Taiqi He

Thousands of the world’s languages are in danger of extinction—a tremendous threat to cultural identities and human language diversity. Interlinear Glossed Text (IGT) is a form of linguistic annotation that can support documentation and resource creation for these languages’ communities. IGT typically consists of (1) transcriptions, (2) morphological segmentation, (3) glosses, and (4) free translations to a majority language. We propose Wav2Gloss: a task in which these four annotation components are extracted automatically from speech, and introduce the first dataset to this end, Fieldwork: a corpus of speech with all these annotations, derived from the work of field linguists, covering 37 languages, with standard formatting, and train/dev/test splits. We provide various baselines to lay the groundwork for future research on IGT generation from speech, such as end-to-end versus cascaded, monolingual versus multilingual, and single-task versus multi-task approaches.

Interlinear glossing provides a vital type of morphosyntactic annotation, both for linguists and language revitalists, and numerous conventions exist for representing it formally and computationally. Some of these formats are human readable; others are machine readable. Some are easy to edit with general-purpose tools. Few represent non-concatentative processes like infixation, reduplication, mutation, truncation, and tonal overwriting in a consistent and formally rigorous way (on par with affixation). We propose an annotation convention—Generalized Glossing Guidelines (GGG) that combines all of these positive properties using an Item-and-Process (IP) framework. We describe the format, demonstrate its linguistic adequacy, and compare it with two other interlinear glossed text annotation schemes.

Natural language processing models based on machine learning (ML-NLP models) have been developed to solve practical problems, such as interpreting an Internet search query. These models are not intended to reflect human language comprehension mechanisms, and the word representations used by ML-NLP models and human brains might therefore be quite different. However, because ML-NLP models are trained with the same kinds of inputs that humans must process, and they must solve many of the same computational problems as the human brain, ML-NLP models and human brains may end up with similar word representations. To distinguish between these hypotheses, we used representational similarity analysis to compare the representational geometry of word representations in two ML-NLP models with the representational geometry of the human brain, as indexed with event-related potentials (ERPs). Participants listened to stories while the electroencephalogram was recorded. We extracted averaged ERPs for each of the 100 words that occurred most frequently in the stories, and we calculated the similarity of the neural response for each pair of words. We compared this 100 × 100 similarity matrix to the 100 × 100 similarity matrix for the word pairs according to two ML-NLP models. We found significant representational similarity between the neural data and each ML-NLP model, beginning within 250 ms of word onset. These results indicate that ML-NLP systems that are designed to solve practical technology problems have a representational geometry that is correlated with that of the human brain, presumably because both are influenced by the structural properties and statistics of language.

Cross-lingual language tasks typically require a substantial amount of annotated data or parallel translation data. We explore whether language representations that capture relationships among languages can be learned and subsequently leveraged in cross-lingual tasks without the use of parallel data. We generate dense embeddings for 29 languages using a denoising autoencoder, and evaluate the embeddings using the World Atlas of Language Structures (WALS) and two extrinsic tasks in a zero-shot setting: cross-lingual dependency parsing and cross-lingual natural language inference.