A Weighted Polynomial Information Gain Kernel for Resolving Prepositional Phrase Attachment Ambiguities with Support Vector Machines
Proceedings of the Eighteenth International Joint Conference on Artificial Intelligence. Pages 133-139.
© 2003 International Joint Conferences on Artificial Intelligence
Abstract
We introduce a new kernel for Support Vector Machine learning in a natural language setting. As a case study to incorporate domain knowledge into a kernel, we consider the problem of resolving Prepositional Phrase attachment ambiguities. The new kernel is derived from a distance function that proved to be succesful in memory-based learning. We start with the Simple Overlap Metric from which we derive a Simple Overlap Kernel and extend it with Information Gain Weighting. Finally, we combine it with a polynomial kernel to increase the dimensionality of the feature space. The closure properties of kernels guarantee that the result is again a kernel. This kernel achieves high classification accuracy and is efficient in both time and space usage. We compare our results with those obtained by memory-based and other learning methods. They make clear that the proposed kernel achieves a higher classification accuracy.
- [Aha eta/., 1991] D. Aha, D. Kibler, and M. Albert. Instance based learning algorithms. Machine Learning, 6:37-66, 1991.
- [Brown eta/., 1999] Michael P.S. Brown, William Noble Grundy, David Lin, Nello Cristianini, Charles Sugnet, Manuel Ares, and David Haussler. Support vector machine classification of microarray gene expression data, 1999. [Chih-Chung and Chi-Jen, 2002] Chang Chih-Chung and Lin Chi-Jen. Libsvm: a library for support vector ma- chines, 2002.
- [Collins and Brooks, 1995] M.J Collins and J. Brooks. Prepositional phrase attachment through a backed-off model. In Proceedings of the Third Workshop on Very Large Corpora, Cambridge, 1995.
- [Cost and Salzberg, 1993] Scott Cost and Steven Salzberg. A weighted nearest neighbour algorithm for learning with symbolic features . Machine Learning, 10:57-78, 1993.
- [Cristianini and Shawe-Taylor, 2000] Nello Cristianini and John Shawe-Taylor. An Introduction to Support Vec- tor Machines and other Kernel-based Learning Methods. Cambridge University Press, 2000.
- [Daelemans eta/., 2002] Walter Daelemans, Jakub Zavrel, Ko van der Sloot, and Antal van den Bosch. Timbl: Tilburg memory-based learner, version 4.3. Technical re- port, Til burg University and University of Antwerp, 2002.
- [Marcus eta/., 1993] M. Marcus, Santorini B, and M.A. Marcinkiewicz. Building a large annotated corpus of en- glish: The penn treebank. Computational Linguistics, 19(2):313-330, 1993.
- [Mitchell, 1997] Tom Mitchell. Machine Learning. The McGraw-Hill Companies, Inc., 1997.
- [Ramon, 2002] Jan Ramon. Clustering and instance based learning in first order logic. PhD thesis, K.U. Leuven, Belgium, 2002. [Ratnaparkhi eta/., 1994]
- Adwait Ratnaparkhi, J. Reynar, and S. Roukos. A maximum entropy model for prepo- sitional phrase attachment. In Proceedings of the ARPA Workshop on Human Language Technology, Plainsboro, NJ, 1994.
- [Ratnaparkhi, 1998] Adwait Ratnaparkhi. Maximum En- tropy Models for Natural Language Ambiguity Resolution. PhD thesis, University of Pennsylvania, Philadelphia, PA, 1998.
- [Sang, 2002] Erik F. Tjong Kim Sang. Introduction to the conll-2002 shared task: Language-independent named en- tity recognition. In Proceedings of CoNLL-2002, Taipei, Taiwan, pages 155-158, 2002.
- [Stetina and Nagao, 1997] Jiri Stetina and Makoto Nagao. Corpus based pp attachment ambiguity resolution with a semantic dictionary., 1997. Kyoto University, Yoshida Honmachi, Kyoto 606, Japan.
