Professional Documents
Culture Documents
Similarity Measures
A. A. P. Ribeiro1 , L. Zhao1 , A. A. Macedo1
Abstract. This article presents a comprehensive review on knowledge-based semantic similarity measures and
similarity measures for text collections widely used in the literature. Such similarity measures have been applied in
various types of tasks ranging from Information Retrieval to Natural Language Processing. Taking advantage of this
review, we can systematize and compare them according to the type of data for structured, semi-structured, and
multimedia data retrieval. For each type of data, usually there are different ways to measure the similarity. Therefore,
the present review also contributes to the scientific community in such a way that it makes easier the comparison of
semantic measures in terms of their evaluations, the selection of semantic measures according to a specific usage context,
and the summarization of theoretical findings related to semantic measures.
Categories and Subject Descriptors: H.2.8 [Database Management]: Database Applications; I.2.6 [Artificial Intel-
ligence]: Learning
1. INTRODUCTION
Data similarity measures have long been studied due to their vast applications in computer science
and in other branches of science. In the last years, textual similarity measures permanently over-
played in a wide range of research areas, acting on information extraction and processing, specially, in
tasks of information retrieval, textual classification, clustering of documents, topic detection, tracking,
generation of questions, question-answering applications, machine translation, textual summarization,
data mining, Web search, clustering and recommender systems. Classic similarity measures can be
classified into two approaches: (i) similarity measures based on features and (ii) similarity measures
based on links. The first group of measures calculates the similarity between objects by considering
their feature vectors. On the other hand, the similarities based on links define measures in terms of
the structural composition of a graph. The similarities based on features are the textual similarity al-
gorithms that can be further classified into the following four categories: (i) similarity measure based
on strings, (ii) similarity measures based on terms, (iii) similarity measures based on collections and
(iv) similarity measures based on knowledge. This article is going to deal with the knowledge-based
and collection-based similarity.
Most of the similarity measures based on features are defined in vector space, for example, the
Cosine is the most traditional measure calculates the similarity between two non-zero vectors in an
inner product space [Salton 1989]; the Manhattan distance is the distance between two objects in a
grid based on a rigidly horizontal and/or vertical path [Krause 2012]; the Euclidean distance measures
the distance between two points in Euclidean space [Greub 1967]; the Jaccard is a statistic similarity
measure used for comparing the similarity and diversity of sample sets [Jaccard 1901]; additionally,
Copyright
2017
c Permission to copy without fee all or part of the material printed in KDMiLe is granted provided that
the copies are not made or distributed for commercial advantage, and that notice is given that copying is by permission
of the Sociedade Brasileira de Computao.
the Dice measure is defined as twice the number of common terms in the compared strings divided by
the total number of terms [Dice 1945]; the Overlap Matching coefficient counts the number of similar
terms (dimensions) [Ukkonen 1990]. Overlap is similar to the Dice coefficient, but it measures how
similar two strings are in terms of the number of common bigrams; Charikar Similarity is a similarity
measure related to the Jaccard index that measures the overlap between two sets and it is defined as
the size of the intersection divided by the smaller of the size of the two sets [Charikar 2002]. Similarity
measures applied to specific domains (such as [Ehsani and Drabls 2016]) are not considered.
This article presents a review on similarity measures for text collections and knowledge-based se-
mantic measures widely used in the information retrieval and information extraction literature. The
remaining sections of this article are organized as follows: Section 2 presents the similarity measured
reviewed by the authors, Section 3 details a comparative study on the similarity measures, and Section
4 concludes with final remarks and future work.
2. SIMILARITY MEASURES
Usually, similarity measures for text document collection are presented as semantic similarity coeffi-
cients that quantify the similarity between textual information (words, sentences, paragraphs, docu-
ments, etc) based on information obtained from corpora1 . Similarity measures at concept level is a
kind of semantic similarity to identify the similarity between words using information extracted from
semantic networks2 . The next two subsection present, respectively, similarity measures for text col-
lection and knowledge-based semantic measures. Usually, similarity measures for text collections are
called statical similarity measures and knowledge-based similarity measures are semantic measures.
1A corpus is a large collection of textual documents which is mainly used for information extraction, information
retrieval, natural language researches.
