# Antoniak 2018: Evaluating the Stability of Embedding-based Word Similarities

> Maria Antoniak, David Mimno, 2018. 1 records on Mysteries Department cite it. The inductive-methods survey cites it for bootstrap stability checks on…

Source: https://mysteriesdept.com/sources/SRC0D8MPZN023JK/evaluating-the-stability-of-embedding-based-word  
Publisher: Mysteries Department  
Licence: CC BY 4.0 (credit Mysteries Department and link the page)  

---

[Sources](https://mysteriesdept.com/sources)[Voynich Manuscript](https://mysteriesdept.com/mysteries/voynich-manuscript)Antoniak 2018

**0**

# Evaluating the Stability of Embedding-based Word Similarities

Maria Antoniak, David Mimno, 2018. _Transactions of the Association for Computational Linguistics_.

Shows that nearest-neighbor similarities from word embeddings shift considerably with small changes to the training corpus, more so for small corpora, and recommends averaging over bootstrap samples rather than trusting one model.

**Voynich Manuscript** Articledoi: **[10.1162/tacl_a_00008](https://doi.org/10.1162/tacl_a_00008)** Checked against the original0

## What it's used for here

The inductive-methods survey cites it for bootstrap stability checks on embeddings trained on the small Voynich corpus.

## Cited by on this site 1 record

[1 methods surveys]

Inductive methods

SRC

## Cite it

```
Antoniak, M., & Mimno, D. (2018). Evaluating the Stability of Embedding-based Word Similarities. Transactions of the Association for Computational Linguistics. https://doi.org/10.1162/tacl_a_00008
```

[Read it free at doi.org](https://doi.org/10.1162/tacl_a_00008)[PDF stored here · 14 pages](https://mysteriesdept.com/sources/SRC0D8MPZN023JK/evaluating-the-stability-of-embedding-based-word/pdf)

The PDF is the copy doi.org gives away, kept here so the link can't break. Downloads are open to beta members.

Key

**SRC** 0D8MPZN023JK

Kind

Journal article

Rights

CC-BY-4.0

Checked

Against the published article

Added by

[apollo](https://mysteriesdept.com/u/apollo)

The record

Bibliographic records are CC0. Take this one as CSL-JSON, BibTeX or RIS.

[CSL-JSON](https://mysteriesdept.com/sources/SRC0D8MPZN023JK/evaluating-the-stability-of-embedding-based-word/csl.json)