SIGMORPHON
Organizers: Garrett Nicolai, Eleanor Chodroff, Çağrı Çöltekin, Fred Mailhot
Workshop Papers
Authors: Matan Eyal, Hila Noga, Roee Aharoni, Idan Szpektor, Reut Tsarfaty
Recent work attributes progress in NLP to large language models (LMs) with increased model size and large quantities of pretraining data. Despite this, current state-of-the-art LMs for Hebrew are both under-parameterized and under-trained compared to LMs in other languages. Additionally, previous work on pretrained Hebrew LMs focused on encoder-only models. While the encoder-only architecture is beneficial for classification tasks, it does not cater well for sub-word prediction tasks, such as Named Entity Recognition, when considering the morphologically rich nature of Hebrew. In this paper we argue that sequence-to-sequence generative architectures are more suitable for large LMs in morphologically rich languages (MRLs) such as Hebrew. We demonstrate this by casting tasks in the Hebrew NLP pipeline as text-to-text tasks, for which we can leverage powerful multilingual, pretrained sequence-to-sequence models as mT5, eliminating the need for a separate, specialized, morpheme-based, decoder. Using this approach, our experiments show substantial improvements over previously published results on all existing Hebrew NLP benchmarks. These results suggest that multilingual sequence-to-sequence models present a promising building block for NLP for MRLs.
Go to PaperAuthors: Gulinigeer Abudouwaili, Kahaerjiang Abiderexiti, Nian Yi, Aishan Wumaier
Due to the lack of data resources, rule-based or transfer learning is mainly used in the morphological tagging of low-resource languages. However, these methods require expert knowledge, ignore contextual features, and have error propagation. Therefore, we propose a joint morphological tagger for low-resource agglutinative languages to alleviate the above challenges. First, we represent the contextual input with multi-dimensional features of agglutinative words. Second, joint training reduces the direct impact of part-of-speech errors on morphological features and increases the indirect influence between the two types of labels through a fusion mechanism. Finally, our model separately predicts part-of-speech and morphological features. Part-of-speech tagging is regarded as sequence tagging. When predicting morphological features, two-label adjacency graphs are dynamically reconstructed by integrating multilingual global features and monolingual local features. Then, a graph convolution network is used to learn the higher-order intersection of labels. A series of experiments show that the proposed model in this paper is superior to other comparative models.
Go to PaperAuthors: Bhargav Shandilya, Alexis Palmer
Linguistic analysis is a core task in the process of documenting, analyzing, and describing endangered and less-studied languages. In addition to providing insight into the properties of the language being studied, having tools to automatically label words in a language for grammatical category and morphological features can support a range of applications useful for language pedagogy and revitalization. At the same time, most modern NLP methods for these tasks require both large amounts of data in the language and compute costs well beyond the capacity of most research groups and language communities. In this paper, we present a gloss-to-gloss (g2g) model for linguistic analysis (specifically, morphological analysis and part-of-speech tagging) that is lightweight in terms of both data requirements and computational expense. The model is designed for the interlinear glossed text (IGT) format, in which we expect the source text of a sentence in a low-resource language, a translation of that sentence into a language of wider communication, and a detailed glossing of the morphological properties of each word in the sentence. We first produce silver standard parallel glossed data by automatically labeling the high-resource translation. The model then learns to transform source language morphological labels into output labels for the target language, mediated by a structured linguistic representation layer. We test the model on both low-resource and high-resource languages, and find that our simple CNN-based model achieves comparable performance to a state-of-the-art transformer-based model, at a fraction of the computational cost.
Go to PaperAuthors: Yves Scherrer
This paper evaluates various character alignment methods on the task of sentence-level standardization of dialect transcriptions. We compare alignment methods from different scientific traditions (dialectometry, speech processing, machine translation) and apply them to Finnish, Norwegian and Swiss German dialect datasets. In the absence of gold alignments, we evaluate the methods on a set of characteristics that are deemed undesirable for the task. We find that trained alignment methods only show marginal benefits to simple Levenshtein distance. On this particular task, eflomal outperforms related methods such as GIZA++ or fast_align by a large margin.