2 A semantic network is a graph to knowledge representation of semantic relations between concepts (nodes).
2005]. GLSA extends LSA focusing in term vectors instead of documents as LSA. GLSA is not based
on bag-of-words3 . It exploits pair-wise term similarities to compute a representation for terms. GLSA
demands a measure of semantic similarity between terms and a method of dimensionality reduction.
GLSA combines any similarity measure and any reduction of dimensionality.
Explicit Semantic Analysis (ESA) uses a corpus of documents as a knowledge base, it represents
the individual words or entire documents as vectorial representations of text documents such as HAL,
LSA and GSA [Gabrilovich and Markovitch 2007]. The ESA is a measure to calculate semantic
relationships between any pair of documents, any corpora including Wikipedia articles and the Open
Directory Project [Egozi et al. 2011].
The documents are represented as centroids of vectors representing its words. The words are
represented as a column vector in the Term-Frequency and Inverse Term-Frequency (TF-IDF) array
of the text corpus. The terms or texts of ESA are portrayed by vectors with high dimension. Each
element of the vector represents the pair TF-IDF between terms of documents. The semantic similarity
between two terms or texts is expressed by the cosine measure between the corresponding vectors.
However, unlike LSA, ESA deals with human-readable labels transforming them into concepts that
make up the vector space. The conversions are possible thanks to the use of a knowledge base [Egozi
et al. 2011; Gabrilovich and Markovitch 2007]. The scheme is extended from single words to multi-
words documents by simply summing the arrays of all words in the documents [Gabrilovich and
Markovitch 2007]. The semantic relatedness of the words is given by a numeric estimation.
Cross-Language Explicit Semantic Analysis (CLESA) is a multilingual generalization of ESA
[Potthast et al. 2008]. CLESA manipulates documents that are aligned with a multilingual reference
collection that corresponds documents as vectors of concepts without considering the languages. The
relationships between two documents in different languages are calculated by cosine considering the
vector space. A document written in a specific language is represented as ESA vector by using an
index document collection in the language. The similarity between a document and a document from
another language is quantified in the concept space, by computing the cosine similarity between both.
Pointwise Mutual Information - Information Retrieval (PMI-IR) is a method to calculate
similarities between pairs of words, using the AltaVistas Advanced Search Query [Friedman 2004],
calculating the probabilities of similarity the Alta Vista calculates the similarity [Turney 2001]. This
probability is based on the proximity of the pair of words in Web pages, considering greater proximity
greater similarity. The PMI-IR algorithm, like LSA, is based on co-occurrence [Manning et al. 1999].
The core idea is that a word is characterized by the neighborhood it has [Firth 1957]. There are
many different measures of the degree to which two words co-occur [Manning et al. 1999]. Therefore,
the ratio between one probability to another is a measure of the degree of statistical dependence.
Once the equation is symmetrical, it is the amount of information that acquires about the presence
that explains the term mutual information.
Second-order Co-occurrence - Pointwise Mutual Information (SCO-PMI) is a semantic
similarity measure that applies PMI-IR to sort the list of neighboring words of two target words being
compared in a collection [Islam and Inkpen 2008; 2006]. The advantage of using SCO-PMI is that
it calculates the similarity between two words that they do not co-occur frequently, but the same
neighboring words are co-occurring. The preprocessed word pairs are taken to calculate semantic
word similarity using SCO-PMI. It is a corpus-based method for determining the semantic similarity
of two target words. Evaluation result shows that the method outperforms several competing corpus-
based methods. This method focuses on measuring the similarity between two target words. After
finding the similarity between all words in the document, the retrieval of similar information can be
performed to user query too.
3 Bag-of-words
names a model that ignores context, semantic and order of words, simplifying computational efforts. As
vocabulary may potentially run into millions, this model faces scalability challenges.
Normalized Google Distance (NGD) is a semantic similarity measure using the number of
hits returned by the Google search engine for a given set of keywords [Cilibrasi and Vitanyi 2007].
This algorithm returns the keywords with the same meanings or similar meanings based on natural
language processing. The words are similar if that tend to be near" in distance units of Google, while
words with different meanings tend to be more distant.
