Dijkstra-WSA: a Graph-Based Approach to Word Sense Alignment
Total Page:16
File Type:pdf, Size:1020Kb
Dijkstra-WSA: A Graph-Based Approach to Word Sense Alignment Michael Matuschek ‡ and Iryna Gurevych †‡ Ubiquitous Knowledge Processing Lab (UKP-DIPF), German† Institute for Educational Research and Educational Information Schloßstr. 29, 60486 Frankfurt, Germany ‡Ubiquitous Knowledge Processing Lab (UKP-TUDA), Department of Computer Science, Technische Universitat¨ Darmstadt Hochschulstr. 10, 64289 Darmstadt, Germany http://www.ukp.tu-darmstadt.de Abstract best for all purposes, as different LSRs cover differ- ent words, senses and information types. In this paper, we present Dijkstra-WSA, a These considerations have sparked increasing re- novel graph-based algorithm for word sense search efforts in the area of word sense alignment alignment. We evaluate it on four different pairs of lexical-semantic resources with dif- (WSA). It has been shown that aligned resources ferent characteristics (WordNet-OmegaWiki, can indeed lead to better performance than using the WordNet-Wiktionary, GermaNet-Wiktionary resources individually. Examples include seman- and WordNet-Wikipedia) and show that it tic parsing using FrameNet (FN), WN, and VerbNet achieves competitive performance on 3 out (VN) (Shi and Mihalcea, 2005), word sense disam- of 4 datasets. Dijkstra-WSA outperforms the biguation using an alignment of WN and Wikipedia state of the art on every dataset if it is com- (WP) (Navigli and Ponzetto, 2012) and semantic bined with a back-off based on gloss similar- role labeling using a combination of PropBank, VN ity. We also demonstrate that Dijkstra-WSA is not only flexibly applicable to different re- and FN in the SemLink project (Palmer, 2009). sources but also highly parameterizable to op- Some of these approaches to WSA either rely heav- timize for precision or recall. ily on manual labor (e.g. Shi and Mihalcea (2005)) or on information which is only present in few resources such as the most frequent sense (MFS) 1 Introduction (Suchanek et al., 2008). This makes it difficult to Lexical-semantic resources (LSRs) are a corner- apply them to a larger set of resources. stone for many Natural Language Processing (NLP) In earlier work, we presented the large-scale re- applications such as word sense disambiguation source UBY (Gurevych et al., 2012). It contains (WSD) and information extraction. However, the nine resources in two languages which are mapped growing demand for large-scale resources in dif- to a uniform representation using the LMF standard ferent languages is hard to meet. The Princeton (Eckle-Kohler et al., 2012). They are thus struc- WordNet (WN) (Fellbaum, 1998) is widely used for turally interoperable. UBY contains pairwise sense English, but for most languages corresponding re- alignments between a subset of these resources, and sources are considerably smaller or missing. Col- this work also presented a framework for creat- laboratively constructed resources like Wiktionary ing alignments based on the similarity of glosses (WKT) and OmegaWiki (OW) provide a viable op- (Meyer and Gurevych, 2011). However, it is not tion for such cases and seem especially suitable clear to what extent this approach can be applied to for smaller languages (Matuschek et al., 2013), but resources which lack this kind of information (see there are still considerable gaps in coverage which Section 3). need to be filled. A related problem is that there In summary, aligning senses is a key requirement usually does not exist a single resource which works for semantic interoperability of LSRs to increase the 151 Transactions of the Association for Computational Linguistics, 1 (2013) 151–164. Action Editor: Patrick Pantel. Submitted 12/2012; Revised 2/2013; Published 5/2013. c 2013 Association for Computational Linguistics. Downloaded from http://www.mitpressjournals.org/doi/pdf/10.1162/tacl_a_00217 by guest on 28 September 2021 coverage and effectiveness in NLP tasks. Still, exist- two LSRs. A pair of aligned senses denote the same ing efforts are mostly focused on specific types of re- meaning. E.g., the two senses of letter “The conven- sources (most often requiring glosses) or application tional characters of the alphabet used to represent scenarios. In this paper, we propose an approach to speech” and “A symbol in an alphabet, bookstave” alleviate this and present Dijkstra-WSA, a novel, ro- (taken from WN and WKT, respectively) are clearly bust algorithm for word sense alignment which is equivalent and should be aligned. applicable to a wide variety of resource pairs and languages. For the first time, we apply a graph-based 2.2 Evaluation Resources algorithm which works on full graph representations For the evaluation of Dijkstra-WSA, we align four of both resources to word sense alignment. This en- pairs of LSRs used in previous work, namely WN- ables us to take a more