Repository logo
 

Linear mappings: semantic transfer from transformer models for cognate detection and coreference resolution

Date

2022

Authors

Nath, Abhijnan, author
Krishnaswamy, Nikhil, advisor
Blanchard, Nathaniel, committee member
King, Emily J., committee member

Journal Title

Journal ISSN

Volume Title

Abstract

Embeddings or vector representations of language and their properties are useful for understanding how Natural Language Processing technology works. The usefulness of embeddings, however, depends on how contextualized or information-rich such embeddings are. In this work, I apply a novel affine (linear) mapping technique first established in the field of computer vision to embeddings generated from large Transformer-based language models. In particular, I study its use in two challenging linguistic tasks: cross-lingual cognate detection and cross-document coreference resolution. Cognate detection for two Low-Resource Languages (LRL), Assamese and Bengali, is framed as a binary classification problem using semantic (embedding-based), articulatory, and phonetic features. Linear maps for this task are extrinsically evaluated on the extent of transfer of semantic information between monolingual as well as multi-lingual models including those specialized for low-resourced Indian languages. For cross-document coreference resolution, whole-document contextual representations are generated for event and entity mentions from cross- document language models like CDLM and other BERT-variants and then linearly mapped to form coreferring clusters based on their cosine similarities. I evaluate my results on gold output based on established coreference metrics like BCUB and MUC. My findings reveal that linearly transforming vectors from one model's embedding space to another carries certain semantic information with high fidelity thereby revealing the existence of a canonical embedding space and its geometric properties for language models. Interestingly, even for a much more challenging task like coreference resolution, linear maps are able to transfer semantic information between "lighter" models or less contextual models and "larger" models with near-equivalent performance or even improved results in some cases.

Description

Rights Access

Subject

coreference resolution
low-resource languages
transformer
language models
affine transformation
semantics

Citation

Associated Publications