Similarity through the Co-occurrence Distribution (DISCO) assumes that words with sim-
ilar meanings occur in a similar context [Kolb 2009]. Large text collections are statistically analyzed
to obtain similarity of distribution. The DISCO is a method that calculates the similarity of distri-
bution between the words using a simple context window of a size of approximately three words for
the count of co-occurrences. When two words are submitted to the calculation of the exact similarity,
DISCO simply retrieves its vectors from the indexed data and calculates the similarity according to
the Lin measure [Lin 1998], presented next. If the most similar word according to the distribution
is requested, DISCO returns the second order of the word vector. DISCO has two main similarity
measures: DISCO1 and DISCO2. The DISCO1 calculates the first order similarity between two input
words according to the word arrangement sets. The DISCO2 calculates the second order similarity
between two input words, according to the distribution of similar words.
Knowledge-based measures try to identify the degree of similarity among concepts by using algorithms
supported by lexical resources and/or semantic networks. The similarity measures based on knowledge
can be separated into two groups: the semantic relatedness measures and the semantic-based measures.
The semantic relatedness measures indicate the strength of the semantic interactions between ob-
jects since there are no constraints on the quality of the considered semantic links. Semantic re-
latedness similarities measures is a category of relationships between two words, incorporating a
bigger range of relationships between concepts such as is_type_of, is_one_exemple_specific_of,
is_part_of, is_the_opposite_of [Patwardhan et al. 2003]. The most used examples of semantic
relatedness measures are Resnik (RES) [Resnik 1995; KG and SADASIVAM 2017], Lin (LIN) [Lin
1998], Jiang & Conrath (JCN) [Jiang and Conrath 1997], St.Onge (HSO) [Hirst et al. 1998],
Lesk (LESK) [Banerjee and Pedersen 2002], and Pairs of Vectors (Vectors) [Patwardhan 2003].
RES is a measure of semantic similarity based on the notion of information content by considering an
is a taxonomy. The value of RES is the information content of the Least Common Subsumer 4 [Resnik
1995; KG and SADASIVAM 2017].
LIN suggests the semantic similarity between two topics in a taxonomy [Lin 1998]. LIN is defined
as a function of the meaning shared by the topics and the meaning of each of the individual topics.
The meaning shared by two topics can be recognized by looking at the lowest common ancestor,
which corresponds to the most specific common classification of the two topics. Once this common
classification is identified, the meaning shared by two topics can be measured by the amount of
information needed to state the commonality of the two topics. The semantic similarity is defined
in terms of the hierarchical taxonomy. The disadvantage is to capture the semantic relationships in
non-hierarchical components.
JCN is a hybrid similarity measure that mixes words or concepts. It combines a taxonomy structure
with measures based on corpus. Exploiting the best of both, the taxonomy helps to guarantee seman-
tics, while the statistical approach ensures the evidence of the distribution of the exploited corpus.
LIN and JCN increase the information content of the Least Common Subsumer by considering the
sum of the content of concepts. LIN scales the content of the Least Common Subsumer, while JCN
assigns the difference between the sum and the information content of the Least Common Subsumer.
The measure HSO finds lexical chains of strings relating two meanings of a word [Hirst et al. 1998].
This measure calculates relatedness between concepts using the path distance between the concept
nodes, number of changes in direction of the path connecting two concepts and the allowableness of
the path. When the relation between concepts is close, they are semantically related to each other
[Choudhari 2012].
The LESK discovers overlaps in the glossary of synsets of WordNet (it is presented next). The
relativeness score is the sum of the squares of the overlap lengths. On the other hand, the measure
Vector creates for each word of the specific glossary of WordNet [Patwardhan 2003]. After, it rep-
resents each glossary/concept as a vector that is the mean of the co-occurrence vectors. Vector was
developed as a measure of semantic relatedness that represents concepts using context vectors, and it
is able to establish relatedness by measuring the angle between these vectors. This measure combines
the information from a dictionary with statistical information derived from large corpora of text. In
other words, semantic relatedness is then measured simply as the nearness of the two vectors in the
multidimensional space (the cosine of two normalized vectors). One of the strengths of this measure is
that the basic idea of the Vector measure can be used with any dictionary, disregarding the WordNet.
The semantic-based measures consist of four categories: (i) the semantic similarity measures the
taxonomic term relationships to extract the similarity; (ii) the semantic distance is the inverse of
the semantic relativeness; (iii) the semantic dissimilarity is the inverse semantic similarity; and (iv)
the taxonomic distance is related to the dissimilarity. In the literature, the most cited examples of
semantic-based measures are Leacock & Chodorow (LCH) [Leacock and Chodorow 1998], Wu &
Palmer (WUP) [Wu and Palmer 1994], and Path Length (Path) [Wu and Palmer 1994].
