Although the advancements of pre-trained Large Language Models have significantly accelerated recent progress in NLP, their ever-increasing size poses significant challenges for conventional fine-tuning, especially in memory-intensive tasks. We investigate the potential of Parameter-Efficient Fine-Tuning, focusing on Low-Rank Adaptation (LoRA), in the domain of multilingual summarization, a task that is both challenging (due to typically long inputs), and relatively unexplored. We conduct an extensive study across different data availability scenarios, including high- and low-data settings, and cross-lingual transfer, leveraging models of different sizes. Our findings reveal that LoRA is competitive with full fine-tuning when trained with high quantities of data, and excels in low-data scenarios and cross-lingual transfer. We also study different strategies for few-shot cross-lingual transfer, finding that continued LoRA tuning outperforms full fine-tuning and the dynamic composition of language-specific LoRA modules.

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.

Standard autoregressive language models perform only polynomial-time computation to compute the probability of the next symbol. While this is attractive, it means they cannot model distributions whose next-symbol probability is hard to compute. Indeed, they cannot even model them well enough to solve associated easy decision problems for which an engineer might want to consult a language model. These limitations apply no matter how much computation and data are used to train the model, unless the model is given access to oracle parameters that grow superpolynomially in sequence length. Thus, simply training larger autoregressive language models is not a panacea for NLP. Alternatives include energy-based models (which give up efficient sampling) and latent-variable autoregressive models (which give up efficient scoring of a given string). Both are powerful enough to escape the above limitations.

We introduce neural finite state transducers (NFSTs), a family of string transduction models defining joint and conditional probability distributions over pairs of strings. The probability of a string pair is obtained by marginalizing over all its accepting paths in a finite state transducer. In contrast to ordinary weighted FSTs, however, each path is scored using an arbitrary function such as a recurrent neural network, which breaks the usual conditional independence assumption (Markov property). NFSTs are more powerful than previous finite-state models with neural features (Rastogi et al., 2016.) We present training and inference algorithms for locally and globally normalized variants of NFSTs. In experiments on different transduction tasks, they compete favorably against seq2seq models while offering interpretable paths that correspond to hard monotonic alignments.

We introduce neural particle smoothing, a sequential Monte Carlo method for sampling annotations of an input string from a given probability model. In contrast to conventional particle filtering algorithms, we train a proposal distribution that looks ahead to the end of the input string by means of a right-to-left LSTM. We demonstrate that this innovation can improve the quality of the sample. To motivate our formal choices, we explain how neural transduction models and our sampler can be viewed as low-dimensional but nonlinear approximations to working with HMMs over very large state spaces.