Gazpacho and summer rash: lexical relationships from temporal patterns of web search queries

Enrique Alfonseca, Massimiliano Ciaramita, Keith B. HallEMNLP 2009 · OpenReview

Abstract

In this paper, we investigate the temporal patterns of web search queries to reveal semantic and pragmatic relationships between words. By analyzing five years of search logs, we demonstrate that temporal profiles provide a unique dimension of meaning that complements traditional distributional similarity. Our evaluations show that time-series similarity significantly improves query suggestion performance, achieving state-of-the-art results. However, temporal data alone proves to be a weak indicator for broad topical categorization.

Introduction: The Temporal Dimension of Meaning

Traditional Natural Language Processing (NLP) typically relies on "distributional similarity"—the idea that words appearing in similar textual contexts share similar meanings. While powerful, this approach often misses real-world context. For instance, the terms "gazpacho" and "summer rash" rarely appear in the same sentence, yet they are deeply linked by human behavior: both spike in popularity during the summer months.

This paper explores the "wisdom of the crowds" by treating the timing of search queries as a "temporal heartbeat." We hypothesize that these patterns can automatically discover lexical relationships that text-based models might overlook.

Methodology: Extracting Temporal Signatures

To test this, we mapped a vocabulary of nouns from WordNet 3.0 to five years of anonymized Google search logs (2004–2009). The process involved four key steps:

  1. Temporal Signatures: We generated a time series for each phrase, normalized to account for internet growth and weekly cycles.
  2. Correlation Analysis: We used the Pearson correlation coefficient to identify pairs with highly similar "heartbeats" (correlation > 0.9).
  3. Graph Clustering: Related terms were grouped using graph-cut algorithms.
  4. Fourier Analysis: Discrete Fourier Transforms (DFT) were applied to smooth noise and isolate periodic signals, such as yearly seasonal trends.
A typical temporal signature extracted from Google search logs, showing periodic spikes over a five-year period.

Discovering Lexical Relationships

Our analysis revealed that temporal similarity captures both semantic (linguistic) and pragmatic (behavioral) relationships. For example, "Aids" and "Welt" (German for "World") correlate because of "World Aids Day," a link that is purely event-driven rather than linguistic.

As shown in the table below, the temporal method identifies a wide range of relationship types, from direct synonyms to structurally unrelated pairs that are linked by real-world events.

Breakdown of lexical relationship types identified through high temporal correlation.

Results and Applications

Query Suggestion

The most successful application of this data was in query suggestion. Using only temporal time-series similarity, we achieved a Precision@1 of 0.74, outperforming the standard distributional similarity baseline of 0.72. Most importantly, these signals are complementary. When combined, the system reached a Precision@1 of 0.79, suggesting that "when" we search provides information that "what" we write does not.

Query Categorization

Conversely, temporal data struggled with broad categorization. Sorting queries into high-level buckets like "Shopping" or "Computers" yielded only 50% accuracy. This suggests that while temporal signals are excellent for identifying specific entity relationships (e.g., an actor and their movies, which showed a 0.88 correlation), they are too specific to define broad topical categories.

Conclusion and Future Work

Temporal patterns in search queries offer a rich, underutilized source of semantic information. While limited by the need for high-volume, multi-year data, the potential for improving search features is clear. Future research will look into sub-day intervals to further refine these relationships and apply temporal context to solve complex problems like entity disambiguation—distinguishing between "Jaguar" the car and "Jaguar" the animal based on the timing of the search.

Summary updated

Abstract

In this paper we investigate temporal patterns of web search queries. We carry out several evaluations to analyze the properties of temporal profiles of queries, revealing promising semantic and pragmatic relationships between words. We focus on two applications: query suggestion and query categorization. The former shows a potential for time-series similarity measures to identify specific semantic relatedness between words, which results in state-of-the-art performance in query suggestion while providing complementary information to more traditional distributional similarity measures. The query categorization evaluation suggests that the temporal profile alone is not a strong indicator of broad topical categories.
BibTeX
@inproceedings{alfonseca2009gazpacho,
  author = {Enrique Alfonseca and Massimiliano Ciaramita and Keith B. Hall},
  title = {{Gazpacho and summer rash: lexical relationships from temporal patterns of web search queries}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2009},
  url = {https://openreview.net/forum?id=HJVEuzfOZH},
}
Figure