The LCH measures the length of the shortest path between two concepts using node-counting and
exploiting the maximum depth of the taxonomy [Leacock and Chodorow 1998]. It returns a score
indicating the similarity between the different meanings of a word. It considers the shortest path
connecting the meanings and the maximum depth of the taxonomy in which the meanings occurs. On
the other hand, the WUP measures the similarity between two meanings of a word considering the
depth of the two meanings in the taxonomy and its Least Common Subsumer [Wu and Palmer 1994] .
WUP is a prototype lexical selection system called UNICON, that represents the English and Chinese
verbs based on a set of shared semantic domains and the selection information is also included in these
representations without exact matching. The concepts are organized into hierarchical structures to
form an interlingua conceptual base. The input to the system is the source verb argument structure.
The measure Path quantifies the similarity between two meanings of a word based on the shortest
path linking between the two meanings in the taxonomy is_a" (hyperonyms/homonyms) [Wu and
Palmer 1994]. This measure is inversely proportional to the number of nodes along the shortest path
between the concepts. The shortest possible path occurs when two concepts are the same, in which
case the length is 1. Thus, the maximum similarity value is 1.
Overall, the presented approaches are distinct algorithms used to calculate semantic similarity
between concepts. The semantic similarity can be improved by using human knowledge to generate a
more accurate measure. Human knowledge usually is expressed in dictionaries, taxonomies, ontologies
and concept networks.
The WordNet is the most popular concept network to be used to measure similarity based on
knowledge [Miller et al. 1990]. This concept network is a large graph, or a lexical database, where
each node represents a real world concept that are English words classified as noun, verb, adjective
and adverb. All words are grouped into sets of cognitive synonyms (synsets) and each one expresses
a distinct concept, for example, the concept-object like a house, or an entity like a teacher, or an
abstract concept like art, and so on. Every node of the network consists of a set of synonyms words
called synsets. Each synset represents the real world concept associated with that node and it has
associated with it a gloss (short definition or description of the real world concept). The synsets
and the glosses are similar to the content of an ordinary dictionary such as the synonyms and the
definitions, respectively. Synsets are interlinked by means of conceptual-semantic and lexical relations.
Each link or edge describes a relationship between the real world concepts represented by the synsets
that are linked. Types of relationships are opposite of, is a member of, causes, pertains to, is a
kind of, is a part of and others. The network of relations between word senses present in WordNet
encodes a vast amount of human knowledge, giving rise to a great number of possibilities of knowledge
representation used for various tasks. WordNet has been manipulated by different approaches in order
to automatically extract its association relations and to interpret these associations in terms of a set
of conceptual relations, such as the DOLCE foundational ontology [Gangemi et al. 2003]. In terms of
limitations, WordNet does not present the etymology or the pronunciation of words and it is basically
composed of the everyday English word.
Agirre & Rigau developed a notion of conceptual density to create their algorithm for Word
Sense Disambiguation [Agirre and Rigau 1997]. They used the context of a given word along with the
hierarchy of is-a relations in WordNet to find the exact sense of the word. It divides the network
hierarchy of WordNet into sub-hierarchies and each of the senses of the ambiguous word belongs to
one sub-hierarchy. The conceptual density for each sub-hierarchy is then calculated using a conceptual
density formula which, intuitively, describes the amount of space occupied by the context words in
each of the sub-hierarchies.
There is no standard way to evaluate similarity measures without the agreement of human judgments.
Hence, computational similarity measures must rate the similarity of a set of word pairs, after it
is necessary to look at how well their ratings correlate with human ratings of the same pairs. To
evaluate measures, it is also useful to compare them by considering important features used by the
most applications. Thereafter, assumptions can empower the scientific community in making easier
the selection of measures according to a usage scenario, and in helping the summarization of findings
related to the measures. Here, we present a comparative study of similarity measures.
Table I presents similarity measures and the main requirements of similarity algorithms. The
similarity measures are the nineteen presented measures and the classical measures such as Cosine,
Jaccard, Dice Measure, Euclidean distance, Overlap Coefficient and Manhattan distance. The main
requirements are Changeable Granularity (what is the granularity of information to be manipulated
word, paragraph, document, etc), Partial Matching, Ranking Relevance Allowed, Terms Weights (if
there is a term-weighting scheme), Easy Implementation, Size Document Dependency, Dependency
of Ordered Terms, Semantic Sensibility (what kind of knowledge-based resource is used), and the
implemented approach by the measure.
