Suurten kielimallien generoiman tekstin kontrollointi monikielisessä ympäristössä latenttiavaruuden ohjausmenetelmillä
Year: 2024
Type: dissertation
Abstract: Numerous real-world applications of large language models (LLMs) require controlling the generated text to align with specific attributes. Furthermore, as LLMs are increasingly deployed in multilingual contexts, some completely new attributes emerge that must be controlled, along with novel challenges in controlling the same attributes as in English-specific settings. One of the new attributes that emerge is ensuring that the generated text is in a specific language. In this thesis, we show that a family of techniques referred to as latent space steering methods can be used to control the language of the text generated by an LLM, using Finnish and English as the case study. More specifically, we construct steering vectors from the intermediate representations of translation pairs and use these steering vectors to generate Finnish answers to challenging English questions, managing to shift the language in all responses to Finnish with a small degradation in the quality of the answers. In addition to ensuring that the generated text is in a specific language, another unique challenge to multilingual settings emerges from constructing the steering vectors for other attributes due to the creation requiring examples of the desired behaviour, which can be difficult to obtain with less represented languages. We provide one solution to this by showing that the effectiveness of some style steering vectors generalizes across languages. To demonstrate this, we use steering vectors for two styles constructed from English prompts to control the style of Finnish text generation. We observe that while there is a drop in the effectiveness of the steering vectors for the other style, the steering vectors for the other style achieve a near-perfect generalization performance. By further exploring these latent space steering methods in multilingual settings, we contribute to the important path of making the advancements in LLMs more accessible for the entire world.
Source: Aaltodoc (Aalto University)
Author: Leino, Julius
Institution: Aalto University
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7025044678
