Discourse segmentation received increased attention in the past years, however the majority of studies have focused on written genres and with high-resource languages. This paper investigates discourse segmentation of a Taiwan Southern Min spontaneous speech corpus. We compare the fine-tuning a Language Model (LLM using two approaches: supervised, thanks to a high-quality annotated dataset, and weakly-supervised, requiring only a small amount of manual labeling. The corpus used here is transcribed with both Chinese characters and romanized transcription. This allows us to compare the impact of the written form on the discourse segmentation task. Additionally, the dataset includes manual prosodic breaks labeling, allowing an exploration of the role prosody can play in contemporary discourse segmentation systems grounded in LLMs. In our study, the supervised approach outperforms weak-supervision ; character-based version demonstrated better scores compared to the romanized version; and prosodic information proved to be an interesting source to increase discourse segmentation performance.

BabyLM paves the way for a range of experiments aimed at better understanding language models (LMs) and the differences and similarities between human and artificial language learning. However, the current framework is limited to the English language and a narrow but significant range of evaluation metrics, primarily focused on syntax, semantics, and pragmatics. In this paper, we propose some steps towards extending the framework to other languages, specifically Mandarin Chinese and French, leveraging existing linguistic resources for these languages. Additionally, we advocate for greater exploration of genre variations within subcorpora for training LMs, as well as for the adoption of additional evaluation metrics with different underlying principles. Our proposal consists of using high-quality spontaneous speech corpora as a source for extracting production-related variables, which the models are then fine-tuned to predict. We hypothesize that these production-related features offer insights into the language processing mechanisms underlying the data and that cognitively sensitive models should outperform others in predicting these features. Specifically, we propose focusing on the prediction of phenomena such as speech reductions, prosodic prominences, sequences co-occurring with listeners’ backchannels, and disfluencies. To illustrate our approach, we present an example involving the prediction of speech reductions in spontaneous speech in two different languages (French and English), using models trained on 10 million tokens from different data source mixtures. Although the results are preliminary, they suggest that this task can characterize models for predicting human language processing.

We introduce The Benchmark of Linguistic Minimal Pairs (BLiMP),1 a challenge set for evaluating the linguistic knowledge of language models (LMs) on major grammatical phenomena in English. BLiMP consists of 67 individual datasets, each containing 1,000 minimal pairs—that is, pairs of minimally different sentences that contrast in grammatical acceptability and isolate specific phenomenon in syntax, morphology, or semantics. We generate the data according to linguist-crafted grammar templates, and human aggregate agreement with the labels is 96.4%. We evaluate n-gram, LSTM, and Transformer (GPT-2 and Transformer-XL) LMs by observing whether they assign a higher probability to the acceptable sentence in each minimal pair. We find that state-of-the-art models identify morphological contrasts related to agreement reliably, but they struggle with some subtle semantic and syntactic phenomena, such as negative polarity items and extraction islands.

Though state-of-the-art sentence representation models can perform tasks requiring significant knowledge of grammar, it is an open question how best to evaluate their grammatical knowledge. We explore five experimental methods inspired by prior work evaluating pretrained sentence representation models. We use a single linguistic phenomenon, negative polarity item (NPI) licensing, as a case study for our experiments. NPIs like any are grammatical only if they appear in a licensing environment like negation (Sue doesn’t have any cats vs. *Sue has any cats). This phenomenon is challenging because of the variety of NPI licensing environments that exist. We introduce an artificially generated dataset that manipulates key features of NPI licensing for the experiments. We find that BERT has significant knowledge of these features, but its success varies widely across different experimental methods. We conclude that a variety of methods is necessary to reveal all relevant aspects of a model’s grammatical knowledge in a given domain.

Tree-structured neural network architectures for sentence encoding draw inspiration from the approach to semantic composition generally seen in formal linguistics, and have shown empirical improvements over comparable sequence models by doing so. Moreover, adding multiplicative interaction terms to the composition functions in these models can yield significant further improvements. However, existing compositional approaches that adopt such a powerful composition function scale poorly, with parameter counts exploding as model dimension or vocabulary size grows. We introduce the Lifted Matrix-Space model, which uses a global transformation to map vector word embeddings to matrices, which can then be composed via an operation based on matrix-matrix multiplication. Its composition function effectively transmits a larger number of activations across layers with relatively few model parameters. We evaluate our model on the Stanford NLI corpus, the Multi-Genre NLI corpus, and the Stanford Sentiment Treebank and find that it consistently outperforms TreeLSTM (Tai et al., 2015), the previous best known composition function for tree-structured models.