A comparative analysis shows that the classical vector space models (measures from line 1 to line
6) are used just for a very raw text level and they have no addition of information enriched by any
semantic nor ontologies. On the other hand, the statical and semantic similarity measures (from line
7 to line 19) have semantic aggregated in the process.
The classical measures and the semantic measures allow the modification of the level of granularity
of the manipulated information. However, the statistical measures are not flexible in terms of that.
The classical measures are dependent on the size and terms and they do not aggregate semantic. The
statistical measures are not so dependent, but they usually have a higher computational cost and the
semantic is focused on the collection. Finally, the semantic or knowledge-based measures are full of
the specified requirements, but they can be a laborious task of implementation. In many cases, they
are specific-domain.
The columns are: 1: Changeable Granularity; 2: Partial Matching; 3: Ranking Relevance Allowed; 4: Terms Weights;
5: Easy Implementation; 6: Size Document Dependency; 7: Dependency of Terms; 8 Semantic Sensibility.
The commonly used collection/knowledge-based similarity measures and corpus-based similarity mea-
sures have been covered in this review article. Up to now, the similarity measure is still an important
subject under study and new ones are continuously proposed. Nevertheless, the future tendencies are
appointing to the representation of text and its relation on networks. An example is the WordNet and
ontologies, which have shown the effectiveness of the representation of textual data as a network. This
tendency is supported by the whole new branch of network science, working on large-scale graphs with
non-trivial topological structures, the complex networks. Now, we are working on this direction. As
a result, the presented similarity measures are being augmented by network measures and they will
be quantitative and qualitatively compared. Formulas and more detailed descriptions will be included
and an application will carry out all measures.
5. ACKNOWLEDGEMENTS
The authors would like to thank CAPES (1569180), FAPESP (2016/13206-4) and CNPq (302031/2016-
2) for their financial support.
REFERENCES
Agirre, E. and Rigau, G. A proposal for word sense disambiguation using conceptual distance. AMSTERDAM
STUDIES IN THE THEORY AND HISTORY OF LINGUISTIC SCIENCE SERIES 4 , 1997.
Banerjee, S. and Pedersen, T. An adapted lesk algorithm for word sense disambiguation using wordnet. In
International Conference on Intelligent Text Processing and Computational Linguistics. Springer, pp. 136145, 2002.
Burgess, C. and Lund, K. Multiple constraints in syntactic ambiguity resolution: A connectionist account of psy-
cholinguistic data. COGSCI-94, Atlanta, GA, 1994.
Charikar, M. S. Similarity estimation techniques from rounding algorithms. In Proceedings of the thiry-fourth annual
ACM symposium on Theory of computing. ACM, pp. 380388, 2002.
Choudhari, M. Extending the hirst and st-onge measure of semantic relatedness for the unified medical language
system. Ph.D. thesis, University of Minnesota, 2012.
Cilibrasi, R. L. and Vitanyi, P. M. The google similarity distance. IEEE Transactions on knowledge and data
engineering 19 (3): 370383, 2007.
Dice, L. R. Measures of the amount of ecologic association between species. Ecology 26 (3): 297302, 1945.
Egozi, O., Markovitch, S., and Gabrilovich, E. Concept-based information retrieval using explicit semantic
analysis. ACM Transactions on Information Systems (TOIS) 29 (2): 8, 2011.
Ehsani, R. and Drabls, F. Topoicsim: a new semantic similarity measure based on gene ontology. BMC bioinfor-
matics 17 (1): 296, 2016.
Firth, J. R. A synopsis of linguistic theory, 1930-1955, 1957.
Friedman, B. G. Web search savvy: Strategies and shortcuts for online research. Psychology Press, 2004.
Gabrilovich, E. and Markovitch, S. Computing semantic relatedness using wikipedia-based explicit semantic
analysis. In IJcAI. Vol. 7. pp. 16061611, 2007.
Gangemi, A., Navigli, R., and Velardi, P. The ontowordnet project: extension and axiomatization of concep-
tual relations in wordnet. In OTM Confederated International Conferences" On the Move to Meaningful Internet
Systems". Springer, pp. 820838, 2003.
