Search for: All records

Instruction tuning has remarkably advanced large language models (LLMs) in understand- ing and responding to diverse human instruc- tions. Despite the success in high-resource lan- guages, its application in lower-resource ones faces challenges due to the imbalanced foun- dational abilities of LLMs across different lan- guages, stemming from the uneven language distribution in their pre-training data. To tackle this issue, we propose pivot language guided generation (PLUG), an approach that utilizes a high-resource language, primarily English, as the pivot to enhance instruction tuning in lower-resource languages. It trains the model to first process instructions in the pivot language, and then produce responses in the target lan- guage. To evaluate our approach, we introduce a benchmark, X-AlpacaEval, of instructions in 4 languages (Chinese, Korean, Italian, and Spanish), each annotated by professional trans- lators. Our approach demonstrates a significant improvement in the instruction-following abili- ties of LLMs by 29% on average, compared to directly responding in the target language alone. Further experiments validate the versatility of our approach by employing alternative pivot languages beyond English to assist languages where LLMs exhibit lower proficiency. 

Taxonomies serve many applications with a structural representation of knowledge. To incorporate emerging concepts into existing taxonomies, the task of taxonomy completion aims to find suitable positions for emerging query concepts. Previous work captured homogeneous token-level interactions inside a concatenation of the query concept term and definition using pre-trained language mod- els. However, they ignored the token-level interactions between the term and definition of the query concepts and their related concepts. In this work, we propose to capture heterogeneous token-level interactions between the different textual components of concepts that have different types of relations. We design a relation-aware mutual attention module (RAMA) to learn such interactions for taxonomy completion. Experimental results demonstrate that our new taxonomy completion framework based on RAMA achieves the state-of-the-art performance on six taxonomy datasets.