The ROBOT TALK corpus was created with the aim of serving as a language sample to perform the quantitative and qualitative contrastive linguistic analyses in order to answer the main question of the project: is it possible to distinguish whether a text in Spanish has been generated by an LLM or by a person using linguistic features of the text?

This is a comparable monitor corpus in Spanish. It is composed of author-comparable texts (human, Bard, Claude, GPT-3.5-Turbo, GPT-4, Mixtral) of three main genres (news, film reviews and scientific articles specialised in linguistics).

 

  Sample of the corpus

 

Characteristics of the corpus

• Text written in Spanish
• Comparables by author
  • humano
  • Bard
  • Claude
  • GPT-3.5-Turbo
  • GPT-4
  • Mixtral
• Sources:
  • Scientific journals in linguistics
  • Online news
  • Film review websites
• Genres
  • Scientific articles
  • News
  • Film reviews

Sources of the corpus

• Scientific articles on linguistics
  • RSEL, Revista de investiación Lingüística, Revista electrónica de lingüística aplicada, Sintagma, Círculo de Lingüística Aplicada a la Comunicación, Asterisco, …
• News
  • RTVE, EFE
• Film review
  • Filmaffinity

 

Description of the corpus

Comparable corpus	Author	Human	Bard	Claude	GPT-3.5-Turbo	GPT-4	Mixtral	No. of texts by genre
Genre ot the text	Scientific articles	144	90	0	90	95	90	509
	News	171	171	60	111	171	111	795
	Film reviews	160	160	65	95	160	95	735
Total no. of text		475	421	125	296	426	296	2039