Greub, W. H. Linear Algebra: 3d Ed. Springer, 1967.
Hirst, G., St-Onge, D., et al. Lexical chains as representations of context for the detection and correction of
malapropisms. WordNet: An electronic lexical database vol. 305, pp. 305332, 1998.
Islam, A. and Inkpen, D. Second order co-occurrence pmi for determining the semantic similarity of words. In Proc.of
the International Conference on Language Resources and Evaluation, Genoa, Italy. pp. 10331038, 2006.
Islam, A. and Inkpen, D. Semantic text similarity using corpus-based word similarity and string similarity. ACM
Transactions on Knowledge Discovery from Data (TKDD) 2 (2): 10, 2008.
Jaccard, P. Etude comparative de la distribution florale dans une portion des Alpes et du Jura. Impr. Corbaz, 1901.
Jiang, J. J. and Conrath, D. W. Semantic similarity based on corpus statistics and lexical taxonomy. arXiv preprint
cmp-lg/9709008 , 1997.
KG, S. and SADASIVAM, G. S. Modified heuristic similarity measure for personalization using collaborative filtering
technique. Appl. Math 11 (1): 307315, 2017.
Kolb, P. Experiments on the difference between semantic similarity and relatedness. In Proceedings of the 17th Nordic
Conference on Computational Linguistics-NODALIDA09, 2009.
Krause, E. F. Taxicab geometry: An adventure in non-Euclidean geometry. Courier Corporation, 2012.
Landauer, T. K. and Dumais, S. T. A solution to platos problem: The latent semantic analysis theory of acquisition,
induction, and representation of knowledge. Psychological review 104 (2): 211, 1997.
Leacock, C. and Chodorow, M. Combining local context and wordnet sense similarity for word sense identification.
wordnet, an electronic lexical database, 1998.
Lin, D. Extracting collocations from text corpora. In First workshop on computational terminology. Citeseer, pp.
5763, 1998.
Lund, K. and Burgess, C. Producing high-dimensional semantic spaces from lexical co-occurrence. Behavior Research
Methods, Instruments, & Computers 28 (2): 203208, 1996.
Lund, K., Burgess, C., and Atchley, R. A. Semantic and associative priming in high-dimensional semantic space.
In Proceedings of the 17th annual conference of the Cognitive Science Society. Vol. 17. pp. 660665, 1995.
Manning, C. et al. Foundations of statistical natural language processing. Vol. 999. MIT Press, 1999.
Matveeva, I., Levow, G.-A., Farahat, A., and Royer, C. Generalized latent semantic analysis for term represen-
tation. In Proc. of RANLP, 2005.
Miller, G. A., Beckwith, R., Fellbaum, C., Gross, D., and Miller, K. J. Introduction to wordnet: An on-line
lexical database. International journal of lexicography 3 (4): 235244, 1990.
Osgood, C., Suci, G., and Tannenbaum, P. The measurement of meaning. University of Illinois Press, 1964.
Patwardhan, S. Incorporating dictionary and corpus information into a context vector measure of semantic related-
ness. Ph.D. thesis, University of Minnesota, Duluth, 2003.
Patwardhan, S., Banerjee, S., and Pedersen, T. Using measures of semantic relatedness for word sense disam-
biguation. In International Conference on Intelligent Text Processing and Computational Linguistics. Springer, pp.
241257, 2003.
Potthast, M., Stein, B., and Anderka, M. A wikipedia-based multilingual retrieval model. In European Conference
on Information Retrieval. Springer, pp. 522530, 2008.
Resnik, P. Using information content to evaluate semantic similarity in a taxonomy. In Proceedings of the 14th
International Joint Conference on Artificial Intelligence - Volume 1. IJCAI95. Morgan Kaufmann Publishers Inc.,
San Francisco, CA, USA, pp. 448453, 1995.
Salton, G. Automatic text processing. addison welsley. Reading, Massachusetts vol. 4, 1989.
Turney, P. Mining the web for synonyms: Pmi-ir versus lsa on toefl, 2001.
Ukkonen, E. A linear-time algorithm for finding approximate shortest common superstrings. Algorithmica 5 (1):
313323, 1990.
Wu, Z. and Palmer, M. Verbs semantics and lexical selection. In Proceedings of the 32nd annual meeting on
Association for Computational Linguistics. Association for Computational Linguistics, pp. 133138, 1994.