We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations

DBLP - CS Bibliography


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Computation and Language

Title: Demystifying Neural Language Models' Insensitivity to Word-Order

Abstract: Recent research analyzing the sensitivity of natural language understanding models to word-order perturbations have shown that the state-of-the-art models in several language tasks may have a unique way to understand the text that could seldom be explained with conventional syntax and semantics. In this paper, we investigate the insensitivity of natural language models to word-order by quantifying perturbations and analysing their effect on neural models' performance on language understanding tasks in GLUE benchmark. Towards that end, we propose two metrics - the Direct Neighbour Displacement (DND) and the Index Displacement Count (IDC) - that score the local and global ordering of tokens in the perturbed texts and observe that perturbation functions found in prior literature affect only the global ordering while the local ordering remains relatively unperturbed. We propose perturbations at the granularity of sub-words and characters to study the correlation between DND, IDC and the performance of neural language models on natural language tasks. We find that neural language models - pretrained and non-pretrained Transformers, LSTMs, and Convolutional architectures - require local ordering more so than the global ordering of tokens. The proposed metrics and the suite of perturbations allow a systematic way to study the (in)sensitivity of neural language understanding models to varying degree of perturbations.
Comments: 11 pages, 13 figure + appendix
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
Cite as: arXiv:2107.13955 [cs.CL]
  (or arXiv:2107.13955v1 [cs.CL] for this version)

Submission history

From: Louis Clouatre [view email]
[v1] Thu, 29 Jul 2021 13:34:20 GMT (6040kb,D)

Link back to: arXiv, form interface, contact.