- [Vapnik, 1998] Vladimir Vapnik. The Nature of Statistical Learning Theory. Springer-Verlag, New York, 1998.
- [Veenstra eta/., 2000] Jorn Veenstra, Antal van den Bosch, Sabine Buchholz, Walter Daelemans, and Jakub Zavrel. Memory-based word senese disambiguation. Computers and the Humanities, 34: 1-2: 171 -177, 2000.
- [Zavrel eta/., 1997] Jakub Zavrel, Walter Daelemans, and Jorn Veenstra. Resolving pp attachment ambiguities with memory-based learning. In Proceedings CoNNL, Madrid, pages 136-144. Computational Linguistics, Tilburg Uni- versity, 1997.
References were extracted automatically from the PDF and may contain errors.
Full text extracted automatically from the PDF; figures, tables and equations may differ. The PDF is the version of record.
1 Introduction
An important issue in natural language analysis is the resolu tion of structural ambiguity. A sentence is said to be struc turally ambiguous when it can be assigned to more than one syntactic structure [Zavrel et al., 1997]. In Prepositional Phrase (PP) attachment one wants to disambiguate between cases where it is uncertain whether the PP attaches to the verb or to the noun.
Example 1 Consider the following two sentences:
- I bought the shirt with pockets.
- I washed the shirt with soap.
In sentence I, with modifies the noun shirt because with pockets ( PP) describes the shirt. In sentence 2 however; with modifies the verb washed because with soap (PP) describes how the shirt is washed [Ratnaparkhi, 1998].
This type of attachment ambiguity is easy for people to resolve because they can use their world knowledge [Stetina and Nagao, 1997]. A computer program usually cannot rely on that kind of knowledge.
· Author funded by a doctoral grant of the institute for advance ment of scientific technological research in Flanders (IWT).
This problem has already been tackled using memory based learning like for example k-nearest neighbours. Here, the training examples are first stored in memory and classi fication of a new example is done based on the closest ex ample stored in memory. Therefore, one needs a function that expresses the distance or similarity between examples. There already exist several dedicated distance functions to solve all kind of natural language problems using memory based learning [Veenstra et al., 2000; Zavrel et al., 1997; Daelemans et al., 2002] .
We will use a Support Vector Machine (SVM) to tackle the problem of PP attachment disambiguation. Central to SVM learning is the kernel function K : X x X -+ ~ where X contains the examples and the kernel K calculates an inner product in a second space, the feature space F . This product expresses how similar examples are.
Our goal is to combine the power of SVMs with the dis tance functions that are well-suited for the probem for which they were designed. Deriving a distance from a kernel is straightforward, see Section 2.1. However, deriving a kernel from a distance is not trivial since kernels must satisfy some extra conditions, i.e. being a kernel is a much stronger con dition than being a distance. In this paper we will describe a method that shows how such dedicated distance functions can be used as a basis for designing kernels that sequentially can be used in SVM learning.
We use the PP attachment problem as a case study to illus trate our approach. As a starting point we take the Overlap Metric that has been succesfully used in memory-based learn ing for the same problem [Zavrel et al., 1997].
Section 2 will give a short overview of the theory of SVMs together with some theorems and definitions that are needed in Section 4. Based on [Zavrel et al., 1997], section 3 gives an overview of metrics developed for memory-based learning applied to the PP attachment problem. In Section 4 the new kernels will be introduced. Finally Sections 5 and 6 give some experimental results and a conclusion of this work.
2 Support Vector Machines
For simplicity, in our explanation we will consider the case of binary classification only, i.e. we consider an input space X with input vectors x and a target space D = { 1, -1}. The goal of the SVM is to assign every x to one of two classes D = { 1, -1}. The decision boundary that separates the in put vectors belonging to different classes is usually an arbi trary n- !-dimensional manifold if the input space X is n dimensional.
