Page 41 - 2024S

34
Figure 1: Proposed System Diagram.
for precise odor mixing and release. Figure 2
shows the olfactory display device used in our
experiments.
Figure 2: Olfactory display device used in the
experiment. This 13-component system allows
for precise odor mixing and release through
rapid solenoid valve switching.
This olfactory display’s user-friendly interface the sender, system, user, or assistant. The sys-
allows us to specify odor mixtures directly,
streamlining the process of synchronizing ol-
factory cues with visual and auditory content.
Its compact size and ability to rapidly switch
between multiple odors make it ideal for our
research on integrating olfactory experiences
with text-based odor detection.
UEC Int’l Mini-Conference No.52
The device’s precision and flexibility enable
the creation of complex, time-synchronized odor
profiles, enhancing the immersive quality of mul-
timedia experiences while maintaining practical-
ity for various research applications.
3.2 Text Analysis by Language Mod-
els
Using large language models (LLMs) such as
gpt-3.5 or gpt-4 provided by OpenAI [8], our
system analyzes subtitles to detect contextual
and semantic cues that suggest specific odors.
This involves understanding nuanced expres-
sions and indirect references to odors, which are
challenging but crucial for accurate odor pre-
diction. To steer LLMs produce desired result,
we apply modern prompt engineering technique
that will be introduced in the next section, and
test in different models like: Claude, Gemini.
3.3 Prompt Engineering
Prompt engineering is employed to refine
the input to language models, ensuring that
the output is highly relevant and accurate.
Techniques such as the Chain-of-Thought
prompting have proven effective in enhancing
the specificity and relevance of model responses
for our application [9]. In addition, we ask the
model to first write out the reason and then
give its prediction, this way it can deal with
more complicate case and produce reasonable
result.
For instance, Figure 3 shows how we use Ope-
nAI Chat API with our crafted prompt to get
the desired odor prediction. From the documen-
tation, the API request body consists of a series
of message objects that represent the chat his-
tory. Each message object includes the role of
tem message defines model’s behavior and con-
text for the conversation, while user prompts
carry the input messages from the user. The as-
sistant role messages are the model’s responses
to the user prompts, formulated based on the
given context and conversation history. We in-
serted 1 set of conversation to the message body