abstract perspective and re- OW (Gurevych et al., 2012), WN-WKT (Meyer and duce the problem of identifying equivalent senses to Gurevych, 2011), GN-WKT (Henrich et al., 2011) the problem of matching nodes in these graphs. Also and WN-WP (Niemann and Gurevych, 2011). Our for the first time, we comparatively evaluate a WSA goal is to cover resources with different character- algorithm on a variety of different datasets with dif- istics: Expert-built (WN, GN) and collaboratively ferent characteristics. constructed LSRs (WP, WKT, OW), resources in The key properties of Dijkstra-WSA are: different languages (English and German) and also Robustness The entities within the LSRs which resources with few sense descriptions (GN) or se- are to be aligned (usually senses or synsets) are mod- mantic relations (WKT). We contrastively discuss eled as nodes in the graph. These nodes are con- the results of the Dijkstra-WSA algorithm on these nected by an edge if they are semantically related. different datasets and relate the results to the prop- While, for instance, semantic relations lend them- erties of the LSRs involved. Moreover, using exist- selves very well to deriving edges, different possi- ing datasets ensures comparability to previous work bilities for graph construction are equally valid as which discusses only one dataset at a time. the algorithm is agnostic to the origin of the edges. WordNet (WN) (Fellbaum, 1998) is a lexical re- Language-independence No external resources source for the English language created at Princeton such as corpora or other dictionaries are needed; the University. The resource is organized in sets of syn- graph construction and alignment only rely on the onymous words (synsets) which are represented by information from the considered LSRs. glosses (sometimes accompanied by example sen- Flexibility The graph construction as well as the tences) and organized in a hierarchy. The latest ver- actual alignment are highly parameterizable to ac- sion 3.0 contains 117,659 synsets. commodate different requirements regarding preci- Wikipedia (WP) is a freely available, multilin- sion or recall. gual online encyclopedia. WP can be edited by ev- The rest of this paper is structured as follows: ery Web user, which causes rapid growth: By Febru- In Section 2 we give a precise problem description ary 2013 the English WP contained over 4,000,000 and introduce the resources covered in our experi- article pages. Each article usually describes a dis- ments, in Section 3 we discuss some related work, tinct concept, and articles are connected by hyper- while our graph-based algorithm Dijkstra-WSA is links within the article texts. presented in Section 4. We describe an evaluation Wiktionary (WKT) is the dictionary pendant to on four datasets with different properties, including WP. By February 2013 the English WKT contained an error analysis, in Section 5 and conclude in Sec- over 3,200,000 article pages, while the German edi- tion 6, pointing out directions for future work. tion contained over 200,000 ones. For each word, multiple senses can be encoded. Similar to WN, 2 Notation and Resources they are represented by a gloss and usage exam- ples. There also exist hyperlinks to synonyms, hy- 2.1 Problem Description pernyms, meronyms etc. The targets of these rela- A word sense alignment, or alignment for short, is tions are not senses, however, but merely lexemes formally defined as a list of pairs of senses from (i.e. the relations are not disambiguated). 152 Downloaded from http://www.mitpressjournals.org/doi/pdf/10.1162/tacl_a_00217 by guest on 28 September 2021 LSRs P /R/F1/Acc. Approach Meyer and Gurevych (2011) WN-WKT 0.67/0.65/0.66/0.91 Gloss similarity + Machine learning Niemann and Gurevych (2011) WN-WP 0.78/0.78/0.78/0.95 Gloss similarity + Machine learning Henrich et al. (2011) GN-WKT 0.84/0.85/0.84/0.94 Pseudo-gloss overlap de Melo and Weikum (2010) WN-WP 0.86/NA/NA/NA Gloss/article overlap Laparra et al. (2010) FN-WN 0.79/0.79/0.79/NA Dijkstra-SSI+ (WSD algorithm) Navigli (2009) WN 0.64/0.64/0.64/NA Graph-based WSD of WN glosses Ponzetto and Navigli (2009) WN-WP NA/NA/NA/0.81 Graph-based, only for WP categories Navigli and Ponzetto (2012) WN-WP 0.81/0.75/0.78/0.83 Graph-based WSA using WN relations Table 1: Summary of various approaches to WSA. “NA” stands for “Not Available”. OmegaWiki (OW) is a freely editable online glosses, which are not present in every case (e.g. for dictionary like WKT. However, there do not ex- VN). Moreover, as it involves supervised machine ist distinct language editions as OW is organized learning, it requires the initial effort of manually an- in language-independent concepts (“Defined Mean- notating a sufficient amount of training data. Hen- ings”) to which lexicalizations in various languages rich et al. (2011) use a similar approach for align- are attached. These can be considered as multilin- ing GN and WKT.