Conference Paper

Parser Combination by Reparsing.

DOI: 10.3115/1614049.1614082 Conference: Human Language Technology Conference of the North American Chapter of the Association of Computational Linguistics, Proceedings, June 4-9, 2006, New York, New York, USA
Source: DBLP

ABSTRACT We present a novel parser combination scheme that works by reparsing input sen- tences once they have already been parsed by several different parsers. We apply this idea to dependency and constituent parsing, generating results that surpass state-of-the- art accuracy levels for individual parsers.

0 Followers
 · 
84 Views
  • Source
    [Show abstract] [Hide abstract]
    ABSTRACT: Consider the scenario where votes from multiple experts utilizing different data modalities or mod-eling assumptions are available for a given prediction task. The task of combining these signals with the goal of obtaining a better prediction is ubiquitous in Information Retrieval (IR), Natural Lan-guage Processing (NLP) and many other areas. In IR, for instance, meta-search aims to combine the outputs of multiple search engines to produce a better ranking. In NLP, aggregation of the outputs of computer systems generating natural language translations [7], syntactic dependency parses [8], identifying intended meanings of words [1], and others has received considerable recent attention. Most existing learning approaches to aggregation address the supervised setting. However, for com-plex prediction tasks such as these, data annotation is a very labor intensive and time consuming process. In this line of work, we first derive a mathematical and algorithmic framework for learning to com-bine predictions from multiple signals without supervision. In particular, we use the extended Mal-lows formalism (e.g. [5, 4]) for modeling aggregation, and derive an unsupervised learning proce-dure for estimating the model parameters [2]. While direct application of the learning framework can be computationally expensive in general, we propose alternatives to keep learning and infer-ence tractable. The intuition behind our approach is that the agreement between signals can serve to estimate their relative quality, which can in turn be used to induce aggregation. Indeed, higher quality signals are better at generating labels close (defined in terms of a distance function) to cor-rect prediction and thus will tend to agree with one another, whereas the poor ones will not. The key assumption we make is that predictions induced by signals are conditionally independent given the true prediction. We demonstrate the effectiveness of our framework on the tasks of aggregating permutations and aggregating top-k lists. In many practical applications, the relative quality of the constituent signals is unlikely to remain the same across different domains. Consider, for example, the meta-search task we mentioned earlier. The relative quality of the search engines is likely to depend on the type of the query issued: one may specialize on ranking product reviews while others on ranking scientific documents. Therefore, we extend our aggregation formalism to explicitly model such latent variability in the quality of the constituent signals [3]. We again instantiate the extended framework on aggregating permutations and top-k lists and experimentally demonstrate (Figure 1, left) that it is capable of learning a better aggregation than our type agnostic model if the variability is indeed present in the data.
  • [Show abstract] [Hide abstract]
    ABSTRACT: Dependency parsing has made many advancements in recent years, in particular for English. There are a few dependency parsers that achieve comparable accuracy scores with each other but with very different types of errors. This paper examines creating a new dependency structure through ensemble learning using a hybrid of the outputs of various parsers. We combine all tree outputs into a weighted edge graph, using 4 weighting mechanisms. The weighted edge graph is the input into our ensemble system and is a hybrid of very different parsing techniques (constituent parsers, transition-based dependency parsers, and a graph-based parser). From this graph we take a maximum spanning tree. We examine the new dependency structure in terms of accuracy and errors on individual part-of-speech values. The results indicate that using a greater number of more varied parsers will improve accuracy results. The combined ensemble system, using 5 parsers based on 3 different parsing techniques, achieves an accuracy score of 92.58%, beating all single parsers on the Wall Street Journal section 23 test set. Additionally, the ensemble system reduces the average relative error on selected POS tags by 9.82%.
    Proceedings of the Workshop on Innovative Hybrid Approaches to the Processing of Textual Data; 04/2012
  • Source

Preview

Download
1 Download
Available from