A Robust Approach to Aligning Heterogeneous Lexical Resources

Our approach for aligning lexical resources exploits the graph structure of each resource. Therefore, we assume that a lexical resource $L$ can be represented as an undirected graph $G=\left(V,E\right)$ where $V$ is the set of nodes, i.e., the concepts defined in the resource, and $E$ is the set of undirected edges, i.e., semantic relations between concepts. Each concept $c\in V$ is associated with a set of lexicalizations ${ℒ}_G\left(c\right)=\left\{w_1,w_2,\mathrm{\dots },w_n\right\}$. For instance, WordNet can be readily represented as an undirected graph $G$ whose nodes are synsets and edges are modeled after the relations between synsets defined in WordNet (e.g., hypernymy, meronymy, etc.), and ${ℒ}_G$ is the mapping between each synset node and the set of synonyms which express the concept. However, other resources such as Wiktionary do not provide semantic relations between concepts and, therefore, have first to be transformed into semantic networks before they can be aligned using our alignment algorithm. We explain in Section 3 how a semi-structured resource which does not exhibit a graph structure can be transformed into a semantic network.

Given a pair of lexical resources $L_1$ and $L_2$, we align each concept in $L_1$ by mapping it to its corresponding concept(s) in the target lexicon $L_2$. Algorithm 2 formalizes the alignment process: the algorithm takes as input the semantic graphs $G_1$ and $G_2$ corresponding to the two resources, as explained above, and produces as output an alignment in the form of a set $A$ of concept pairs. The algorithm iterates over all concepts $c_1\in V_1$ and, for each of them, obtains the set of concepts $C\subset V_2$, which can be considered as alignment candidates for $c_1$ (line 2). For a concept $c_1$, alignment candidates in $G_2$ usually consist of every concept $c_2\in V_2$ that shares at least one lexicalization with $c_1$ in the same part of speech tag, i.e., ${ℒ}_{G_1}\left(c_1\right)\cap {ℒ}_{G_2}\left(c_2\right)\ne \mathrm{\varnothing }$ [31, 20]. Once the set of target candidates $C$ for a source concept $c_1$ is obtained, the alignment task can be cast as that of identifying those concepts in $C$ to which $c_1$ should be aligned. To do this, the algorithm calculates the similarity between $c_1$ and each $c_2\in C$ (line 2). If their similarity score exceeds a certain value denoted by $\theta$ (line 2), the two concepts $c_1$ and $c_2$ are aligned and the pair $\left(c_1,c_2\right)$ is added to $A$ (line 2).

Different resource alignment techniques usually vary in the way they compute the similarity of a pair of concepts across two resources (line 2 in Algorithm 2). In the following, we present our novel approach for measuring the similarity of concept pairs.

[t!] Lexical Resource Aligner {algorithmic}[1] \REQUIREgraphs $H=(V_H,E_H)$, $G_1=(V_1,E_1)$ and $G_2=(V_2,E_2)$, the similarity threshold $\theta$, and the combination parameter $\beta$ \ENSURE$A$, the set of all aligned concept pairs

$A\leftarrow \mathrm{\varnothing }$

concept $c_1\in V_1$ \STATE$C$ $\leftarrow$ getCandidates$(c_1,V_2)$ \FORALLconcept $c_2\in C$ \STATE$sim$ $\leftarrow$ calculateSimilarity$(H,G_1,G_2,c_1,c_2,\beta )$ \IF$sim>\theta$ \STATE$A$ $\leftarrow$ $A$ $\cup$ $\{(c_1,c_2)\}$ \ENDIF\ENDFOR

$A$

To enable a comparison with the state of the art, we followed Matuschek and Gurevych (2013) and performed an alignment of WordNet synsets (wn) to three different collaboratively-constructed resources: Wikipedia (wp), Wiktionary (wt), and OmegaWiki (ow). We utilized the DKPro software [35, 9] to access the information in the foregoing three resources. For wp, wt, ow we used the dump versions $20090822$, $20131002$, and $20131115$, respectively.

