Understanding the Semantic Structure of Noun Phrase Queries

Xiao LiACL 2010 · OpenReview

Abstract

Determining the semantic intent of web queries not only involves identifying their semantic class, which is a primary focus of previous works, but also understanding their semantic structure. In this work, we formally define the semantic structure of noun phrase queries as comprised of intent heads and intent modifiers. We present methods that automatically identify these constituents as well as their semantic roles based on Markov and semi-Markov conditional random fields. We show that the use of semantic features and syntactic features significantly contribute to improving the understanding performance.

Introduction

Web search queries often take the form of "telegraphic" noun phrases—keyword strings that lack formal grammar but carry dense semantic information. For instance, a query like "Johnny Depp movies 2010" is essentially a shorthand request for a structured database lookup. While traditional search engines excel at "intent classification" (e.g., identifying this as a movie-related query), providing direct, structured answers requires a deeper understanding of the query's internal structure. This paper proposes a framework to decompose such queries into their constituent semantic roles.

Semantic Framework: Intent Heads and Modifiers

The researchers define the semantic structure of a query through two primary roles:

  1. Intent Head (IH): The specific attribute or information the user is seeking (e.g., "cast," "director," "salary").
  2. Intent Modifier (IM): The attribute values that serve as constraints or filters for the search (e.g., "Alice in Wonderland," "2010").
Examples of Intent Heads (IH) and Intent Modifiers (IM) across Movie, Job, and National Park domains.

Methodology

To automate the identification of these roles, the study compares two discriminative probabilistic models: Linear-chain Conditional Random Fields (CRFs) and Semi-Markov CRFs. While standard CRFs label individual words, Semi-Markov CRFs label word segments. This is crucial for web queries where semantic units are often multi-word phrases (e.g., "The Godfather").

The effectiveness of these models relies on three types of features:

  • Lexical Features: Basic word identities.
  • Semantic Features: Leveraging external lexicons (databases of titles, actors, etc.) with fuzzy matching to handle variations.
  • Syntactic Features: Utilizing Part-of-Speech (POS) tags and chunking to identify structural cues, such as the likely position of an Intent Head.

Experimental Results

The models were evaluated across three domains: Movie, Job, and National Park. The results demonstrate that Semi-Markov CRFs consistently outperform standard CRFs by better capturing the phrase-level nature of query constituents.

Performance comparison of Semi-Markov CRF models across different feature combinations and domains.

As shown in the results, the inclusion of semantic and syntactic features provides a substantial performance boost. In the Movie domain, the transition from basic lexical features to a full feature set (B1-B6) increased sentence accuracy significantly. Across all domains, the Semi-Markov CRF achieved an average accuracy of 76.1% and a segment F1-score of 86.7%, proving that a learned statistical model significantly outperforms manual linguistic rules.

Conclusion

This research highlights that understanding the "telegraphic" language of search requires moving beyond simple classification to structural decomposition. By modeling queries as sequences of semantic segments and integrating rich lexical, semantic, and syntactic features, search engines can more accurately map keyword strings to structured data. Future work will focus on expanding lexicon coverage and handling more complex, non-noun phrase query types.

Summary updated

Abstract

Determining the semantic intent of web queries not only involves identifying their semantic class, which is a primary focus of previous works, but also understanding their semantic structure. In this work, we formally define the semantic structure of noun phrase queries as comprised of intent heads and intent modifiers. We present methods that automatically identify these constituents as well as their semantic roles based on Markov and semi-Markov conditional random fields. We show that the use of semantic features and syntactic features significantly contribute to improving the understanding performance.
BibTeX
@inproceedings{li2010understanding,
  author = {Xiao Li},
  title = {{Understanding the Semantic Structure of Noun Phrase Queries}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2010},
  url = {https://openreview.net/forum?id=ByVMtog_-r},
}
Figure