Retrieval of Temporal Event Sequences from Textual Descriptions

TL;DR

This work tackles temporal event sequence retrieval from textual descriptions by introducing TESRBench, a diverse benchmark of real-world sequences and synthesized descriptions, and TPP-Embedding, a model that embeds both sequences and descriptions into a shared temporal-aware embedding space using temporal encodings and LLMs within a TPP framework. A contrastive learning objective ties description and sequence embeddings, enhanced by 4-bit quantization and LoRA for efficiency. Across TESRBench and multi-domain settings, TPP-Embedding consistently outperforms traditional text-based baselines in MRR and Recall@K, demonstrating robust retrieval across temporally structured, multi-type event data. The work advances practical TESR by supplying a standardized benchmark and a scalable, domain-adaptive retrieval model with strong temporal-semantic alignment for real-world applications.

Abstract

Retrieving temporal event sequences from textual descriptions is crucial for applications such as analyzing e-commerce behavior, monitoring social media activities, and tracking criminal incidents. To advance this task, we introduce TESRBench, a comprehensive benchmark for temporal event sequence retrieval (TESR) from textual descriptions. TESRBench includes diverse real-world datasets with synthesized and reviewed textual descriptions, providing a strong foundation for evaluating retrieval performance and addressing challenges in this domain. Building on this benchmark, we propose TPP-Embedding, a novel model for embedding and retrieving event sequences. The model leverages the TPP-LLM framework, integrating large language models (LLMs) with temporal point processes (TPPs) to encode both event texts and times. By pooling representations and applying a contrastive loss, it unifies temporal dynamics and event semantics in a shared embedding space, aligning sequence-level embeddings of event sequences and their descriptions. TPP-Embedding demonstrates superior performance over baseline models across TESRBench datasets, establishing it as a powerful solution for the temporal event sequence retrieval task.

Figures (5)

• Figure 1: TPP-Embedding framework with one TESRBench example, where the model embeds both textual descriptions and temporal event sequences using a shared TPP-LLM framework, applies pooling to generate fixed-length representations, and uses contrastive learning with similarity scores to align matching pairs for effective event sequence retrieval.
• Figure 2: Instructions for generating objective summaries of event sequences, focusing on the order, timing, and general trends without including specific numbers or timestamps.
• Figure 3: TPP-Embedding architecture, illustrating the embedding process for a event sequence through the integration of temporal and text representations, followed by processing with a large language model and a pooling layer to generate a fixed-length sequence representation.
• Figure 4: Comparison of average MRRs with standard deviations on TESRBench in event sequence retrieval.
• Figure 5: Comparison of average Recall@5 with standard deviations on TESRBench in event sequence retrieval.