators, but not all. This indicates that while translation has the potential to outperform zero-shot approaches, creating MT models that generate accurate task-specific translations may not be straightforward.

During conversations, people align to one another over time, by using similar words, concepts, and syntax. This helps form a shared understanding of the conversational content and is associated with increased engagement and satisfaction. It also affects conversation outcomes: e.g., when talking to language learners, an above normal level of linguistic alignment of parents or language teachers is correlated with faster language acquisition. These benefits make human-like alignment an important property of dialogue systems, which has often been overlooked by the NLP community. In order to fill this gap, we ask: (RQ1) Due to the importance for engagement and satisfaction, to what degree do state-of-the-art dialogue systems align to adult users? (RQ2) With a potential application to child language acquisition in mind, do systems, similar to parents, show high levels of alignment during conversations with children? Our experiments show that ChatGPT aligns to adults at roughly human levels, while Llama2 shows elevated alignment. However, when responding to a child, both systems’ alignment is below human levels.

Hyperparameter tuning, the process of searching for suitable hyperparameters, becomes more difficult as the computing resources required to train neural networks continue to grow. This topic continues to receive little attention and discussion—much of it hearsay—despite its obvious importance. We attempt to formalize hyperparameter sensitivity using two metrics: similarity-based sensitivity and performance-based sensitivity. We then use these metrics to quantify two such claims: (1) transformers are more sensitive to hyperparameter choices than LSTMs and (2) transformers are particularly sensitive to batch size. We conduct experiments on two different character-level sequence-to-sequence tasks and find that, indeed, the transformer is slightly more sensitive to hyperparameters according to both of our metrics. However, we do not find that it is more sensitive to batch size in particular.

The differences between cloze-task language model (LM) probing with 1) expert-made templates and 2) naturally-occurring text have often been overlooked. Here, we evaluate 16 different LMs on 10 probing English datasets – 4 template-based and 6 template-free – in general and biomedical domains to answer the following research questions: (RQ1) Do model rankings differ between the two approaches? (RQ2) Do models’ absolute scores differ between the two approaches? (RQ3) Do the answers to RQ1 and RQ2 differ between general and domain-specific models? Our findings are: 1) Template-free and template-based approaches often rank models differently, except for the top domain- specific models. 2) Scores decrease by up to 42% Acc@1 when comparing parallel template-free and template-based prompts. 3) Perplexity is negatively correlated with accuracy in the template-free approach, but, counter-intuitively, they are positively correlated for template-based probing. 4) Models tend to predict the same answers frequently across prompts for template-based probing, which is less common when employing template-free techniques.

Prior work has uncovered a set of common problems in state-of-the-art context-based question answering (QA) systems: a lack of attention to the context when the latter conflicts with a model’s parametric knowledge, little robustness to noise, and a lack of consistency with their answers. However, most prior work focus on one or two of those problems in isolation, which makes it difficult to see trends across them. We aim to close this gap, by first outlining a set of – previously discussed as well as novel – desiderata for QA models. We then survey relevant analysis and methods papers to provide an overview of the state of the field. The second part of our work presents experiments where we evaluate 15 QA systems on 5 datasets according to all desiderata at once. We find many novel trends, including (1) systems that are less susceptible to noise are not necessarily more consistent with their answers when given irrelevant context; (2) most systems that are more susceptible to noise are more likely to correctly answer according to a context that conflicts with their parametric knowledge; and (3) the combination of conflicting knowledge and noise can reduce system performance by up to 96%. As such, our desiderata help increase our understanding of how these models work and reveal potential avenues for improvements.

Canonical morphological segmentation is the process of analyzing words into the standard (aka underlying) forms of their constituent morphemes.This is a core task in endangered language documentation, and NLP systems have the potential to dramatically speed up this process. In typical language documentation settings, training data for canonical morpheme segmentation is scarce, making it difficult to train high quality models. However, translation data is often much more abundant, and, in this work, we present a method that attempts to leverage translation data in the canonical segmentation task. We propose a character-level sequence-to-sequence model that incorporates representations of translations obtained from pretrained high-resource monolingual language models as an additional signal. Our model outperforms the baseline in a super-low resource setting but yields mixed results on training splits with more data. Additionally, we find that we can achieve strong performance even without needing difficult-to-obtain word level alignments. While further work is needed to make translations useful in higher-resource settings, our model shows promise in severely resource-constrained settings.

Current natural language processing (NLP) research tends to focus on only one or, less frequently, two dimensions – e.g., performance, interpretability, or efficiency – at a time, which may lead to suboptimal conclusions. Work on adapter modulesfocuses on improving performance and efficiency, with no investigation of unintended consequences on other aspects such as fairness. To address this gap, we conduct experiments on three text classification datasets by either (1) finetuning all parameters or (2) using adapter modules. Regarding performance and efficiency, we confirm prior findings that the accuracy of adapter-enhanced models is roughly on par with that of fully finetuned models, while training time is substantially reduced. Regarding fairness, we show that adapter modules result in mixed fairness across sensitive groups. Further investigation reveals that, when the standard finetuned model exhibits limited biases, adapter modules typically do not introduce extra bias. On the other hand, when the finetuned model exhibits increased bias, the use of adapter modules poses the potential danger of amplifying these biases to a significant extent. Our findings highlight the need for a case-by-case evaluation rather than a one-size-fits-all judgment.