One of the basic ideas behind SVMs is to have a mapping <ll from the original input space X into a high-dimensional feature space F that is a Hilbert space, i.e. a complete vector space provided with an inner product. Separation of the trans formed feature vectors <ll(x) in F is done linearly, i.e. by a hyperplane. Cover's theorem states that any consistent train ing set can be made linear separable provided the dimension ofF is high enough.
However, transforming the training set T S {x1,x2, ... ,x;, ... ,XN} C X into such a higher-dimensional space incurs both computational and learning-theoretic problems. The high dimensionality of F makes it very expensive both in terms of memory and time to represent the feature vectors <ll(x;) corresponding to the training vectors x;. Moreover, separating the data in this way exposes the learning system to the risk of overfitting the data if the separating hyperplane is not chosen properly.
SVMs sidestep both difficulties [Vapnik, 1998]. First, overfitting is avoided by choosing the unique maximum mar gin hyperplane among all possible hyperplanes that can sep arate the data in F. This hyperplane maximizes the distance to the closest data points.
Second, the maximum margin hyperplane in F can be rep resented entirely in terms of training vectors x; E T S and a kernelK.
Definition: A kernel is a function K : X x X -t 1R so that forallxandyinX,K(x,y) =< <ll(x),<ll(y) >where<llisa (non-linear) mapping from the input space X into the Hilbert space F provided with the inner product< ., . > [Cristianini and Shawe-Taylor, 2000].
To be more precise, once we have chosen a kernel K we can represent the maximal margin hyperplane (or decision boundary) by a linear equation in x
where the training vectors x; and their class labels d; are given. The decision boundary is determined completely once we know the a; and b. They are obtained by optimizing a convex quadratic objective function with linear constraints. Moreover, most of the a; prove to be zero. By definition the vectors x; corresponding with non-zero a; are called the sup port vectors SV and this set consists of those data points that lie closest to the hyperplane and thus are the most difficult to classify.
In order to classify a new point Xnew. one determines the sign of
If this sign is positive Xnew belongs to class 1, if negative to class -1, if zero Xnew lies on the decision boundary. Note that we have now restricted the summation to the set SV of support vectors because the other a; are zero anyway.
To conclude, SVM can sidestep the above two difficulties because neither the feature space F nor the map <ll from the input space X into F are explicitly defined, they are replaced by the kernel K that operates on vectors of the input space X.
2.1 Some Properties of Kernels
The selection of an appropriate kernel K is the most impor tant design decision in SVMs since it implicitly defines the feature space F and the map <11. A SVM will work correctly even if we don't know the exact form of the features that are used in F. Moreover, the kernel expresses prior knowledge about the patterns being modelled, encoded as a similarity measure between two vectors [Brown et al., 1999].
But not all maps over X x X are kernels. Since a kernel K is related to an inner product, cfr. the definition above, it has to satisfy some conditions that arise naturally from the defini tion of an inner product and are given by Mercer's theorem: the map must be continuous and positive definite [Vapnik, 1998].
In this paper we will use following methods to construct kernels [Cristianini and Shawe-Taylor, 2000]:
Ml Making kernels from kernels: Based on the fact that ker nels satisfy a number of closure properties. In this case, the Mercer conditions follow naturally from the closure properties of kernels.
M2 Making kernels from features: Start from the features of the input vectors and obtain a kernel by working out their inner product. A feature is a component of the in put vector. In this case, the Mercer conditions follow naturally from the definition of an inner product.
The set of kernels is closed under certain operations, this allows us to make new kernels from existing ones (see M1). We mention only the closure property used in the paper:
Closure properties of kernels: Let <ll : X -t 1Rn be a map and K1 be a kernel over 1Rn x 1Rn then the following function is a kernel K(x, y) = K1(<ll(x), <ll(y)).
