Back to Basics for Monolingual Alignment: Exploiting Word Similarity and Contextual Evidence

Back to Basics for Monolingual Alignment: Exploiting Word Similarity and Contextual Evidence Md Arafat Sultan†, Steven Bethard‡ and Tamara Sumner† †Institute of Cognitive Science and Department of Computer Science University of Colorado Boulder ‡Department of Computer and Information Sciences University of Alabama at Birmingham arafat.sultan@colorado.edu, bethard@cis.uab.edu, sumner@colorado.edu Abstract We present a simple, easy-to-replicate monolin- gual aligner that demonstrates state-of-the-art performance while relying on almost no su- pervision and a very small number of external resources. Based on the hypothesis that words with similar meanings represent potential pairs for alignment if located in similar contexts, we propose a system that operates by finding such pairs. In two intrinsic evaluations on alignment test data, our system achieves F1 scores of 88– 92%, demonstrating 1–3% absolute improve- ment over the previous best system. Moreover, in two extrinsic evaluations our aligner out- performs existing aligners, and even a naive application of the aligner approaches state-of- the-art performance in each extrinsic task. 1 Introduction Monolingual alignment is the task of discovering and aligning similar semantic units in a pair of sentences expressed in a natural language. Such alignments pro- vide valuable information regarding how and to what extent the two sentences are related. Consequently, alignment is a central component of a number of important tasks involving text comparison: textual entailment recognition, textual similarity identifica- tion, paraphrase detection, question answering and text summarization, to name a few. The high utility of monolingual alignment has spawned significant research on the topic in the re- cent past. Major efforts that have treated alignment as a standalone problem (MacCartney et al., 2008; Thadani and McKeown, 2011; Yao et al., 2013a) are primarily supervised, thanks to the manually aligned corpus with training and test sets from Microsoft Re- search (Brockett, 2007). Primary concerns of such work include both quality and speed, due to the fact that alignment is frequently a component of larger NLP tasks. Driven by similar motivations, we seek to devise a lightweight, easy-to-construct aligner that produces high-quality output and is applicable to various end tasks. Amid a variety of problem formulations and ingenious approaches to alignment, we take a step back and examine closely the effectiveness of two frequently made assumptions: 1) Related semantic units in two sentences must be similar or related in their meaning, and 2) Commonalities in their se- mantic contexts in the respective sentences provide additional evidence of their relatedness (MacCartney et al., 2008; Thadani and McKeown, 2011; Yao et al., 2013a; Yao et al., 2013b). Alignment, based solely on these two assumptions, reduces to finding the best combination of pairs of similar semantic units in sim- ilar contexts. Exploiting existing resources to identify similarity of semantic units, we search for robust techniques to identify contextual commonalities. Dependency trees are a commonly used structure for this purpose. While they remain a central part of our aligner, we expand the horizons of dependency-based alignment beyond exact matching by systematically exploiting the notion of “type equivalence” with a small hand- crafted set of equivalent dependency …

Content truncated to 500 words