Large language models (LLMs) and small language models (SLMs) are being adopted at remarkable speed, although their safety still remains a serious concern. With the advent of multilingual S/LLMs, the question now becomes a matter of scale: can we expand multilingual safety evaluations of these models with the same velocity at which they are deployed? To this end, we introduce RTP-LX, a human-transcreated and human-annotated corpus of toxic prompts and outputs in 28 languages. RTP-LX follows participatory design practices, and a portion of the corpus is especially designed to detect culturally-specific toxic language. We evaluate 10 S/LLMs on their ability to detect toxic content in a culturally-sensitive, multilingual scenario. We find that, although they typically score acceptably in terms of accuracy, they have low agreement with human judges when scoring holistically the toxicity of a prompt; and have difficulty discerning harm in context-dependent scenarios, particularly with subtle-yet-harmful content (e.g. microaggressions, bias). We release this dataset to contribute to further reduce harmful uses of these models and improve their safe deployment.

We describe a method for automatically generating Lexical Transfer Rules (LTRs) from word equivalences using transfer rule templates. Templates are skeletal LTRs, unspecified for words. New LTRs are created by instantiating a template with words, provided that the words belong to the appropriate lexical categories required by the template. We define two methods for creating an inventory of templates and using them to generate new LTRs. A simpler method consists of extracting a finite set of templates from a sample of hand coded LTRs and directly using them in the generation process. A further method consists of abstracting over the initial finite set of templates to define higher level templates, where bilingual equivalences are defined in terms of correspondences involving phrasal categories. Phrasal templates are then mapped onto sets of lexical templates with the aid of grammars. In this way an infinite set of lexical templates is recursively defined. New LTRs are created by parsing input words, matching a template at the phrasal level and using the corresponding lexical categories to instantiate the lexical template. The definition of an infinite set of templates enables the automatic creation of LTRs for multi-word, non-compositional word equivalences of any cardinality.