Improved Fully Unsupervised Parsing with Zoomed Learning

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 684–693, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics Improved Fully Unsupervised Parsing with Zoomed Learning Roi Reichart ICNC The Hebrew University roiri@cs.huji.ac.il Ari Rappoport Institute of computer science The Hebrew University arir@cs.huji.ac.il Abstract We introduce a novel training algorithm for unsupervised grammar induction, called Zoomed Learning. Given a training set T and a test set S, the goal of our algorithm is to identify subset pairs Ti, Si of T and S such that when the unsupervised parser is trained on a training subset Ti its results on its paired test subset Si are better than when it is trained on the entire training set T . A successful ap- plication of zoomed learning improves overall performance on the full test set S. We study our algorithm’s effect on the leading algorithm for the task of fully unsupervised parsing (Seginer, 2007) in three different En- glish domains, WSJ, BROWN and GENIA, and show that it improves the parser F-score by up to 4.47%. 1 Introduction Grammar induction is the task of learning grammati- cal structure from plain text without human supervi- sion. The task is of great importance both for the understanding of human language acquisition and since its output can be used by NLP applications, avoiding the costly and error prone creation of man- ually annotated corpora. Many recent works have addressed the task (e.g. (Klein and Manning, 2004; Seginer, 2007; Cohen and Smith, 2009; Headden et al., 2009)) and its importance has increased due to the recent availability of huge corpora. A basic challenge to this research direction is how to utilize training data in the best possible way. Klein and Manning (2004) report results for their dependency model with valence (DMV) for unsupervised dependency parsing when it is trained and tested on the same corpus (both when sentence length restriction is imposed, such as for WSJ10, and when it is not, such as for the entire WSJ). To- day’s best unsupervised dependency parsers, which are rooted in this model, train on short sentences only: both Headen et al., (2009) and Cohen and Smith (2009) train on WSJ10 even when the test set includes longer sentences. Recently, Spitkovsky et al., (2010) demonstrated that training the DMV model on sentences of up to 15 words length yields better results on the entire section 23 of WSJ (with no sentence length restric- tion) than training with the entire WSJ corpus. In contrast to these dependency models, the Seginer constituency parser achieves its best perfor- mance when trained on the entire WSJ corpus ei- ther if sentence length restriction is imposed on the test corpus or not. The sentence length restriction training protocol of (Spitkovsky et al., 2010), harms this parser. When the parser is trained with the entire WSJ corpus its F-score performance on the WSJ10, WSJ20 and the entire WSJ corpora are 76, 64.8 and 56.7 respectively. When training is done with …

Content truncated to 500 words