In recent years machine translation has become very successful for high-resource language pairs. This has also sparked new interest in research on the automatic translation of low-resource languages, including Indigenous languages. However, the latter are deeply related to the ethnic and cultural groups that speak (or used to speak) them. The data collection, modeling and deploying machine translation systems thus result in new ethical questions that must be addressed. Motivated by this, we first survey the existing literature on ethical considerations for the documentation, translation, and general natural language processing for Indigenous languages. Afterward, we conduct and analyze an interview study to shed light on the positions of community leaders, teachers, and language activists regarding ethical concerns for the automatic translation of their languages. Our results show that the inclusion, at different degrees, of native speakers and community members is vital to performing better and more ethical research on Indigenous languages.

Recent advances in NLP have led to a rise in inter-disciplinary and application-oriented research. While this demonstrates the growing real-world impact of the field, research papers frequently feature experiments that do not account for the complexities of realistic data and environments. To explore the extent of this gap, we investigate the relationship between the real-world motivations described in NLP papers and the models and evaluation which comprise the proposed solution. We first survey papers from the NLP Applications track from ACL 2020 and EMNLP 2020, asking which papers have differences between their stated motivation and their experimental setting, and if so, mention them. We find that many papers fall short of considering real-world input and output conditions due to adopting simplified modeling or evaluation settings. As a case study, we then empirically show that the performance of an educational dialog understanding system deteriorates when used in a realistic classroom environment.

With a growing focus on morphological inflection systems for languages where high-quality data is scarce, training data noise is a serious but so far largely ignored concern. We aim at closing this gap by investigating the types of noise encountered within a pipeline for truly unsupervised morphological paradigm completion and its impact on morphological inflection systems: First, we propose an error taxonomy and annotation pipeline for inflection training data. Then, we compare the effect of different types of noise on multiple state-of-the- art inflection models. Finally, we propose a novel character-level masked language modeling (CMLM) pretraining objective and explore its impact on the models’ resistance to noise. Our experiments show that various architectures are impacted differently by separate types of noise, but encoder-decoders tend to be more robust to noise than models trained with a copy bias. CMLM pretraining helps transformers, but has lower impact on LSTMs.

With recent advances in large language models (LLMs), the concept of automatically generating children’s educational materials has become increasingly realistic. Working toward the goal of age-appropriate simplicity in generated educational texts, we first examine the ability of several popular LLMs to generate stories with properly adjusted lexical and readability levels. We find that, in spite of the growing capabilities of LLMs, they do not yet possess the ability to limit their vocabulary to levels appropriate for younger age groups. As a second experiment, we explore the ability of state-of-the-art lexical simplification models to generalize to the domain of children’s stories and, thus, create an efficient pipeline for their automatic generation. In order to test these models, we develop a dataset of child-directed lexical simplification instances, with examples taken from the LLM-generated stories in our first experiment. We find that, while the strongest-performing current lexical simplification models do not perform as well on material designed for children due to their reliance on large language models behind the scenes, some models that still achieve fairly strong results on general data can mimic or even improve their performance on children-directed data with proper fine-tuning, which we conduct using our newly created child-directed simplification dataset.

Large multilingual models have inspired a new class of word alignment methods, which work well for the model’s pretraining languages. However, the languages most in need of automatic alignment are low-resource and, thus, not typically included in the pretraining data. In this work, we ask: How do modern aligners perform on unseen languages, and are they better than traditional methods? We contribute gold-standard alignments for Bribri–Spanish, Guarani–Spanish, Quechua–Spanish, and Shipibo-Konibo–Spanish. With these, we evaluate state-of-the-art aligners with and without model adaptation to the target language. Finally, we also evaluate the resulting alignments extrinsically through two downstream tasks: named entity recognition and part-of-speech tagging. We find that although transformer-based methods generally outperform traditional models, the two classes of approach remain competitive with each other.

Advances in conversational AI systems, powered in particular by large language models, have facilitated rapid progress in understanding and generating dialog. Typically, task-oriented or open-domain dialog systems have been designed to work with two-party dialog, i.e., the exchange of utterances between a single user and a dialog system. However, modern dialog systems may be deployed in scenarios such as classrooms or meetings where conversational analysis of multiple speakers is required. This survey will present research around computational modeling of “multi-party dialog”, outlining differences from two-party dialog, challenges and issues in working with multi-party dialog, and methods for representing multi-party dialog. We also provide an overview of dialog datasets created for the study of multi-party dialog, as well as tasks that are of interest in this domain.

This paper provides an overview of the first shared task on choosing beneficial instances for machine translation, conducted as part of the CoCo4MT 2023 Workshop at MTSummit. This shared task was motivated by the need to make the data annotation process for machine translation more efficient, particularly for low-resource languages for which collecting human translations may be difficult or expensive. The task involved developing methods for selecting the most beneficial instances for training a machine translation system without access to an existing parallel dataset in the target language, such that the best selected instances can then be manually translated. Two teams participated in the shared task, namely the Williams team and the AST team. Submissions were evaluated by training a machine translation model on each submission’s chosen instances, and comparing their performance with the chRF++ score. The system that ranked first is by the Williams team, that finds representative instances by clustering the training data.

Neural models have drastically advanced state of the art for machine translation (MT) between high-resource languages. Traditionally, these models rely on large amounts of training data, but many language pairs lack these resources. However, an important part of the languages in the world do not have this amount of data. Most languages from the Americas are among them, having a limited amount of parallel and monolingual data, if any. Here, we present an introduction to the interested reader to the basic challenges, concepts, and techniques that involve the creation of MT systems for these languages. Finally, we discuss the recent advances and findings and open questions, product of an increased interest of the NLP community in these languages.

