A Conversational Movie Search System Based on Conditional Random Fields

Jingjing Liu, Scott Cyphers, Panupong Pasupat, Ian McGraw, James R. GlassInterspeech 2012 · DOI

Abstract

In the early 2010s, movie search was largely limited to basic keyword matching, often failing to interpret complex natural language queries. This paper introduces MovieBrowser, a system that leverages Semi-Markov Conditional Random Fields (Semi-CRFs) and Latent Dirichlet Allocation (LDA) to provide a conversational search experience. By parsing queries into semantic slots and expanding them through topic modeling and phonetic matching, MovieBrowser achieves significant improvements in search relevance, typically placing the desired movie in the top two results.

Introduction

While streaming services revolutionized content delivery, search mechanisms remained rudimentary. A query like "a funny movie about bridesmaids starring Kera Nightley" often yielded irrelevant results because systems could not distinguish between genres, plots, and actors. MovieBrowser addresses these challenges by bridging the gap between natural human speech and structured database queries using statistical modeling and semantic interpretation.

System Architecture

The MovieBrowser pipeline integrates several components to move from raw speech to refined search results. The architecture involves a dialogue manager, a search engine, and a feedback loop incorporating crowdsourced data from Amazon Mechanical Turk (AMT) for model training and annotation.

Overall system architecture of MovieBrowser, including the crowdsourcing and search components.

Methodology

The core of the system relies on three primary techniques to handle the nuances of natural language:

  1. Semi-Markov CRFs: Unlike standard CRFs that label word-by-word, Semi-CRFs label segments of text. This allows for joint segmentation and classification, which is crucial for identifying multi-word entities like movie titles (e.g., "West Side Story") as single semantic units.
  2. Topic Modeling (LDA): To address data sparsity in movie databases, Latent Dirichlet Allocation was used to cluster words from 2 million plots and 8,000 reviews. This enables query expansion, mapping a user's intent (e.g., "funny") to related database tags (e.g., "comedy").
  3. Phonetic Search: Using the Lucene search engine and the Metaphone algorithm, the system converts words into phonetic codes. This provides robustness against misspellings or errors in Automatic Speech Recognition (ASR), such as "Kera" vs. "Keira."
Semantic parsing and query expansion pipeline for a natural language movie query.

Experimental Results

The system was evaluated using two datasets: a synthetic-natural dataset (4,384 sentences) and a free-form natural language dataset (5,200 sentences) collected via crowdsourcing.

  • Parsing Accuracy: The Semi-CRF model achieved an F1-score of approximately 66% on free-form queries, rising to nearly 80% when lexical features (dictionaries of actors and titles) were included for structured queries.
  • Search Relevance: In a study of 1,000 search events, the average rank of the first relevant movie was 0.65 (where 0 is the top result). This indicates that the system almost always places the correct movie in the #1 or #2 spot.
  • Model Comparison: Semi-CRFs consistently outperformed standard CRFs across both datasets, proving that segment-level modeling is superior for natural language search queries.

Conclusion

MovieBrowser demonstrates the efficacy of combining segment-level statistical modeling with semantic generalization for conversational search. By moving beyond literal keyword matching, the system provides a more intuitive interface for discovering content. Future work will focus on scaling these models to larger databases and integrating ASR uncertainty directly into the probabilistic parsing process.

Summary updated

BibTeX
@inproceedings{liu2012conversational,
  author = {Jingjing Liu and Scott Cyphers and Panupong Pasupat and Ian McGraw and James R. Glass},
  title = {{A Conversational Movie Search System Based on Conditional Random Fields}},
  booktitle = {Interspeech},
  year = {2012},
  doi = {10.21437/interspeech.2012-563},
}
Figure