Another interesting property is that we can derive a dis tance d from a kernel K as follows:
Distance derived from a kernel: Let K be a kernel over X x X, then d defined as d(x,y) = ,jK(x,x)- 2K(x,y) + K(y,y) is a distance on X and consequentially (X, d) is a metric space.
3 Metrics for Memory-based Learning
In this section, we will focus on the distance functions [Zavrel et al., 1997; Cost and Salzberg, 1993] used for memory-based learning with symbolic values. First, we will have a look at the Simple Overlap Metric (SOM) and next we will discuss Information Gain Weighting (IGW). Memory-based learning is a class of machine learning techniques where training in stances are stored in memory first and classification of new instances later on is based on the distance (or similarity) be tween the new instance and the closest training instances that have already been stored in memory. A well-known example of memory-based learning is k-nearest neighbours classifica tion. We will not go into further detail, the literature offers many books and articles that provide a comprehensive intro duction to memory-based learning, e.g. [Mitchell, 1997]. For our purpose, the important thing to remember is that we will be working with symbolic values like strings over an alphabet of characters. Instances are n-dimensional vectors a and b in some universe nn. The components of the vectors in nn are strings, i.e. a; , b; E :E*, with :E* the set of all strings over some alphabet :E. Notice that for n = 1, nn becomes n and the elements of n are strings in :E*. In order for the inner products defined below to be meaningful, the set of strings n has to be finite. As a result the set of vectors nn is also finite as required.
3.1 Simple Overlap Metric
The most basic metric for vectors with symbolic values is the Simple Overlap Metric(SOM) [Zavrel et al., 1997]:
with
and
Here dsoM(a, b) is the distance between the vectors a and b, represented by n features and t5 is the distance per feature. The k-nearest neighbour algorithm equipped with this metric is called IB1 [Aha et al., 1991]. The IB1 algorithm simply counts the number of (mis)matching feature values in both vectors. This is a reasonable choice if we have no information about the importance of the different features. But if we do have information about feature importance then we can add linguistic bias to weight the different features.
3.2 Information Gain Weighting
Information Gain Weighting (IGW) measures for every fea ture i separately how much information it contributes to our knowledge of the correct class label. The Information Gain (I G) of a feature i is measured by calculating the entropy be tween the cases with and without knowledge of the value of that feature:
Where C is the set of class labels, Vi is the set of values for feature i, H(C) = - L:cEC P(c)log2P(c) is the entropy of the class labels and si(i) is the split info. The resulting weights w; can be used to extend Equation 2 with weights (see Equation 6). The k-nearest neigbour algorithm equipped with this metric is called IB 1-IG [Zavrel et al., 1997] and the corresponding distance called diG:
4 A Kernel for Natural Language settings
In this section, we will describe the kernel that we have con structed from the IGW metric d1a from the previous sec tion. A kernel will be a function K : nn X nn -t ~ and the mapping to the feature space F will be of the form <I> : nn -t ~nn. We will begin with the simple, one dimensional, unweighted case (Section 4.1 ). Next, we will extend it to n dimensions (Section 4.2) and add IGW (Sec tion 4.3) and finally in Section 4.4 the Polynomial Informa tion Gain (PIG) kernel will be introduced.
4.1 A Simple Overlap Kernel (SOK) in one dimension
If we only want to consider vectors in a one-dimensional space then we have to take a closer look at Equations 3 and 4. To derive a kernel from this distance function we will use M2 from Section 2.1 . Consider the inner product space ~n of functions f : n -t ~, the inner product of two functions f and g is defined as:
Next, let ¢soK : n -t ~n with
be a feature mapping to the feature space F . The function t5a is defined as t5a(x) = 1 when x = a and 0 otherwise (see Equation 4). Starting from the features and the above de fined feature mapping we will now work out the inner prod uct to come to the corresponding kernel ksoK : f1 X f1 -t ~:
ksoK(a, b) = < ¢soK(a), ¢soK(b) >
We don'thave to proof that the kernel ksoK satisfies the Mer cer conditions because this follows naturally from the defini tion of the inner product. We started from the features, de fined a mapping ¢soK on the features and we worked out the inner product between them, see M2 from Section 2.1 . How ever, to show that the kernel really corresponds to the distance function given in Equations 3 and 4 we have to verify the dis tance formula for kernels given in Section 2.1:
The last step is justified by the fact that 1 - t5a (b) is either 0 or 1 [Ramon, 2002].
