Multiple Feature-Sets Method for Dependency Parsing
0
Citation
32
Reference
10
Related Paper
Abstract:
This paper presents a simple and effective approach to improve dependency parsing by exploiting multiple feature-sets. Traditionally, features are extracted by applying the feature templates to all the word pairs(first-order features)and word tuples(second-order features). In this pa per, we show that exploiting different feature templates for different word pairs and word tuples achieves significant improvement overbaseline parsers. First, we train a text chunker using a freely available implementation of the first-order linear conditional random fields model. Then we build a clause-chunk tree for a given sentence based on chunking information and punctuation marks. Finally, we extract features for dependency parsing according to multiple feature-sets. We extend the projective parsing algorithms of McDonald[20] and Carreras[1] for our case, experimental results show that our approach significantly outperform the baseline systems without increasing complexity. Given correct chunking information, we improve from baseline accuracies of 91.36% and 92.20% to 93.19% and 93.89%, respectively.Keywords:
Dependency grammar
Chunking (psychology)
Feature (linguistics)
Word order
Punctuation
S-attributed grammar
A new Chinese chunking algorithm is proposed based on conditional random fields and semantic features. Through the analysis of Chinese chunking task and its sequential characteristics, conditional random fields that combine various kinds of features were applied. Semantic features were utilized to further improve the chunking performance. Experimental results on the Chinese chunking corpus of Microsoft Research Asia show that the algorithm achieves impressive accuracy of 92.52% in terms of the F-score.
Chunking (psychology)
Cite
Citations (0)
Treebank
Chunking (psychology)
Cite
Citations (5)
Dependency grammar
Sequence labeling
Named Entity Recognition
Feature (linguistics)
Cite
Citations (0)
In this paper, we propose a variety of Long Short-Term Memory (LSTM) based models for sequence tagging. These models include LSTM networks, bidirectional LSTM (BI-LSTM) networks, LSTM with a Conditional Random Field (CRF) layer (LSTM-CRF) and bidirectional LSTM with a CRF layer (BI-LSTM-CRF). Our work is the first to apply a bidirectional LSTM CRF (denoted as BI-LSTM-CRF) model to NLP benchmark sequence tagging data sets. We show that the BI-LSTM-CRF model can efficiently use both past and future input features thanks to a bidirectional LSTM component. It can also use sentence level tag information thanks to a CRF layer. The BI-LSTM-CRF model can produce state of the art (or close to) accuracy on POS, chunking and NER data sets. In addition, it is robust and has less dependence on word embedding as compared to previous observations.
Sequence (biology)
Cite
Citations (3,264)
To improve the accuracy of Chinese chunking and utilize the semantic information of words,a new Chinese chunking method is proposed based on conditional random fields and semantic classes.Through the analysis of Chinese chunking task and its sequential characteristics,conditional random fields that could incorporate various types of features were applied to overcome the label bias problem.Semantic features were utilized to improve the chunking performance.Experimental results show that the algorithm achieves impressive accuracy of 92.77% in terms of the F-score.A further experiment indicates the effects of feature template selection and training data′s scales on the aspect of chunking performance.
Chunking (psychology)
Semantic feature
Cite
Citations (18)
This paper proposes a distributed strategy for Chinese text chunking on the basis Conditional Random Fields(CRFs) and Error-driven technique.First eleven types of Chinese chunks are divided into different groups to build CRFs model respectively.Then,the error-driven technique is applied over CRFs chunking results for further modification.Finally,a method is described to deal with the conflicting chunking according to the F-measure values.The experimental results show that this approach is effective,outperforming the single CRFs-based approach,distributed method and other hybrid approaches in the open test by achieving reaches 94.90%,91.00%,and 92.91% in recall,precision,and F-measure respectively.
CRFS
Chunking (psychology)
Cite
Citations (1)
Chinese chunking is defined as a task to automatically segment Chinese sentences into small chunks which hold semantic meanings. To improve the performance of Chinese chunking, we propose an approach to use frequently used words (FUW) for Chinese chunking. We use conditional random fields for chunking, and modified the training corpus according to the frequency of the words in it. Finally we devise an experiment to evaluate how the number of FUW affects the performance of conditional random fields. The experiment shows that the approach can be helpful for Chinese chunking, but the count of FUW should be carefully selected for high performance and low training cost.
Chunking (psychology)
Cite
Citations (0)
We describe an algorithm for Japanese analysis that does both base phrase chunking and dependency parsing simultaneously in linear-time with a single scan of a sentence. In this paper, we show a pseudo code of the algorithm and evaluate its performance empirically on the Kyoto University Corpus. Experimental results show that the proposed algorithm with the voted perceptron yields reasonably good accuracy.
Chunking (psychology)
Dependency grammar
Phrase
Base (topology)
Cite
Citations (4)
CRFS
Chunking (psychology)
Agglutinative language
Sequence labeling
Cite
Citations (7)
Dependency parsing has gained more and more interest in natural language processing in recent years due to its simplicity and general applicability for diverse languages. The international conference of computational natural language learning (CoNLL) has organized shared tasks on multilingual dependency parsing successively from 2006 to 2009, which leads to extensive progress on dependency parsing in both theoretical and practical perspectives. Meanwhile, dependency parsing has been successfully applied to machine translation, question answering, text mining, etc. To date, research on dependency parsing mainly focuses on data-driven supervised approaches and results show that the supervised models can achieve reasonable performance on in-domain texts for a variety of languages when manually labeled data is provided. However, relatively less effort is devoted to parsing out-domain texts and resource-poor languages, and few successful techniques are bought up for such scenario. This tutorial will cover all these research topics of dependency parsing and is composed of four major parts. Especially, we will survey the present progress of semi-supervised dependency parsing, web data parsing, and multilingual text parsing, and show some directions for future work. In the first part, we will introduce the fundamentals and supervised approaches for dependency parsing. The fundamentals include examples of dependency trees, annotated treebanks, evaluation metrics, and comparisons with other syntactic formulations like constituent parsing. Then we will introduce a few mainstream supervised approaches, i.e., transition-based, graph-based, easy-first, constituent-based dependency parsing. These approaches study dependency parsing from different perspectives, and achieve comparable and state-of-the-art performance for a wide range of languages. Then we will move to the hybrid models that combine the advantages of the above approaches. We will also introduce recent work on efficient parsing techniques, joint lexical analysis and dependency parsing, multiple treebank exploitation, etc. In the second part, we will survey the work on semi-supervised dependency parsing techniques. Such work aims to explore unlabeled data so that the parser can achieve higher performance. This tutorial will present several successful techniques that utilize information from different levels: whole tree level, partial tree level, and lexical level. We will discuss the advantages and limitations of these existing techniques. In the third part, we will survey the work on dependency parsing techniques for domain adaptation and web data. To advance research on out-domain parsing, researchers have organized two shared tasks, i.e., the CoNLL 2007 shared task and the shared task of syntactic analysis of non-canonical languages (SANCL 2012). Both two shared tasks attracted many participants. These participants tried different techniques to adapt the parser trained on WSJ texts to out-domain texts with the help of large-scale unlabeled data. Especially, we will present a brief survey on text normalization, which is proven to be very useful for parsing web data. In the fourth part, we will introduce the recent work on exploiting multilingual texts for dependency parsing, which falls into two lines of research. The first line is to improve supervised dependency parser with multilingual texts. The intuition behind is that ambiguities in the target language may be unambiguous in the source language. The other line is multilingual transfer learning which aims to project the syntactic knowledge from the source language to the target language.
Dependency grammar
S-attributed grammar
Top-down parsing language
Dependency graph
Cite
Citations (0)