The main focus of Bitext recent consulting projects has been the generation of instructional prompts for AI models. These projects typically involve training models that can answer questions by extracting information from financial reports and tables and performing calculations. Given an answer, we generate a prompt that is linguistically correct, as well as the set of calculation steps to reach the answer – both the questions and the steps are validated by financial experts for accuracy and relevance.
Most of our projects have three ingredients:
- Linguistic structure: our linguists prepare the necessary linguistic data and tools
- Scientific/specialist structure: our linguists work with subject-matter experts, typically to create ontologies to map domain knowledge
- Tagging/annotation: our software generates tagged data by combining linguistic data and ontologies, and the subject matter experts validate the output
As an example, we’ve worked on projects with medical/healthcare experts, to validate the creation and annotation of linguistic resources, incorporating knowledge such as semantic equivalence between texts and paraphrasing.
We have also worked in multiple languages (German, French, Japanese…) and language variants.