4.2 Extending the SOK ton Dimensions
In this section we consider vectors a, b E nn, the distance between these vectors is measured by the distance dsoM from Equation 2. Now consider the inner product space ~on (the space of functions f : nn --+ ~) comparable to the one in Equation 7, with the following inner product:
We begin by defining a mapping <I? SOK : nn --+ ~on:
We will again use M2 from Section 2.1 to derive a ker nel for dsoM· We work out the inner product between the feature vectors <l?soK(a) and <l?soK(b) to come to the corresponding kernel KsoK : nn X nn --+ ~: KsoK(a, b) = < <l?soK(a), <l?soK(b) >
Here the function 8a is defined as 8a(x) = L~=l 8a, (x;) with 8a, (x;) like in Equations 4 and 8. Notice that this corresponds to the inner product between the vectors a (¢soK(a1), ... , ¢soK(an)) and b (¢soK(b1), ... , ¢soK(bn)). So, alternatively we can write KsoK: nn X nn--+ ~:
We don't have to proof that the kernel KsoK is a valid kernel as this follows naturally from the definition of the inner product. However, we again have to show that it is a kernel that corresponds to the distance function dsoM from Equation 2. In the following we will show that the kernel KsoK in fact corresponds to the square root of dso M, we will call this distance function dr:
However, this does not impose any problems for the kernel we are aiming to develop.
4.3 Adding Information Gain to the SOK
Next, we will introduce IGW into the kernel. We will con struct a kernel based on the distance function d1a from Equa tion 6. We have to define new mappings ¢and <I? to cope with the weights:
We will start again by working out the inner prod uct for the )-dimensional case k1a : f1 X f1 --+ ~: kw(a;, b;) = < ¢w(ai), ¢w(b;) >
The function kw is guaranteed to be a valid kernel, this follows naturally from the definition of the inner product (see M2 from Section 2.1). In the same way as for K soK we can show that K1a corresponds to the inner product between the vectors a (¢w(al), ... , ¢w(an)) and b = (¢w(b1), ... , ¢w(bn)). yielding the following result for the kernel Kw : nn X nn --+ ~:
Also, in the same way as we did before we can show for k1a that it corresponds to a distance function 810 : f1 X f1 --+ ~ with 8w(a;, bi) = w;8(a;, b;) with 8 as defined in Equa tion 3. Furthermore, in the same way as we did for KsoK in Section 4.2 we can show that the kernel Kw corresponds to the square root of the distance function dw from Equation 6.
4.4 A Polynomial Information Gain Kernel
In this section, we will increase the dimensionality of the fea ture space F by making use of a polynomial kernel K paly· We will start this section by giving an example [Cristianini and Shawe-Taylor, 2000]:
Example 2 (Polynomial kernel): mensional input space X, with Consider a two di vectors x, y E X :
which is equivalent to an inner product between feature vec tors of the form
We call the kernel Kpaly a polynomial kernel because it maps the vectors from the input space X to the feature space F of all monomials of degree 2.
Next, we will derive the Polynomial Information Gain (PIG) kernel KPJa, making use of M1 from Section 2.1. We will do this based on the closure property mentioned in Section 2.1. Consider the feature mapping <I? IG from Equa tion 11, the kernel KPia will be defined by a combination of this feature mapping and the polynomial kernel Kpaly from the above example.
as defined in Equation II and
From the closure properties of kernels, it follows naturally that Kpw indeed is a valid kernel which calculates the in ner product of two vectors transformed by a feature mapping if>pw: nn-+ ~om.