In this work, we present the results of the AmericasNLP 2023 Shared Task on Machine Translation into Indigenous Languages of the Americas. This edition of the shared task featured eleven language pairs, one of which – Chatino-Spanish – uses a newly collected evaluation dataset, consisting of professionally translated text from the legal domain. Seven teams participated in the shared task, with a total of 181 submissions. Additionally, we conduct a human evaluation of the best system outputs, and compare them to the best submissions from the prior shared task. We find that this analysis agrees with the quantitative measures used to rank submissions, which shows further improvements of 9.64 ChrF on average across all languages, when compared to the prior winning system.

Pretrained multilingual models enable zero-shot learning even for unseen languages, and that performance can be further improved via adaptation prior to finetuning. However, it is unclear how the number of pretraining languages influences a model’s zero-shot learning for languages unseen during pretraining. To fill this gap, we ask the following research questions: (1) How does the number of pretraining languages influence zero-shot performance on unseen target languages? (2) Does the answer to that question change with model adaptation? (3) Do the findings for our first question change if the languages used for pretraining are all related? Our experiments on pretraining with related languages indicate that choosing a diverse set of languages is crucial. Without model adaptation, surprisingly, increasing the number of pretraining languages yields better results up to adding related languages, after which performance plateaus. In contrast, with model adaptation via continued pretraining, pretraining on a larger number of languages often gives further improvement, suggesting that model adaptation is crucial to exploit additional pretraining languages.

Pretrained multilingual models are able to perform cross-lingual transfer in a zero-shot setting, even for languages unseen during pretraining. However, prior work evaluating performance on unseen languages has largely been limited to low-level, syntactic tasks, and it remains unclear if zero-shot learning of high-level, semantic tasks is possible for unseen languages. To explore this question, we present AmericasNLI, an extension of XNLI (Conneau et al., 2018) to 10 Indigenous languages of the Americas. We conduct experiments with XLM-R, testing multiple zero-shot and translation-based approaches. Additionally, we explore model adaptation via continued pretraining and provide an analysis of the dataset by considering hypothesis-only models. We find that XLM-R’s zero-shot performance is poor for all 10 languages, with an average performance of 38.48%. Continued pretraining offers improvements, with an average accuracy of 43.85%. Surprisingly, training on poorly translated data by far outperforms all other methods with an accuracy of 49.12%.

Neural networks have long been at the center of a debate around the cognitive mechanism by which humans process inflectional morphology. This debate has gravitated into NLP by way of the question: Are neural networks a feasible account for human behavior in morphological inflection?We address that question by measuring the correlation between human judgments and neural network probabilities for unknown word inflections. We test a larger range of architectures than previously studied on two important tasks for the cognitive processing debate: English past tense, and German number inflection. We find evidence that the Transformer may be a better account of human behavior than LSTMs on these datasets, and that LSTM features known to increase inflection accuracy do not always result in more human-like behavior.

The field of natural language processing (NLP) has grown over the last few years: conferences have become larger, we have published an incredible amount of papers, and state-of-the-art research has been implemented in a large variety of customer-facing products. However, this paper argues that we have been less successful than we *should* have been and reflects on where and how the field fails to tap its full potential. Specifically, we demonstrate that, in recent years, **subpar time allocation has been a major obstacle for NLP research**. We outline multiple concrete problems together with their negative consequences and, importantly, suggest remedies to improve the status quo. We hope that this paper will be a starting point for discussions around which common practices are – or are *not* – beneficial for NLP research.

Morphologically-rich polysynthetic languages present a challenge for NLP systems due to data sparsity, and a common strategy to handle this issue is to apply subword segmentation. We investigate a wide variety of supervised and unsupervised morphological segmentation methods for four polysynthetic languages: Nahuatl, Raramuri, Shipibo-Konibo, and Wixarika. Then, we compare the morphologically inspired segmentation methods against Byte-Pair Encodings (BPEs) as inputs for machine translation (MT) when translating to and from Spanish. We show that for all language pairs except for Nahuatl, an unsupervised morphological segmentation algorithm outperforms BPEs consistently and that, although supervised methods achieve better segmentation scores, they under-perform in MT challenges. Finally, we contribute two new morphological segmentation datasets for Raramuri and Shipibo-Konibo, and a parallel corpus for Raramuri–Spanish.

Automatic morphological processing can aid downstream natural language processing applications, especially for low-resource languages, and assist language documentation efforts for endangered languages. Having long been multilingual, the field of computational morphology is increasingly moving towards approaches suitable for languages with minimal or no annotated resources. First, we survey recent developments in computational morphology with a focus on low-resource languages. Second, we argue that the field is ready to tackle the logical next challenge: understanding a language’s morphology from raw text alone. We perform an empirical study on a truly unsupervised version of the paradigm completion task and show that, while existing state-of-the-art models bridged by two newly proposed models we devise perform reasonably, there is still much room for improvement. The stakes are high: solving this task will increase the language coverage of morphological resources by a number of magnitudes.

Neural machine translation (MT) systems have been shown to perform poorly on low-resource language pairs, for which large-scale parallel data is unavailable. Making the data annotation process faster and cheaper is therefore important to ensure equitable access to MT systems. To make optimal use of a limited annotation budget, we present CHIA (choosing instances to annotate), a method for selecting instances to annotate for machine translation. Using an existing multi-way parallel dataset of high-resource languages, we first identify instances, based on model training dynamics, that are most informative for training MT models for high-resource languages. We find that there are cross-lingual commonalities in instances that are useful for MT model training, which we use to identify instances that will be useful to train models on a new target language. Evaluating on 20 languages from two corpora, we show that training on instances selected using our method provides an average performance improvement of 1.59 BLEU over training on randomly selected instances of the same size.

Human–computer conversation has long been an interest of artificial intelligence and natural language processing research. Recent years have seen a dramatic improvement in quality for both task-oriented and open-domain dialogue systems, and an increasing amount of research in the area. The goal of this work is threefold: (1) to provide an overview of recent advances in the field of open-domain dialogue, (2) to summarize issues related to ethics, bias, and fairness that the field has identified as well as typical errors of dialogue systems, and (3) to outline important future challenges. We hope that this work will be of interest to both new and experienced researchers in the area.

Language documentation encompasses translation, typically into the dominant high-resource language in the region where the target language is spoken. To make data accessible to a broader audience, additional translation into other high-resource languages might be needed. Working within a project documenting Kotiria, we explore the extent to which state-of-the-art machine translation (MT) systems can support this second translation – in our case from Portuguese to English. This translation task is challenging for multiple reasons: (1) the data is out-of-domain with respect to the MT system’s training data, (2) much of the data is conversational, (3) existing translations include non-standard and uncommon expressions, often reflecting properties of the documented language, and (4) the data includes borrowings from other regional languages. Despite these challenges, existing MT systems perform at a usable level, though there is still room for improvement. We then conduct a qualitative analysis and suggest ways to improve MT between high-resource languages in a language documentation setting.

Recent advances in natural language processing (NLP) have greatly helped educational applications, for both teachers and students. In higher education, there is great potential to use NLP tools for advancing pedagogical research. In this paper, we focus on how NLP can help understand student experiences in engineering, thus facilitating engineering educators to carry out large scale analysis that is helpful for re-designing the curriculum. Here, we introduce a new task we call response construct tagging (RCT), in which student responses to tailored survey questions are automatically tagged for six constructs measuring transformative experiences and engineering identity of students. We experiment with state-of-the-art classification models for this task and investigate the effects of different sources of additional information. Our best model achieves an F1 score of 48. We further investigate multi-task training on the related task of sentiment classification, which improves our model’s performance to 55 F1. Finally, we provide a detailed qualitative analysis of model performance.

We describe the second SIGMORPHON shared task on unsupervised morphology: the goal of the SIGMORPHON 2021 Shared Task on Unsupervised Morphological Paradigm Clustering is to cluster word types from a raw text corpus into paradigms. To this end, we release corpora for 5 development and 9 test languages, as well as gold partial paradigms for evaluation. We receive 14 submissions from 4 teams that follow different strategies, and the best performing system is based on adaptor grammars. Results vary significantly across languages. However, all systems are outperformed by a supervised lemmatizer, implying that there is still room for improvement.

This paper describes our system for the SIGMORPHON 2021 Shared Task on Unsupervised Morphological Paradigm Clustering, which asks participants to group inflected forms together according their underlying lemma without the aid of annotated training data. We employ agglomerative clustering to group word forms together using a metric that combines an orthographic distance and a semantic distance from word embeddings. We experiment with two variations of an edit distance-based model for quantifying orthographic distance, but, due to time constraints, our system does not improve over the shared task’s baseline system.

Pretrained multilingual models (PMMs) enable zero-shot learning via cross-lingual transfer, performing best for languages seen during pretraining. While methods exist to improve performance for unseen languages, they have almost exclusively been evaluated using amounts of raw text only available for a small fraction of the world’s languages. In this paper, we evaluate the performance of existing methods to adapt PMMs to new languages using a resource available for close to 1600 languages: the New Testament. This is challenging for two reasons: (1) the small corpus size, and (2) the narrow domain. While performance drops for all approaches, we surprisingly still see gains of up to 17.69% accuracy for part-of-speech tagging and 6.29 F1 for NER on average over all languages as compared to XLM-R. Another unexpected finding is that continued pretraining, the simplest approach, performs best. Finally, we perform a case study to disentangle the effects of domain and size and to shed light on the influence of the finetuning source language.

High-performing machine translation (MT) systems can help overcome language barriers while making it possible for everyone to communicate and use language technologies in the language of their choice. However, such systems require large amounts of parallel sentences for training, and translators can be difficult to find and expensive. Here, we present a data collection strategy for MT which, in contrast, is cheap and simple, as it does not require bilingual speakers. Based on the insight that humans pay specific attention to movements, we use graphics interchange formats (GIFs) as a pivot to collect parallel sentences from monolingual annotators. We use our strategy to collect data in Hindi, Tamil and English. As a baseline, we also collect data using images as a pivot. We perform an intrinsic evaluation by manually evaluating a subset of the sentence pairs and an extrinsic evaluation by finetuning mBART (Liu et al., 2020) on the collected data. We find that sentences collected via GIFs are indeed of higher quality.

We present the findings of the LoResMT 2021 shared task which focuses on machine translation (MT) of COVID-19 data for both low-resource spoken and sign languages. The organization of this task was conducted as part of the fourth workshop on technologies for machine translation of low resource languages (LoResMT). Parallel corpora is presented and publicly available which includes the following directions: English↔Irish, English↔Marathi, and Taiwanese Sign language↔Traditional Chinese. Training data consists of 8112, 20933 and 128608 segments, respectively. There are additional monolingual data sets for Marathi and English that consist of 21901 segments. The results presented here are based on entries from a total of eight teams. Three teams submitted systems for English↔Irish while five teams submitted systems for English↔Marathi. Unfortunately, there were no systems submissions for the Taiwanese Sign language↔Traditional Chinese task. Maximum system performance was computed using BLEU and follow as 36.0 for English–Irish, 34.6 for Irish–English, 24.2 for English–Marathi, and 31.3 for Marathi–English.

This paper presents the results of the 2021 Shared Task on Open Machine Translation for Indigenous Languages of the Americas. The shared task featured two independent tracks, and participants submitted machine translation systems for up to 10 indigenous languages. Overall, 8 teams participated with a total of 214 submissions. We provided training sets consisting of data collected from various sources, as well as manually translated sentences for the development and test sets. An official baseline trained on this data was also provided. Team submissions featured a variety of architectures, including both statistical and neural models, and for the majority of languages, many teams were able to considerably improve over the baseline. The best performing systems achieved 12.97 ChrF higher than baseline, when averaged across languages.

In contrast to their word- or sentence-level counterparts, character embeddings are still poorly understood. We aim at closing this gap with an in-depth study of English character embeddings. For this, we use resources from research on grapheme–color synesthesia – a neuropsychological phenomenon where letters are associated with colors –, which give us insight into which characters are similar for synesthetes and how characters are organized in color space. Comparing 10 different character embeddings, we ask: How similar are character embeddings to a synesthete’s perception of characters? And how similar are character embeddings extracted from different models? We find that LSTMs agree with humans more than transformers. Comparing across tasks, grapheme-to-phoneme conversion results in the most human-like character embeddings. Finally, ELMo embeddings differ from both humans and other models.

Linguistically informed analyses of language models (LMs) contribute to the understanding and improvement of such models. Here, we introduce the corpus of Chinese linguistic minimal pairs (CLiMP) to investigate what knowledge Chinese LMs acquire. CLiMP consists of sets of 1000 minimal pairs (MPs) for 16 syntactic contrasts in Chinese, covering 9 major Chinese linguistic phenomena. The MPs are semi-automatically generated, and human agreement with the labels in CLiMP is 95.8%. We evaluate 11 different LMs on CLiMP, covering n-grams, LSTMs, and Chinese BERT. We find that classifier–noun agreement and verb complement selection are the phenomena that models generally perform best at. However, models struggle the most with the ba construction, binding, and filler-gap dependencies. Overall, Chinese BERT achieves an 81.8% average accuracy, while the performances of LSTMs and 5-grams are only moderately above chance level.

Recent work has raised concerns about the inherent limitations of text-only pretraining. In this paper, we first demonstrate that reporting bias, the tendency of people to not state the obvious, is one of the causes of this limitation, and then investigate to what extent multimodal training can mitigate this issue. To accomplish this, we 1) generate the Color Dataset (CoDa), a dataset of human-perceived color distributions for 521 common objects; 2) use CoDa to analyze and compare the color distribution found in text, the distribution captured by language models, and a human’s perception of color; and 3) investigate the performance differences between text-only and multimodal models on CoDa. Our results show that the distribution of colors that a language model recovers correlates more strongly with the inaccurate distribution found in text than with the ground-truth, supporting the claim that reporting bias negatively impacts and inherently limits text-only training. We then demonstrate that multimodal models can leverage their visual training to mitigate these effects, providing a promising avenue for future research.

Intermediate-task training—fine-tuning a pretrained model on an intermediate task before fine-tuning again on the target task—often improves model performance substantially on language understanding tasks in monolingual English settings. We investigate whether English intermediate-task training is still helpful on non-English target tasks. Using nine intermediate language-understanding tasks, we evaluate intermediate-task transfer in a zero-shot cross-lingual setting on the XTREME benchmark. We see large improvements from intermediate training on the BUCC and Tatoeba sentence retrieval tasks and moderate improvements on question-answering target tasks. MNLI, SQuAD and HellaSwag achieve the best overall results as intermediate tasks, while multi-task intermediate offers small additional improvements. Using our best intermediate-task models for each target task, we obtain a 5.4 point improvement over XLM-R Large on the XTREME benchmark, setting the state of the art as of June 2020. We also investigate continuing multilingual MLM during intermediate-task training and using machine-translated intermediate-task data, but neither consistently outperforms simply performing English intermediate-task training.

Explicit mechanisms for copying have improved the performance of neural models for sequence-to-sequence tasks in the low-resource setting. However, they rely on an overlap between source and target vocabularies. Here, we propose a model that does not: a pointer-generator transformer for disjoint vocabularies. We apply our model to a low-resource version of the grapheme-to-phoneme conversion (G2P) task, and show that it outperforms a standard transformer by an average of 5.1 WER over 15 languages. While our model does not beat the the best performing baseline, we demonstrate that it provides complementary information to it: an oracle that combines the best outputs of the two models improves over the strongest baseline by 7.7 WER on average in the low-resource setting. In the high-resource setting, our model performs comparably to a standard transformer.

Pretrained language models have obtained impressive results for a large set of natural language understanding tasks. However, training these models is computationally expensive and requires huge amounts of data. Thus, it would be desirable to automatically detect groups of more or less important examples. Here, we investigate if we can leverage sources of information which are commonly overlooked, Wikipedia categories as listed in DBPedia, to identify useful or harmful data points during pretraining. We define an experimental setup in which we analyze correlations between language model perplexity on specific clusters and downstream NLP task performances during pretraining. Our experiments show that Wikipedia categories are not a good indicator of the importance of specific sentences for pretraining.

In this paper, we describe the findings of the SIGMORPHON 2020 shared task on unsupervised morphological paradigm completion (SIGMORPHON 2020 Task 2), a novel task in the field of inflectional morphology. Participants were asked to submit systems which take raw text and a list of lemmas as input, and output all inflected forms, i.e., the entire morphological paradigm, of each lemma. In order to simulate a realistic use case, we first released data for 5 development languages. However, systems were officially evaluated on 9 surprise languages, which were only revealed a few days before the submission deadline. We provided a modular baseline system, which is a pipeline of 4 components. 3 teams submitted a total of 7 systems, but, surprisingly, none of the submitted systems was able to improve over the baseline on average over all 9 test languages. Only on 3 languages did a submitted system obtain the best results. This shows that unsupervised morphological paradigm completion is still largely unsolved. We present an analysis here, so that this shared task will ground further research on the topic.

We describe the NYU-CUBoulder systems for the SIGMORPHON 2020 Task 0 on typologically diverse morphological inflection and Task 2 on unsupervised morphological paradigm completion. The former consists of generating morphological inflections from a lemma and a set of morphosyntactic features describing the target form. The latter requires generating entire paradigms for a set of given lemmas from raw text alone. We model morphological inflection as a sequence-to-sequence problem, where the input is the sequence of the lemma’s characters with morphological tags, and the output is the sequence of the inflected form’s characters. First, we apply a transformer model to the task. Second, as inflected forms share most characters with the lemma, we further propose a pointer-generator transformer model to allow easy copying of input characters.

In this paper, we present the systems of the University of Stuttgart IMS and the University of Colorado Boulder (IMS–CUBoulder) for SIGMORPHON 2020 Task 2 on unsupervised morphological paradigm completion (Kann et al., 2020). The task consists of generating the morphological paradigms of a set of lemmas, given only the lemmas themselves and unlabeled text. Our proposed system is a modified version of the baseline introduced together with the task. In particular, we experiment with substituting the inflection generation component with an LSTM sequence-to-sequence model and an LSTM pointer-generator network. Our pointer-generator system obtains the best score of all seven submitted systems on average over all languages, and outperforms the official baseline, which was best overall, on Bulgarian and Kannada.

In this paper, we describe two CU-Boulder submissions to the SIGMORPHON 2020 Task 1 on multilingual grapheme-to-phoneme conversion (G2P). Inspired by the high performance of a standard transformer model (Vaswani et al., 2017) on the task, we improve over this approach by adding two modifications: (i) Instead of training exclusively on G2P, we additionally create examples for the opposite direction, phoneme-to-grapheme conversion (P2G). We then perform multi-task training on both tasks. (ii) We produce ensembles of our models via majority voting. Our approaches, though being conceptually simple, result in systems that place 6th and 8th amongst 23 submitted systems, and obtain the best results out of all systems on Lithuanian and Modern Greek, respectively.

While pretrained models such as BERT have shown large gains across natural language understanding tasks, their performance can be improved by further training the model on a data-rich intermediate task, before fine-tuning it on a target task. However, it is still poorly understood when and why intermediate-task training is beneficial for a given target task. To investigate this, we perform a large-scale study on the pretrained RoBERTa model with 110 intermediate-target task combinations. We further evaluate all trained models with 25 probing tasks meant to reveal the specific skills that drive transfer. We observe that intermediate tasks requiring high-level inference and reasoning abilities tend to work best. We also observe that target task performance is strongly correlated with higher-level abilities such as coreference resolution. However, we fail to observe more granular correlations between probing and target task performance, highlighting the need for further work on broad-coverage probing benchmarks. We also observe evidence that the forgetting of knowledge learned during pretraining may limit our analysis, highlighting the need for further work on transfer learning methods in these settings.

We propose the task of unsupervised morphological paradigm completion. Given only raw text and a lemma list, the task consists of generating the morphological paradigms, i.e., all inflected forms, of the lemmas. From a natural language processing (NLP) perspective, this is a challenging unsupervised task, and high-performing systems have the potential to improve tools for low-resource languages or to assist linguistic annotators. From a cognitive science perspective, this can shed light on how children acquire morphological knowledge. We further introduce a system for the task, which generates morphological paradigms via the following steps: (i) EDIT TREE retrieval, (ii) additional lemma retrieval, (iii) paradigm size discovery, and (iv) inflection generation. We perform an evaluation on 14 typologically diverse languages. Our system outperforms trivial baselines with ease and, for some languages, even obtains a higher accuracy than minimally supervised systems.

Neural unsupervised parsing (UP) models learn to parse without access to syntactic annotations, while being optimized for another task like language modeling. In this work, we propose self-training for neural UP models: we leverage aggregated annotations predicted by copies of our model as supervision for future copies. To be able to use our model’s predictions during training, we extend a recent neural UP architecture, the PRPN (Shen et al., 2018a), such that it can be trained in a semi-supervised fashion. We then add examples with parses predicted by our model to our unlabeled UP training data. Our self-trained model outperforms the PRPN by 8.1% F1 and the previous state of the art by 1.6% F1. In addition, we show that our architecture can also be helpful for semi-supervised parsing in ultra-low-resource settings.

We propose a new task in the area of computational creativity: acrostic poem generation in English. Acrostic poems are poems that contain a hidden message; typically, the first letter of each line spells out a word or short phrase. We define the task as a generation task with multiple constraints: given an input word, 1) the initial letters of each line should spell out the provided word, 2) the poem’s semantics should also relate to it, and 3) the poem should conform to a rhyming scheme. We further provide a baseline model for the task, which consists of a conditional neural language model in combination with a neural rhyming model. Since no dedicated datasets for acrostic poem generation exist, we create training data for our task by first training a separate topic prediction model on a small set of topic-annotated poems and then predicting topics for additional poems. Our experiments show that the acrostic poems generated by our baseline are received well by humans and do not lose much quality due to the additional constraints. Last, we confirm that poems generated by our model are indeed closely related to the provided prompts, and that pretraining on Wikipedia can boost performance.

Canonical morphological segmentation consists of dividing words into their standardized morphemes. Here, we are interested in approaches for the task when training data is limited. We compare model performance in a simulated low-resource setting for the high-resource languages German, English, and Indonesian to experiments on new datasets for the truly low-resource languages Popoluca and Tepehua. We explore two new models for the task, borrowing from the closely related area of morphological generation: an LSTM pointer-generator and a sequence-to-sequence model with hard monotonic attention trained with imitation learning. We find that, in the low-resource setting, the novel approaches out-perform existing ones on all languages by up to 11.4% accuracy. However, while accuracy in emulated low-resource scenarios is over 50% for all languages, for the truly low-resource languages Popoluca and Tepehua, our best model only obtains 37.4% and 28.4% accuracy, respectively. Thus, we conclude that canonical segmentation is still a challenging task for low-resource languages.

An intermediate step in the linguistic analysis of an under-documented language is to find and organize inflected forms that are attested in natural speech. From this data, linguists generate unseen inflected word forms in order to test hypotheses about the language’s inflectional patterns and to complete inflectional paradigm tables. To get the data linguists spend many hours manually creating interlinear glossed texts (IGTs). We introduce a new task that speeds this process and automatically generates new morphological resources for natural language processing systems: IGT-to-paradigms (IGT2P). IGT2P generates entire morphological paradigms from IGT input. We show that existing morphological reinflection models can solve the task with 21% to 64% accuracy, depending on the language. We further find that (i) having a language expert spend only a few hours cleaning the noisy IGT data improves performance by as much as 21 percentage points, and (ii) POS tags, which are generally considered a necessary part of NLP morphological reinflection input, have no effect on the accuracy of the models considered here.

Word embeddings typically represent different meanings of a word in a single conflated vector. Empirical analysis of embeddings of ambiguous words is currently limited by the small size of manually annotated resources and by the fact that word senses are treated as unrelated individual concepts. We present a large dataset based on manual Wikipedia annotations and word senses, where word senses from different words are related by semantic classes. This is the basis for novel diagnostic tests for an embedding’s content: we probe word embeddings for semantic classes and analyze the embedding space by classifying embeddings into semantic classes. Our main findings are: (i) Information about a sense is generally represented well in a single-vector embedding – if the sense is frequent. (ii) A classifier can accurately predict whether a word is single-sense or multi-sense, based only on its embedding. (iii) Although rare senses are not well represented in single-vector embeddings, this does not have negative impact on an NLP application whose performance depends on frequent senses.

Language identification for code-switching (CS), the phenomenon of alternating between two or more languages in conversations, has traditionally been approached under the assumption of a single language per token. However, if at least one language is morphologically rich, a large number of words can be composed of morphemes from more than one language (intra-word CS). In this paper, we extend the language identification task to the subword-level, such that it includes splitting mixed words while tagging each part with a language ID. We further propose a model for this task, which is based on a segmental recurrent neural network. In experiments on a new Spanish–Wixarika dataset and on an adapted German–Turkish dataset, our proposed model performs slightly better than or roughly on par with our best baseline, respectively. Considering only mixed words, however, it strongly outperforms all baselines.

Development sets are impractical to obtain for real low-resource languages, since using all available data for training is often more effective. However, development sets are widely used in research papers that purport to deal with low-resource natural language processing (NLP). Here, we aim to answer the following questions: Does using a development set for early stopping in the low-resource setting influence results as compared to a more realistic alternative, where the number of training epochs is tuned on development languages? And does it lead to overestimation or underestimation of performance? We repeat multiple experiments from recent work on neural models for low-resource NLP and compare results for models obtained by training with and without development sets. On average over languages, absolute accuracy differs by up to 1.4%. However, for some languages and tasks, differences are as big as 18.0% accuracy. Our results highlight the importance of realistic experimental setups in the publication of low-resource NLP research results.

Recently, neural network models which automatically infer syntactic structure from raw text have started to achieve promising results. However, earlier work on unsupervised parsing shows large performance differences between non-neural models trained on corpora in different languages, even for comparable amounts of data. With that in mind, we train instances of the PRPN architecture (Shen et al., 2018)—one of these unsupervised neural network parsers—for Arabic, Chinese, English, and German. We find that (i) the model strongly outperforms trivial baselines and, thus, acquires at least some parsing ability for all languages; (ii) good hyperparameter values seem to be universal; (iii) how the model benefits from larger training set sizes depends on the corpus, with the model achieving the largest performance gains when increasing the number of sentences from 2,500 to 12,500 for English. In addition, we show that, by sharing parameters between the related languages German and English, we can improve the model’s unsupervised parsing F1 score by up to 4% in the low-resource setting.

Multi-task learning and self-training are two common ways to improve a machine learning model’s performance in settings with limited training data. Drawing heavily on ideas from those two approaches, we suggest transductive auxiliary task self-training: training a multi-task model on (i) a combination of main and auxiliary task training data, and (ii) test instances with auxiliary task labels which a single-task version of the model has previously generated. We perform extensive experiments on 86 combinations of languages and tasks. Our results are that, on average, transductive auxiliary task self-training improves absolute accuracy by up to 9.56% over the pure multi-task model for dependency relation tagging and by up to 13.03% for semantic tagging.

Morphological segmentation for polysynthetic languages is challenging, because a word may consist of many individual morphemes and training data can be extremely scarce. Since neural sequence-to-sequence (seq2seq) models define the state of the art for morphological segmentation in high-resource settings and for (mostly) European languages, we first show that they also obtain competitive performance for Mexican polysynthetic languages in minimal-resource settings. We then propose two novel multi-task training approaches—one with, one without need for external unlabeled resources—, and two corresponding data augmentation methods, improving over the neural baseline for all languages. Finally, we explore cross-lingual transfer as a third way to fortify our neural model and show that we can train one single multi-lingual model for related languages while maintaining comparable or even improved performance, thus reducing the amount of parameters by close to 75%. We provide our morphological segmentation datasets for Mexicanero, Nahuatl, Wixarika and Yorem Nokki for future research.

Motivated by recent findings on the probabilistic modeling of acceptability judgments, we propose syntactic log-odds ratio (SLOR), a normalized language model score, as a metric for referenceless fluency evaluation of natural language generation output at the sentence level. We further introduce WPSLOR, a novel WordPiece-based version, which harnesses a more compact language model. Even though word-overlap metrics like ROUGE are computed with the help of hand-written references, our referenceless methods obtain a significantly higher correlation with human fluency scores on a benchmark dataset of compressed sentences. Finally, we present ROUGE-LM, a reference-based metric which is a natural extension of WPSLOR to the case of available references. We show that ROUGE-LM yields a significantly higher correlation with human judgments than all baseline metrics, including WPSLOR on its own.

Embedding models typically associate each word with a single real-valued vector, representing its different properties. Evaluation methods, therefore, need to analyze the accuracy and completeness of these properties in embeddings. This requires fine-grained analysis of embedding subspaces. Multi-label classification is an appropriate way to do so. We propose a new evaluation method for word embeddings based on multi-label classification given a word embedding. The task we use is fine-grained name typing: given a large corpus, find all types that a name can refer to based on the name embedding. Given the scale of entities in knowledge bases, we can build datasets for this task that are complementary to the current embedding evaluation datasets in: they are very large, contain fine-grained classes, and allow the direct evaluation of embeddings without confounding factors like sentence context.

Neural part-of-speech (POS) taggers are known to not perform well with little training data. As a step towards overcoming this problem, we present an architecture for learning more robust neural POS taggers by jointly training a hierarchical, recurrent model and a recurrent character-based sequence-to-sequence network supervised using an auxiliary objective. This way, we introduce stronger character-level supervision into the model, which enables better generalization to unseen words and provides regularization, making our encoding less prone to overfitting. We experiment with three auxiliary tasks: lemmatization, character-based word autoencoding, and character-based random string autoencoding. Experiments with minimal amounts of labeled data on 34 languages show that our new architecture outperforms a single-task baseline and, surprisingly, that, on average, raw text autoencoding can be as beneficial for low-resource POS tagging as using lemma information. Our neural POS tagger closes the gap to a state-of-the-art POS tagger (MarMoT) for low-resource scenarios by 43%, even outperforming it on languages with templatic morphology, e.g., Arabic, Hebrew, and Turkish, by some margin.

Machine translation from polysynthetic to fusional languages is a challenging task, which gets further complicated by the limited amount of parallel text available. Thus, translation performance is far from the state of the art for high-resource and more intensively studied language pairs. To shed light on the phenomena which hamper automatic translation to and from polysynthetic languages, we study translations from three low-resource, polysynthetic languages (Nahuatl, Wixarika and Yorem Nokki) into Spanish and vice versa. Doing so, we find that in a morpheme-to-morpheme alignment an important amount of information contained in polysynthetic morphemes has no Spanish counterpart, and its translation is often omitted. We further conduct a qualitative analysis and, thus, identify morpheme types that are commonly hard to align or ignored in the translation process.

Neural state-of-the-art sequence-to-sequence (seq2seq) models often do not perform well for small training sets. We address paradigm completion, the morphological task of, given a partial paradigm, generating all missing forms. We propose two new methods for the minimal-resource setting: (i) Paradigm transduction: Since we assume only few paradigms available for training, neural seq2seq models are able to capture relationships between paradigm cells, but are tied to the idiosyncracies of the training set. Paradigm transduction mitigates this problem by exploiting the input subset of inflected forms at test time. (ii) Source selection with high precision (SHIP): Multi-source models which learn to automatically select one or multiple sources to predict a target inflection do not perform well in the minimal-resource setting. SHIP is an alternative to identify a reliable source if training data is limited. On a 52-language benchmark dataset, we outperform the previous state of the art by up to 9.71% absolute accuracy.

We explore the task of multi-source morphological reinflection, which generalizes the standard, single-source version. The input consists of (i) a target tag and (ii) multiple pairs of source form and source tag for a lemma. The motivation is that it is beneficial to have access to more than one source form since different source forms can provide complementary information, e.g., different stems. We further present a novel extension to the encoder-decoder recurrent neural architecture, consisting of multiple encoders, to better solve the task. We show that our new architecture outperforms single-source reinflection models and publish our dataset for multi-source morphological reinflection to facilitate future research.

Multi-task training is an effective method to mitigate the data sparsity problem. It has recently been applied for cross-lingual transfer learning for paradigm completion—the task of producing inflected forms of lemmata—with sequence-to-sequence networks. However, it is still vague how the model transfers knowledge across languages, as well as if and which information is shared. To investigate this, we propose a set of data-dependent experiments using an existing encoder-decoder recurrent neural network for the task. Our results show that indeed the performance gains surpass a pure regularization effect and that knowledge about language and morphology can be transferred.

We present a semi-supervised way of training a character-based encoder-decoder recurrent neural network for morphological reinflection—the task of generating one inflected wordform from another. This is achieved by using unlabeled tokens or random strings as training data for an autoencoding task, adapting a network for morphological reinflection, and performing multi-task training. We thus use limited labeled data more effectively, obtaining up to 9.92% improvement over state-of-the-art baselines for 8 different languages.

We present a novel cross-lingual transfer method for paradigm completion, the task of mapping a lemma to its inflected forms, using a neural encoder-decoder model, the state of the art for the monolingual task. We use labeled data from a high-resource language to increase performance on a low-resource language. In experiments on 21 language pairs from four different language families, we obtain up to 58% higher accuracy than without transfer and show that even zero-shot and one-shot learning are possible. We further find that the degree of language relatedness strongly influences the ability to transfer morphological knowledge.