Skip to content
All experiencesTemplates
Personal siteDr. Rita Silva

A bibliography you can narrow

Research interests, six publications with facets derived from the entries themselves, talks, teaching, and identifiers that appear only where the author supplied one.

Built from
ProfileHeroPublicationListSocialLinksTimelineBadgeCard
What the browser test drives
Filter the bibliography by year, then by type, then by topic, and watch the live count follow; clear a filter and the full list returns.
Appearance
Graphite with the editorial serif at instrument geometry — a monochrome page where the bibliography is the only colour.
Preview

Instituto Atlântico, Lisbon

Dr. Rita Silva

Computational linguistics, Instituto Atlântico

I study how language models lose information as documents get longer, and what can be measured about that loss without asking the model how it feels.
Publications
6
Citations
233
First paper
2022
Illustrated portrait of Dr. Rita Silva

What I am working on

  • Long-context attentionSparse and hierarchical attention, and the cost of the approximation.
  • Evaluation without a benchmarkMeasurements that survive the next model, and the ones that do not.
  • Morphology and low-resource languagesWhat a tokeniser does to a language it was not trained on.

Publications

Newest first, in the order the venues print them. Filter by year, type or topic; the counts are the entries themselves, not a number I maintain.

6 of 6

  1. Sparse attention degrades at the seventh hop

    Journal article

    R. Silva, M. Okonkwo

    Transactions on Language Technology202624 citations

    DOI 10.0000/fictional.tlt.2026.14

  2. A tokeniser audit for agglutinative languages

    Conference paper

    R. Silva, A. Tan

    Workshop on Low-Resource Language Technology202511 citations

    Best paper, runner-up

    DOI 10.0000/fictional.wlrt.2025.3

  3. Holding the evaluation fixed while the model changes

    Preprint

    A. Tan, R. Silva

    Preprint2025

    Read

  4. Retrieval is not a substitute for context

    Conference paper

    R. Silva

    Conference on Computational Semantics202447 citations

    DOI 10.0000/fictional.ccs.2024.88

  5. What a tokeniser costs a language

    Journal article

    R. Silva, J. Pereira, A. Tan

    Journal of Computational Morphology202363 citations

    DOI 10.0000/fictional.jcm.2023.6

  6. Measuring long-context recall without a needle

    Conference paper

    R. Silva, M. Okonkwo

    Symposium on Evaluation Methods202288 citations

    DOI 10.0000/fictional.sem.2022.9

Talks

  1. The seventh hop2026

    Keynote, Language Technology ConferencePresented the degradation result and the replication attempt that failed.

  2. Auditing a tokeniser in a week2025

    Workshop on Low-Resource Language TechnologyHands-on session with the audit scripts, released under an open licence.

  3. Why retrieval did not replace context2024

    Institute seminar seriesA shorter version of the 2024 paper, aimed at a non-specialist audience.

Teaching

  • Measurement in natural language processingGraduateFifteen weeks on building an evaluation that outlives the model it was written for.
  • Statistical methods for linguistsMastersA course written for people who can read a paper but have never fitted a model.
Elsewhere

Every citation on this page is fictional: the identifiers are shaped like real ones so the layout can be judged, and they resolve nowhere.

Interactive, not live: every action below changes local state in this page. No network request is made, no market data is fetched and no message is sent anywhere.

Luc UI v1.0Academic profiledark theme