This paper was accepted within the Business Monitor at SIGIR 2024.
Digital Assistants (VAs) are necessary Info Retrieval platforms that assist customers accomplish varied duties by means of spoken instructions. The speech recognition system (speech-to-text) makes use of question priors, skilled solely on textual content, to tell apart between phonetically complicated alternate options. Therefore, the technology of artificial queries which can be much like current VA utilization can drastically enhance upon the VA’s abilities-especially for use-cases that don’t (but) happen in paired audio/textual content information.
On this paper, we offer a preliminary exploration of using Massive Language Fashions (LLMs) to generate artificial queries which can be complementary to template-based strategies. We examine whether or not the strategies (a) generate queries which can be much like consumer queries from a preferred VA, and (b) whether or not the generated queries are particular. We discover that LLMs generate extra verbose queries, in comparison with template-based strategies, and reference features particular to the entity. The generated queries are much like VA consumer queries, and are particular sufficient to retrieve the related entity. We conclude that queries generated by LLMs and templates are complementary.