Go to PaperAuthors: Cheonkam Jeong, Dominic Schmitz, Akhilesh Kakolu Ramarao, Anna Stein, Kevin Tang
This paper presents our submission to the SIGMORPHON 2023 task 2 of Cognitively Plausible Morphophonological Generalization in Korean. We implemented both Linear Discriminative Learning and Transformer models and found that the Linear Discriminative Learning model trained on a combination of corpus and experimental data showed the best performance with the overall accuracy of around 83%. We found that the best model must be trained on both corpus data and the experimental data of one particular participant. Our examination of speaker-variability and speaker-specific information did not explain why a particular participant combined well with the corpus data. We recommend Linear Discriminative Learning models as a future non-neural baseline system, owning to its training speed, accuracy, model interpretability and cognitive plausibility. In order to improve the model performance, we suggest using bigger data and/or performing data augmentation and incorporating speaker- and item-specifics considerably.
Go to PaperAuthors: Edith Coates
This paper presents my submission to Track 1 of the 2023 SIGMORPHON shared task on interlinear glossed text (IGT). There are a wide amount of techniques for building and training IGT models (see Moeller and Hulden, 2018; McMillan-Major, 2020; Zhao et al., 2020). I describe my ensembled sequence-to-sequence approach, perform experiments, and share my submission's test-set accuracy. I also discuss future areas of research in low-resource token classification methods for IGT.
Go to PaperAuthors: Gal Astrach, Yuval Pinter
This paper presents the submission by the MeLeL team to the SIGMORPHON–UniMorph Shared Task on Typologically Diverse and Acquisition-Inspired Morphological Inflection Generation Part 3: Models of Acquisition of Inflectional Noun Morphology in Polish, Estonian, and Finnish. This task requires us to produce the word form given a lemma and a grammatical case, while trying to produce the same error-rate as in children. We approach this task with a reduced-size character-based transformer model, multilingual training and an upsampling method to introduce bias.
Go to PaperAuthors: Arya D. McCarthy, Jackson L. Lee, Alexandra DeLucia, Travis Bartley, Milind Agarwal, Lucas F.E. Ashby, Luca Del Signore, Cameron Gibson, Reuben Raff, Winston Wu
Grapheme-to-phoneme conversion is an important component in many speech technologies, but until recently there were no multilingual benchmarks for this task. The third iteration of the SIGMORPHON shared task on multilingual grapheme-to-phoneme conversion features many improvements from the previous year’s task (Ashby et al., 2021), including additional languages, three subtasks varying the amount of available resources, extensive quality assurance procedures, and automated error analyses. Three teams submitted a total of fifteen systems, at best achieving relative reductions of word error rate of 14% in the crosslingual subtask and 14% in the very-low resource subtask. The generally consistent result is that cross-lingual transfer substantially helps grapheme-to-phoneme modeling, but not to the same degree as in-language examples.
Go to PaperAuthors: Leander Girrbach
This paper describes our participation in the Third SIGMORPHON Shared Task on Grapheme-to-Phoneme Conversion (Low-Resource and Cross-Lingual) (McCarthy et al.,2022). Our models rely on different sequence labelling methods. The main model predicts multiple phonemes from each grapheme and is trained using CTC loss (Graves et al., 2006). We find that sequence labelling methods yield worse performance than the baseline when enough data is available, but can still be used when very little data is available. Furthermore, we demonstrate that alignments learned by the sequence labelling models can be easily inspected.
Go to PaperAuthors: Michael Hammond
In this paper we explore a very simple nonneural approach to mapping orthography to phonetic transcription in a low-resource context with transfer data from a related language. We start from a baseline system and focus our efforts on data augmentation. We make three principal moves. First, we start with an HMMbased system (Novak et al., 2012). Second, we augment our basic system by recombining legal substrings in restricted fashion (Ryan and Hulden, 2020). Finally, we limit our transfer data by only using training pairs where the phonetic form shares all bigrams with the target language.
