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: Instance-based Transfer Learning for Multilingual Deep Retrieval

Abstract: Perhaps the simplest type of multilingual transfer learning is instance-based transfer learning, in which data from the target language and the auxiliary languages are pooled, and a single model is learned from the pooled data. It is not immediately obvious when instance-based transfer learning will improve performance in this multilingual setting: for instance, a plausible conjecture is this kind of transfer learning would help only if the auxiliary languages were very similar to the target. Here we show that at large scale, this method is surprisingly effective, leading to positive transfer on all of 35 target languages we tested. We analyze this improvement and argue that the most natural explanation, namely direct vocabulary overlap between languages, only partially explains the performance gains: in fact, we demonstrate target-language improvement can occur after adding data from an auxiliary language with no vocabulary in common with the target. This surprising result is due to the effect of transitive vocabulary overlaps between pairs of auxiliary and target languages.
Subjects: Computation and Language (cs.CL); Information Retrieval (cs.IR); Machine Learning (cs.LG); Machine Learning (stat.ML)
Cite as: arXiv:1911.06111 [cs.CL]
  (or arXiv:1911.06111v1 [cs.CL] for this version)

Submission history

From: Andrew Arnold [view email]
[v1] Fri, 8 Nov 2019 22:23:30 GMT (312kb,D)

Link back to: arXiv, form interface, contact.