The mapping <I> PIG maps vectors a E nn to the space 1R 0 m with m > n. The features of the vectors in ~om are combinations (inner products) of the following form: < ¢>w(ai), <Pw(ai) > and 2 < ¢>w(ai), ¢>w(ai) >. In this way we aim to capture more relevant information that might be present in such combinations. Notice that the inner product here above is the one defined in Equation 7.
5 Experimental Results
In the following section, we will give some experimental re sults we obtained with the kernels Kw and KPJc. First we will describe the PP attachment disambiguation problem in more detail, followed by the experimental set up (Sec tion 5.2). Finally, we will present the results and compare them to other methods that have been applied on the same data set (Section 5.3).
5.1 PP attachment disambiguation problem
If it is uncertain in a sentence whether the preposition attaches to the verb or to the noun then we have a prepositional phrase (PP) attachment problem. For example, in sentence I of Ex ample 1, with modifies the noun shirt because with pockets (PP) describes the shirt. In contrast, in sentence 2, with mod ifies the verb washed because with soap (PP) describes how the shirt is washed [Ratnaparkhi, 1998].
Example 3 For the purpose of PP attachment disambigua tion, sentences 1 and 2 from Example 1 will be reduced to vectors a and b E 0 4 as follows:
1. a = (bought, shirt, with, pockets)
2. b =(washed, shirt, with, soap)
In fact, we only keep those words that are of any importance to the PP attachment problem, i.e. (V( erb ),N( oun),P( reposition),N( oun)).
In the case where sentences are reduced to quadruples as illustrated in Example 3, the human performance is approx imately 88.2% [Ratnaparkhi et al., 1994]. This performance rate gives us an acceptable upper limit for the maximum performance of a computer because it seems unreasonable to expect an algorithm to perform much better than a hu man. As we will show in our experimental results the kernel Kw achieves a classification accuracy up to 82,9%, see Sec tion 5.3. However, in [Zavrel et al., 1997] the IB1-IG attains a maximum classification accuracy of 84.1 %, this is a good indication of the classification accuracy that should be possi ble to obtain with a kernel based on the distance defined in Equation 6.
5.2 Experimental Setup
The experiments have been done with LIBSVM, a C/C++ and Java library for SVMs [Chih-Chung and Chi-Jen, 2002]. The machine we have used is a Pentium III with 256MB RAM memory, running Windows XP. We choose to implement the kernels Kw and KPic in Java.
The type of SVM learning we have used is C-SVM [Chih Chung and Chi-Jen, 2002]. The parameter C controls the model complexity versus the model accuracy and has to be chosen based on the complexity of the problem.
The experiments in this section are conducted on a simpli fied version of the full PP attachment problem (see Example 3). The data consists of four-tuples of words, extracted from the Wall Street Journal Treebank [Marcus et al., 1993] by a group at IBM [Ratnaparkhi et al., 1994]1.
The data set contains 20801 training patterns, 3097 test patterns and an independant validation set of 4093 patterns for parameter optimization. All of the models described be low were trained on all training examples and the results are given for the 3097 test patterns. For the benchmark compari son with other models from the literature, we use only results for which all parameters have been optimized on the valida tion set. For more details concerning the data set we refer to [Zavrel e t at., 1997].
5.3 Results
The table below gives the results obtained with the kernels K1c and KPIG and compares them to other methods taken from the literature·
| Method | Percent correct |
|---|---|
| Kw | 82.9% |
| KPIG | 84.8% |
| IB1 | 83.7% |
| IBl-IG | 84.1% |
| C4.5 | 79.7% |
| Maximum Entropy | 77.7% |
| Transformations | 81.9% |
| Back-off model | 84.1% |
| Late Closure | 59.0% |
| Most Likely for each P | 72.0% |
The results for Kw and KPic were obtamed after opti mization of the parameter C on the validation set. For Kw this optimal value was C = 0.8 and for KPic this optimal value was C = 0.55. The scores ofiBl, IBl-10, C4.5, Late Closure and Most Likely for each P are taken from [Zavrel et al., 1997], the scores from Maximum Entropy are taken from [Ratnaparkhi et al., 1994] and the scores of Transfor mations are taken from [Collins and Brooks, 1995].
