[Paper Review] Large-Scale Noun Compound Interpretation Using Bootstrapping and the Web as a Corpus

Su Nam Kim, Preslav Nakov|arXiv (Cornell University)|Nov 27, 2019

Topic Modeling40 references33 citations

TL;DR

The paper presents a bootstrapping framework that jointly harvests noun compounds (NCs) and fine-grained paraphrasing patterns from the Web to interpret NCs for a target abstract relation (Make 2). It analyzes how fixing either the head or modifier during bootstrapping improves accuracy and scalability.

ABSTRACT

Responding to the need for semantic lexical resources in natural language processing applications, we examine methods to acquire noun compounds (NCs), e.g., "orange juice", together with suitable fine-grained semantic interpretations, e.g., "squeezed from", which are directly usable as paraphrases. We employ bootstrapping and web statistics, and utilize the relationship between NCs and paraphrasing patterns to jointly extract NCs and such patterns in multiple alternating iterations. In evaluation, we found that having one compound noun fixed yields both a higher number of semantically interpreted NCs and improved accuracy due to stronger semantic restrictions.

Motivation & Objective

• Motivate the need for large-scale, interpretable noun compounds (NCs) and their fine-grained paraphrase patterns.
• Develop a bootstrapping method that alternates between extracting NCs from patterns and extracting patterns from NCs, using the Web as a corpus.
• Assess how constraining one noun in the NC (head or modifier) affects extraction quality and coverage.
• Provide a dataset of NCs annotated with both coarse-grained (Make 2) and fine-grained paraphrasing distributions.

Proposed method

• Represent NC semantics as a distribution over paraphrasing verbs and prepositions.
• Use a two-step bootstrapping process: (i) extract NCs that paraphrase given seed patterns; (ii) extract new paraphrasing patterns for the discovered NCs.
• Employ three bootstrapping strategies (loose, strict, NC-only strict) with Web-based queries to harvest NCs and patterns.
• Filter candidates using frequency thresholds and WordNet-based noun checks; disallow seeds or previously extracted NCs; require minimum co-occurrence counts in Web data.
• Iterate up to three times, balancing coverage and semantic accuracy, and generate fine-grained paraphrase distributions for NCs.

Experimental results

Research questions

• RQ1 Can bootstrapping from a small seed set of patterns and NCs scale to hundreds of thousands of NCs interpreted with paraphrase verbs?
• RQ2 Does fixing one noun (head or modifier) in the NC improve extraction accuracy and reduce semantic drift?
• RQ3 What is the trade-off between extraction quantity and accuracy across loose, strict, and NC-only strict bootstrapping regimes?
• RQ4 How informative are fine-grained paraphrase verbs in capturing NC semantics compared to coarse abstract relations?
• RQ5 How does the resulting NC dataset support NLP applications needing NC interpretation?

Key findings

• Strict and NC-only strict bootstrapping yield higher NC and NC-pattern accuracy than loose bootstrapping.
• The number of extracted NCs grows substantially under strict bootstrapping, while NC-only strict bootstrapping accelerates growth due to more pattern instantiations.
• Using a higher co-occurrence threshold (N=10) generally reduces extracted NCs and patterns but can slightly improve accuracy.
• Bootstrapped NCs can be paired with a fine-grained distribution over paraphrasing verbs for each NC.
• Comparative evaluation shows their method achieves higher coverage at the cost of some degradation in accuracy across iterations.