We followed previous work [25, 17] and evaluated the alignment performance in terms of four measures: precision, recall, F1, and accuracy. Precision is the fraction of correct alignment judgments returned by the system and recall is the fraction of alignment judgments in the gold standard dataset that are correctly returned by the system. F1 is the harmonic mean of precision and recall. We also report results for accuracy which, in addition to true positives, takes into account true negatives, i.e., pairs which are correctly judged as unaligned.

Here, we describe how the four semantic graphs for our four lexical resources (i.e., wn, wp, wt, ow) were constructed. As mentioned in Section 2.1.1, we build the wn graph by including all the synsets and semantic relations defined in WordNet (e.g., hypernymy and meronymy) and further populate the relation set by connecting a synset to all the other synsets that appear in its disambiguated gloss. For wp, we used the graph provided by Matuschek and Gurevych (2013), constructed by directly connecting an article (concept) to all the hyperlinks in its first paragraph, together with the category links. Our wn and wp graphs have 118K and 2.8M nodes, respectively, with the average node degree being roughly 9 in both resources.

The other two resources, i.e., wt and ow, do not provide a reliable network of semantic relations, therefore we used our ontologization approach to construct their corresponding semantic graphs. We report, in the following subsection, the experiments carried out to assess the accuracy of our ontologization method, together with the statistics of the obtained graphs for wt and ow.

Having lexical resources represented as semantic networks is highly beneficial. A good example is WordNet, which has been exploited as a semantic network in dozens of NLP tasks [7]. A recent prominent case is Wikipedia [18, 13] which, thanks to its inter-article hyperlink structure, provides a rich backbone for structuring additional information [2, 34, 23, 8]. However, there are many large-scale resources, such as Wiktionary for instance, which by their very nature are not in the form of a graph. This is usually the case with machine-readable dictionaries, where structuring the resource involves the arduous task of connecting lexicographic senses by means of semantic relations. Surprisingly, despite their vast potential, little research has been conducted on the automatic ontologization of collaboratively-constructed dictionaries like Wiktionary and OmegaWiki. Meyer and Gurevych (2012a) and Matuschek and Gurevych (2013) provided approaches for building graph representations of Wiktionary and OmegaWiki. The resulting graphs, however, were either sparse or had a considerable portion of the nodes left in isolation. Our approach, in contrast, aims at transforming a lexical resource into a full-fledged semantic network, hence providing a denser graph with most of its nodes connected.

Aligning lexical resources has been a very active field of research in the last decade. One of the main objectives in this area has been to enrich existing ontologies by means of complementary information from other resources. As a matter of fact, most efforts have been concentrated on aligning the de facto community standard sense inventory, i.e. WordNet, to other resources. These include: the Roget’s thesaurus and Longman Dictionary of Contemporary English [15], FrameNet [16], VerbNet [33] or domain-specific terminologies such as the Unified Medical Language System [4]. More recently, the growth of collaboratively-constructed resources has seen the development of alignment approaches with Wikipedia [32, 2, 34, 31, 25], Wiktionary [20] and OmegaWiki [9]. Last year Matuschek and Gurevych (2013) proposed Dijkstra-WSA, a graph-based approach relying on shortest paths between two concepts when the two corresponding resources graphs were combined by leveraging monosemous linking. Their method when backed off with other definition similarity based approaches [27, 20], achieved state-of-the-art results on the mapping of WordNet to different collaboratively-constructed resources. This approach, however, in addition to setting the threshold for the definition similarity component by means of cross validation, also required other parameters to be tuned, such as the allowed path length ($\lambda$) and the maximum number of edges in a graph. The optimal value for the $\lambda$ parameter varied from one resource pair to another, and even for a specific resource pair it had to be tuned for each configuration. This made the approach dependent on the training data for the specific pair of resources that were to be aligned. Instead of measuring the similarity of two concepts on the basis of their distance in the combined graph, our approach models each concept through a rich vectorial representation we refer to as semantic signature and compares the two concepts in terms of the similarity of their semantic signatures. This rich representation leads to our approach having a good degree of robustness such that it can achieve competitive results even in the absence of training data. This enables our system to be applied effectively for aligning new pairs of resources for which no training data is available, with state-of-the-art performance.