Go to PaperAuthors: Chu-Cheng Lin Lin
We propose a universal grapheme-phoneme transduction model using neuralized finite-state transducers. Many computational models of grapheme-phoneme transduction nowadays are based on the (autoregressive) sequence-to-sequence string transduction paradigm. While such models have achieved state-of-the-art performance, they suffer from theoretical limitations of autoregressive models. On the other hand, neuralized finite-state transducers (NFSTs) have shown promising results on various string transduction tasks. NFSTs can be seen as a generalization of weighted finite-state transducers (WFSTs), and can be seen as pairs of a featurized finite-state machine (‘marked finite-state transducer’ or MFST in NFST terminology), and a string scoring function. Instead of taking a product of local contextual feature weights on FST arcs, NFSTs can employ arbitrary scoring functions to weight global contextual features of a string transduction, and therefore break the Markov property. Furthermore, NFSTs can be formally shown to be more expressive than (autoregressive) seq2seq models. Empirically, joint grapheme-phoneme transduction NFSTs have consistently outperformed vanilla seq2seq models on grapheme-tophoneme and phoneme-to-grapheme transduction tasks for English. Furthermore, they provide interpretable aligned string transductions, thanks to their finite-state machine component. In this talk, we propose a multilingual extension of the joint grapheme-phoneme NFST. We achieve this goal by modeling typological and phylogenetic features of languages and scripts as optional latent variables using a finite-state machine. The result is a versatile graphemephoneme transduction model: in addition to standard monolingual and multilingual transduction, the proposed multilingual NFST can also be used in various controlled generation scenarios, such as phoneme-to-grapheme transduction of an unseen language-script pair. We also plan to release an NFST software package.
Go to PaperAuthors: Dan Garrette
Grapheme-to-phoneme (G2P) conversion is a task that is inherently related to both written and spoken language. Therefore, our submission to the G2P shared task builds off of mSLAM (Bapna et al., 2022), a 600M parameter encoder model pretrained simultaneously on text from 101 languages and speech from 51 languages. For fine-tuning a G2P model, we combined mSLAM’s text encoder, which uses characters as its input tokens, with an uninitialized single-layer RNN-T decoder (Graves, 2012) whose vocabulary is the set of all 381 phonemes appearing in the shared task data. We took an explicitly multilingual approach to modeling the G2P tasks, fine-tuning and evaluating a single model that covered all the languages in each task, and adding language codes as prefixes to the input strings as a means of specifying the language of each example. Our models perform well in the shared task’s “high” setting (in which they were trained on 1,000 words from each language), though they do poorly in the “low” task setting (training on only 100 words from each language). Our models also perform reasonably in the “mixed” setting (training on 100 words in the target language and 1000 words in a related language), hinting that mSLAM’s multilingual pretraining may be enabling useful cross-lingual sharing.
Go to PaperAuthors: Michael Ginn, Sarah Moeller, Alexis Palmer, Anna Stacey, Garrett Nicolai, Mans Hulden, Miikka Silfverberg
This paper presents the findings of the SIGMORPHON 2023 Shared Task on Interlinear Glossing. This first iteration of the shared task explores glossing of a set of six typologically diverse languages: Arapaho, Gitksan, Lezgi, Natügu, Tsez and Uspanteko. The shared task encompasses two tracks: a resource-scarce closed track and an open track, where participants are allowed to utilize external data resources. Five teams participated in the shared task. The winning team Tü-CL achieved a 23.99%-point improvement over a baseline RoBERTa system in the closed track and a 17.42%-point improvement in the open track.
Go to PaperAuthors: Omer Goldman, Khuyagbaatar Batsuren, Salam Khalifa, Aryaman Arora, Garrett Nicolai, Reut Tsarfaty, Ekaterina Vylomova
The 2023 SIGMORPHON–UniMorph shared task on typologically diverse morphological inflection included a wide range of languages: 26 languages from 9 primary language families. The data this year was all lemma-split, to allow testing models' generalization ability, and structured along the new hierarchical schema presented in (Batsuren et al., 2022). The systems submitted this year, 9 in number, showed ingenuity and innovativeness, including hard attention for explainability and bidirectional decoding. Special treatment was also given by many participants to the newly-introduced data in Japanese, due to the high abundance of unseen Kanji characters in its test set.
Go to PaperAuthors: Canaan Breiss, Jinyoung Jo
This paper summarises data collection and curation for Part 2 of the 2023 SIGMORPHON-UniMorph Shared Task 0, which focused on modeling speaker knowledge and generalization of a pair of interacting phonological processes in Korean. We briefly describe how modeling the generalization task could be of interest to researchers in both Natural Language Processing and linguistics, and then summarise the traditional description of the phonological processes that are at the center of the modeling challenge. We then describe the criteria we used to select and code cases of process application in two Korean speech corpora, which served as the primary learning data. We also report the technical details of the experiment we carried out that served as the primary test data.
Go to Paper