Bag of What? Simple Noun Phrase Extraction for Text Analysis

Abram Handler, Matthew Denny, Hanna M. Wallach, Brendan O'ConnorNLP+CSS @ EMNLP 2016 · DOI

Abstract

Traditional "bag-of-words" models often fail to capture the nuanced context provided by multi-word phrases. While n-grams and full syntactic parsers offer alternatives, they suffer from high noise and computational overhead, respectively. This paper presents NPFST (Noun Phrase Finite State Transducer), a method that leverages part-of-speech tagging and finite state transducers to extract high-quality noun phrases. NPFST achieves the high recall of n-gram methods with the low yield of complex parsers, making it an efficient tool for large-scale exploratory text analysis.

Introduction

In social science and text analysis, documents are frequently represented as a "bag-of-words." While computationally efficient, this approach loses critical context. For instance, the unigrams "health" and "care" do not convey the same specific meaning as the phrase "health care."

Existing solutions to capture this context have significant drawbacks. N-gram extraction captures relevant phrases but generates a massive volume of "garbage" sequences (e.g., "the internet website of"). Conversely, full syntactic parsing is accurate but computationally expensive and often fails on noisy text like social media or specialized technical jargon. NPFST aims to provide a "Goldilocks" solution: a method that is faster than a parser, cleaner than n-grams, and requires no manual configuration.

The NPFST Method

The NPFST pipeline consists of three primary stages:

  1. POS Tagging: Text is processed with a standard Part-of-Speech tagger to assign labels such as Noun, Adjective, or Verb.
  2. Grammar Definition: A "FullNP" grammar defines the structure of noun phrases, including sequences of adjectives followed by nouns and coordinated phrases.
  3. Finite State Transducer (FST): The core engine uses an FST to efficiently extract overlapping and nested spans. Unlike "greedy" matchers that only select the longest possible match, the FST can identify sub-phrases within a larger title, preserving important semantic components.
The Finite State Transducer (FST) architecture used by NPFST to identify and extract noun phrases based on POS tags.

Performance and Evaluation

The authors evaluated NPFST across four diverse domains: WNUT 2015 (tweets), BioNLP 2011 (biomedical articles), the New York Times (general news), and a new Congressional Bills Corpus.

A key metric in this evaluation is the trade-off between Recall (the percentage of relevant phrases found) and Yield (the total number of phrases extracted). For effective text analysis, a method should maximize recall while minimizing yield to avoid overwhelming researchers with irrelevant fragments.

Recall vs. Yield across three datasets (WNUT, BioNLP, NYT). NPFST (red square) consistently achieves high recall with significantly lower yield than n-gram methods (blue diamonds 1-6).

As shown in the results, NPFST is approximately 10x faster than full constituent parsing. In biomedical article tests, NPFST processed articles in ~82ms compared to ~825ms for the parser. Across all datasets, NPFST maintained recall typically above 90%, comparable to extracting all n-grams, but with a drastically lower yield.

Insights from the Congressional Case Study

The utility of NPFST was further demonstrated in a case study of U.S. congressional bills (1993–2014). NPFST successfully identified highly partisan phrases that unigram analysis missed. For example, while unigrams might show a general focus on "health," NPFST revealed that Democrats specifically emphasize "mental health" and "victims of domestic violence." This ability to extract nested phrases is crucial, as sub-components of long phrases often carry distinct and important meanings.

Conclusion

NPFST offers a robust and efficient alternative for phrase extraction in text analysis. By balancing recall and yield, it provides researchers with meaningful multi-word units without the noise of n-grams or the cost of full parsing. While currently English-centric and dependent on POS tagger accuracy, the FST framework is language-agnostic and extensible to broader linguistic structures in future work.

Summary updated

BibTeX
@inproceedings{handler2016bag,
  author = {Abram Handler and Matthew Denny and Hanna M. Wallach and Brendan O'Connor},
  title = {{Bag of What? Simple Noun Phrase Extraction for Text Analysis}},
  booktitle = {NLP+CSS @ EMNLP},
  year = {2016},
  doi = {10.18653/v1/w16-5615},
}
Figure