The fact that the kernel Kw performs worse than IB1-IG, although it is equipped with the very same distance metric, may seem somehow surprising. We believe this is because SVMs perform linear separation in the feature space F. The decision boundary of IB 1-IG on the other hand is non-linear. Due to the linearity of the decision boundary of the SVM, some points get misclassified. The number of misclassifica tions is controlled by the parameter C. Choosing a larger value for C forces the SVM to avoid classification errors. However, choosing a value for C that is too high will result in overfitting. Increasing the dimensionality ofF makes lin ear separation easier (Cover's theorem). This can be seen by comparing the classification accuracies between the kernels KIG and KpJG. The results also show that our kernel KPia outperforms all other methods in the comparison.
1 This data set is freely available by ftp from ftp://ftp.cis.upenn.edu/pub/adwaiUPPattachData/.
6 Conclusion
In this paper, we have proposed a method for designing ker nels for SVM learning in natural language settings. As a start ing point we took distance functions that have been used in memory-based learning. The resulting kernel KPia achieves high classification accuracy compared to other methods that have been applied on the same data set. In our approach we started from the vectors of the input space, defined a map ping on those vectors and worked out their inner product in the feature space. All necesary kernel conditions follow nat urally from the definition of the inner product. We used the distance formula for kernels to show that the kernels are ac tually based on the distance functions from Section 3. The experimental results of the kernel K pIG show that increasing the dimensionality in the feature space yields better results. We increased the dimensionality of the feature space by mak ing combinations of the features <I> IG through a polynomial kernel. In this way, we do not only take into account the sim ilarity between corresponding features of vectors, but we also take into account the similarity between non-corresponding features of the vectors. In fact this comes down to taking into account, to a greater extend, the context in wich a word oc curs.
We used the resolvement of PP attachment ambiguities as a case-study to validate our findings, but it is our belief that the kernel K pIG can be used for a wide range of natural language problems. In the future we will apply our kernels in more complex natural language settings and compare the results to other methods that have been applied on the same problems. At the moment we are already running experiments on lan guage independent named-entity recognition [Sang, 2002]. The first results look promising, but it is too early to draw any significant conclusions. We will also try to further ex tend the kernels proposed in this paper to achieve even higher accuracies. Moreover, there are still many distance functions reported in the literature that have not yet been investigated to see whether they are applicable in SVM learning, we intend to investigate such distance functions and if possible derive a kernel from them.
@inproceedings{ijcai2003p19,
title = {A Weighted Polynomial Information Gain Kernel for Resolving Prepositional Phrase Attachment Ambiguities with Support Vector Machines},
author = {Vanschoenwinkel, Bram and Manderick, Bernard},
booktitle = {Proceedings of the Eighteenth International Joint Conference on
Artificial Intelligence, {IJCAI-03}},
publisher = {International Joint Conferences on Artificial Intelligence},
editor = {Georg Gottlob and Toby Walsh},
pages = {133--139},
year = {2003},
url = {https://ijcai.org/proceedings/2003/19},
}
TY - CPAPER TI - A Weighted Polynomial Information Gain Kernel for Resolving Prepositional Phrase Attachment Ambiguities with Support Vector Machines AU - Vanschoenwinkel, Bram AU - Manderick, Bernard T2 - Proceedings of the Eighteenth International Joint Conference on Artificial Intelligence, IJCAI-03 PB - International Joint Conferences on Artificial Intelligence PY - 2003 SP - 133 EP - 139 UR - https://ijcai.org/proceedings/2003/19 ER -Copied
Machine-readable versions of this paper, for tools and AI agents: