Jan Buys

Francois Meyer, Jan Buys

Subwords have become the standard units of text in NLP, enabling efficient open-vocabulary models. With algorithms like byte-pair encoding (BPE), subword segmentation is viewed as a preprocessing step applied to the corpus before training. This can lead to sub-optimal segmentations for low-resource languages with complex morphologies. We propose a subword segmental language model (SSLM) that learns how to segment words while being trained for autoregressive language modelling. By unifying subword segmentation and language modelling, our model learns subwords that optimise LM performance. We train our model on the 4 Nguni languages of South Africa. These are low-resource agglutinative languages, so subword information is critical. As an LM, SSLM outperforms existing approaches such as BPE-based models on average across the 4 languages. Furthermore, it outperforms standard subword segmenters on unsupervised morphological segmentation. We also train our model as a word-level sequence model, resulting in an unsupervised morphological segmenter that outperforms existing methods by a large margin for all 4 languages. Our results show that learning subword segmentation is an effective alternative to existing subword segmenters, enabling the model to discover morpheme-like subwords that improve its LM capabilities.

Khalid N. Elmadani, Francois Meyer, Jan Buys

The paper describes the University of Cape Town's submission to the constrained track of the WMT22 Shared Task: Large-Scale Machine Translation Evaluation for African Languages. Our system is a single multilingual translation model that translates between English and 8 South / South East African Languages, as well as between specific pairs of the African languages. We used several techniques suited for low-resource machine translation (MT), including overlap BPE, back-translation, synthetic training data generation, and adding more translation directions during training. Our results show the value of these techniques, especially for directions where very little or no bilingual training data is available.

Anri Lombard, Simbarashe Mawere, Temi Aina et al.

Decoder-only language models can be adapted to diverse tasks through instruction finetuning, but the extent to which this generalizes at small scale for low-resource languages remains unclear. We focus on the languages of South Africa, where we are not aware of a publicly available decoder-only model that explicitly targets all eleven official written languages, nine of which are low-resource. We introduce MzansiText, a curated multilingual pretraining corpus with a reproducible filtering pipeline, and MzansiLM, a 125M-parameter language model trained from scratch. We evaluate MzansiLM on natural language understanding and generation using three adaptation regimes: monolingual task-specific finetuning, multilingual task-specific finetuning, and general multi-task instruction finetuning. Monolingual task-specific finetuning achieves strong performance on data-to-text generation, reaching 20.65 BLEU on isiXhosa and competing with encoder-decoder baselines over ten times larger. Multilingual task-specific finetuning benefits closely related languages on topic classification, achieving 78.5% macro-F1 on isiXhosa news classification. While MzansiLM adapts effectively to supervised NLU and NLG tasks, few-shot reasoning remains challenging at this model size, with performance near chance even for much larger decoder-only models. We release MzansiText and MzansiLM to provide a reproducible decoder-only baseline and clear guidance on adaptation strategies for South African languages at small scale.

Francois Meyer, Jan Buys

Subword segmentation is typically applied in preprocessing and stays fixed during training. Alternatively, it can be learned during training to optimise the training objective. In this paper we study the learning dynamics of subword segmentation: if a language model can dynamically optimise tokenisation, how do its subwords evolve during pretraining and finetuning? To explore this, we extend the subword segmental language model (SSLM), a framework for learning subwords during training, to support pretraining and finetuning. We train models for three typologically diverse languages to study learning dynamics across the morphological spectrum: Isi-Xhosa is conjunctive (long word forms composed of many morphemes), Setswana is disjunctive (morphemes written as separate words), and English represents a typological middle ground. We analyse subword dynamics from a linguistic perspective, tracking morphology, productivity, and fertility. We identify four stages of subword learning, with the morphologically complex isi-Xhosa exhibiting greater instability. During finetuning, subword boundaries shift to become finer-grained. Lastly, we show that learnable subwords offers a promising approach to improve text generation and cross-lingual transfer for low-resource, morphologically complex languages.

Francois Meyer, Jan Buys

Most data-to-text datasets are for English, so the difficulties of modelling data-to-text for low-resource languages are largely unexplored. In this paper we tackle data-to-text for isiXhosa, which is low-resource and agglutinative. We introduce Triples-to-isiXhosa (T2X), a new dataset based on a subset of WebNLG, which presents a new linguistic context that shifts modelling demands to subword-driven techniques. We also develop an evaluation framework for T2X that measures how accurately generated text describes the data. This enables future users of T2X to go beyond surface-level metrics in evaluation. On the modelling side we explore two classes of methods - dedicated data-to-text models trained from scratch and pretrained language models (PLMs). We propose a new dedicated architecture aimed at agglutinative data-to-text, the Subword Segmental Pointer Generator (SSPG). It jointly learns to segment words and copy entities, and outperforms existing dedicated models for 2 agglutinative languages (isiXhosa and Finnish). We investigate pretrained solutions for T2X, which reveals that standard PLMs come up short. Fine-tuning machine translation models emerges as the best method overall. These findings underscore the distinct challenge presented by T2X: neither well-established data-to-text architectures nor customary pretrained methodologies prove optimal. We conclude with a qualitative analysis of generation errors and an ablation study.

Francois Meyer, Jan Buys

Multilingual modelling can improve machine translation for low-resource languages, partly through shared subword representations. This paper studies the role of subword segmentation in cross-lingual transfer. We systematically compare the efficacy of several subword methods in promoting synergy and preventing interference across different linguistic typologies. Our findings show that subword regularisation boosts synergy in multilingual modelling, whereas BPE more effectively facilitates transfer during cross-lingual fine-tuning. Notably, our results suggest that differences in orthographic word boundary conventions (the morphological granularity of written words) may impede cross-lingual transfer more significantly than linguistic unrelatedness. Our study confirms that decisions around subword modelling can be key to optimising the benefits of multilingual modelling.

Berta Franzluebbers, Donald Dunagan, Miloš Stanojević et al.

Humans understand sentences word-by-word, in the order that they hear them. This incrementality entails resolving temporary ambiguities about syntactic relationships. We investigate how humans process these syntactic ambiguities by correlating predictions from incremental generative dependency parsers with timecourse data from people undergoing functional neuroimaging while listening to an audiobook. In particular, we compare competing hypotheses regarding the number of developing syntactic analyses in play during word-by-word comprehension: one vs more than one. This comparison involves evaluating syntactic surprisal from a state-of-the-art dependency parser with LLM-adapted encodings against an existing fMRI dataset. In both English and Chinese data, we find evidence for multipath parsing. Brain regions associated with this multipath effect include bilateral superior temporal gyrus.

Francois Meyer, Jan Buys

Subword segmenters like BPE operate as a preprocessing step in neural machine translation and other (conditional) language models. They are applied to datasets before training, so translation or text generation quality relies on the quality of segmentations. We propose a departure from this paradigm, called subword segmental machine translation (SSMT). SSMT unifies subword segmentation and MT in a single trainable model. It learns to segment target sentence words while jointly learning to generate target sentences. To use SSMT during inference we propose dynamic decoding, a text generation algorithm that adapts segmentations as it generates translations. Experiments across 6 translation directions show that SSMT improves chrF scores for morphologically rich agglutinative languages. Gains are strongest in the very low-resource scenario. SSMT also learns subwords that are closer to morphemes compared to baselines and proves more robust on a test set constructed for evaluating morphological compositional generalisation.

Stuart Mesham, Luc Hayward, Jared Shapiro et al.

Language models are the foundation of current neural network-based models for natural language understanding and generation. However, research on the intrinsic performance of language models on African languages has been extremely limited, which is made more challenging by the lack of large or standardised training and evaluation sets that exist for English and other high-resource languages. In this paper, we evaluate the performance of open-vocabulary language models on low-resource South African languages, using byte-pair encoding to handle the rich morphology of these languages. We evaluate different variants of n-gram models, feedforward neural networks, recurrent neural networks (RNNs), and Transformers on small-scale datasets. Overall, well-regularized RNNs give the best performance across two isiZulu and one Sepedi datasets. Multilingual training further improves performance on these datasets. We hope that this research will open new avenues for research into multilingual and low-resource language modelling for African languages.

Tumi Moeng, Sheldon Reay, Aaron Daniels et al.

Morphological Segmentation involves decomposing words into morphemes, the smallest meaning-bearing units of language. This is an important NLP task for morphologically-rich agglutinative languages such as the Southern African Nguni language group. In this paper, we investigate supervised and unsupervised models for two variants of morphological segmentation: canonical and surface segmentation. We train sequence-to-sequence models for canonical segmentation, where the underlying morphemes may not be equal to the surface form of the word, and Conditional Random Fields (CRF) for surface segmentation. Transformers outperform LSTMs with attention on canonical segmentation, obtaining an average F1 score of 72.5% across 4 languages. Feature-based CRFs outperform bidirectional LSTM-CRFs to obtain an average of 97.1% F1 on surface segmentation. In the unsupervised setting, an entropy-based approach using a character-level LSTM language model fails to outperforms a Morfessor baseline, while on some of the languages neither approach performs much better than a random baseline. We hope that the high performance of the supervised segmentation models will help to facilitate the development of better NLP tools for Nguni languages.

Peter West, Ari Holtzman, Jan Buys et al.

The principle of the Information Bottleneck (Tishby et al. 1999) is to produce a summary of information X optimized to predict some other relevant information Y. In this paper, we propose a novel approach to unsupervised sentence summarization by mapping the Information Bottleneck principle to a conditional language modelling objective: given a sentence, our approach seeks a compressed sentence that can best predict the next sentence. Our iterative algorithm under the Information Bottleneck objective searches gradually shorter subsequences of the given sentence while maximizing the probability of the next sentence conditioned on the summary. Using only pretrained language models with no direct supervision, our approach can efficiently perform extractive sentence summarization over a large corpus. Building on our unsupervised extractive summarization (BottleSumEx), we then present a new approach to self-supervised abstractive summarization (BottleSumSelf), where a transformer-based language model is trained on the output summaries of our unsupervised method. Empirical results demonstrate that our extractive method outperforms other unsupervised models on multiple automatic metrics. In addition, we find that our self-supervised abstractive model outperforms unsupervised baselines (including our own) by human evaluation along multiple attributes.

Saadia Gabriel, Antoine Bosselut, Jeff Da et al.

We introduce a general framework for abstractive summarization with factual consistency and distinct modeling of the narrative flow in an output summary. Our work addresses current limitations of models for abstractive summarization that often hallucinate information or generate summaries with coherence issues. To generate abstractive summaries with factual consistency and narrative flow, we propose Cooperative Generator -- Discriminator Networks (Co-opNet), a novel transformer-based framework where a generator works with a discriminator architecture to compose coherent long-form summaries. We explore four different discriminator objectives which each capture a different aspect of coherence, including whether salient spans of generated abstracts are hallucinated or appear in the input context, and the likelihood of sentence adjacency in generated abstracts. We measure the ability of Co-opNet to learn these objectives with arXiv scientific papers, using the abstracts as a proxy for gold long-form scientific article summaries. Empirical results from automatic and human evaluations demonstrate that Co-opNet learns to summarize with considerably improved global coherence compared to competitive baselines.

Valerie Hajdik, Jan Buys, Michael W. Goodman et al.

We propose neural models to generate high-quality text from structured representations based on Minimal Recursion Semantics (MRS). MRS is a rich semantic representation that encodes more precise semantic detail than other representations such as Abstract Meaning Representation (AMR). We show that a sequence-to-sequence model that maps a linearization of Dependency MRS, a graph-based representation of MRS, to English text can achieve a BLEU score of 66.11 when trained on gold data. The performance can be improved further using a high-precision, broad coverage grammar-based parser to generate a large silver training corpus, achieving a final BLEU score of 77.17 on the full test set, and 83.37 on the subset of test data most closely matching the silver data domain. Our results suggest that MRS-based representations are a good choice for applications that need both structured semantics and the ability to produce natural language text as output.

Ari Holtzman, Jan Buys, Li Du et al.

Despite considerable advancements with deep neural language models, the enigma of neural text degeneration persists when these models are tested as text generators. The counter-intuitive empirical observation is that even though the use of likelihood as training objective leads to high quality models for a broad range of language understanding tasks, using likelihood as a decoding objective leads to text that is bland and strangely repetitive. In this paper, we reveal surprising distributional differences between human text and machine text. In addition, we find that decoding strategies alone can dramatically effect the quality of machine text, even when generated from exactly the same neural language model. Our findings motivate Nucleus Sampling, a simple but effective method to draw the best out of neural generation. By sampling text from the dynamic nucleus of the probability distribution, which allows for diversity while effectively truncating the less reliable tail of the distribution, the resulting text better demonstrates the quality of human text, yielding enhanced diversity without sacrificing fluency and coherence.

Ari Holtzman, Jan Buys, Maxwell Forbes et al.

Recurrent Neural Networks (RNNs) are powerful autoregressive sequence models, but when used to generate natural language their output tends to be overly generic, repetitive, and self-contradictory. We postulate that the objective function optimized by RNN language models, which amounts to the overall perplexity of a text, is not expressive enough to capture the notion of communicative goals described by linguistic principles such as Grice's Maxims. We propose learning a mixture of multiple discriminative models that can be used to complement the RNN generator and guide the decoding process. Human evaluation demonstrates that text generated by our system is preferred over that of baselines by a large margin and significantly enhances the overall coherence, style, and information content of the generated text.

Jan Buys, Phil Blunsom

Parsing sentences to linguistically-expressive semantic representations is a key goal of Natural Language Processing. Yet statistical parsing has focused almost exclusively on bilexical dependencies or domain-specific logical forms. We propose a neural encoder-decoder transition-based parser which is the first full-coverage semantic graph parser for Minimal Recursion Semantics (MRS). The model architecture uses stack-based embedding features, predicting graphs jointly with unlexicalized predicates and their token alignments. Our parser is more accurate than attention-based baselines on MRS, and on an additional Abstract Meaning Representation (AMR) benchmark, and GPU batch processing makes it an order of magnitude faster than a high-precision grammar-based parser. Further, the 86.69% Smatch score of our MRS parser is higher than the upper-bound on AMR parsing, making MRS an attractive choice as a semantic representation.

Lei Yu, Jan Buys, Phil Blunsom

We introduce an online neural sequence to sequence model that learns to alternate between encoding and decoding segments of the input as it is read. By independently tracking the encoding and decoding representations our algorithm permits exact polynomial marginalization of the latent segmentation during training, and during decoding beam search is employed to find the best alignment path together with the predicted output sequence. Our model tackles the bottleneck of vanilla encoder-decoders that have to read and memorize the entire input sequence in their fixed-length hidden states before producing any output. It is different from previous attentive models in that, instead of treating the attention weights as output of a deterministic function, our model assigns attention weights to a sequential latent variable which can be marginalized out and permits online generation. Experiments on abstractive sentence summarization and morphological inflection show significant performance gains over the baseline encoder-decoders.

Jan Buys, Jan A. Botha

Morphologically rich languages often lack the annotated linguistic resources required to develop accurate natural language processing tools. We propose models suitable for training morphological taggers with rich tagsets for low-resource languages without using direct supervision. Our approach extends existing approaches of projecting part-of-speech tags across languages, using bitext to infer constraints on the possible tags for a given word type or token. We propose a tagging model using Wsabie, a discriminative embedding-based model with rank-based learning. In our evaluation on 11 languages, on average this model performs on par with a baseline weakly-supervised HMM, while being more scalable. Multilingual experiments show that the method performs best when projecting between related language pairs. Despite the inherently lossy projection, we show that the morphological tags predicted by our models improve the downstream performance of a parser by +0.6 LAS on average.

Jan Buys, Phil Blunsom

We propose a simple, scalable, fully generative model for transition-based dependency parsing with high accuracy. The model, parameterized by Hierarchical Pitman-Yor Processes, overcomes the limitations of previous generative models by allowing fast and accurate inference. We propose an efficient decoding algorithm based on particle filtering that can adapt the beam size to the uncertainty in the model while jointly predicting POS tags and parse trees. The UAS of the parser is on par with that of a greedy discriminative baseline. As a language model, it obtains better perplexity than a n-gram model by performing semi-supervised learning over a large unlabelled corpus. We show that the model is able to generate locally and syntactically coherent sentences, opening the door to further applications in language generation.