Antonis Krasakis
@tonylittlewine.bsky.social
PhD student at University of Amsterdam. Working on information retrieval, LLMs and RAG
To do so, we introduce a novel activation function (SNReLU) that preserves the desirable properties of splade models (log saturation effect, sparsity). We find that, given enough training data (needed to adapt to the new act fn), retrieval performance for negations improves by +100%!
January 15, 2025 at 11:57 AM
To do so, we introduce a novel activation function (SNReLU) that preserves the desirable properties of splade models (log saturation effect, sparsity). We find that, given enough training data (needed to adapt to the new act fn), retrieval performance for negations improves by +100%!
First, we construct zero-shot representations
using vector arethmetics designed to encapsulate each compositionality type (eg. intersection or negations). This method often outperforms DR/LSR models trained for compositionality, which struggle for some of these compositionality types
using vector arethmetics designed to encapsulate each compositionality type (eg. intersection or negations). This method often outperforms DR/LSR models trained for compositionality, which struggle for some of these compositionality types
January 15, 2025 at 11:57 AM
First, we construct zero-shot representations
using vector arethmetics designed to encapsulate each compositionality type (eg. intersection or negations). This method often outperforms DR/LSR models trained for compositionality, which struggle for some of these compositionality types
using vector arethmetics designed to encapsulate each compositionality type (eg. intersection or negations). This method often outperforms DR/LSR models trained for compositionality, which struggle for some of these compositionality types