The most intuitive approach, reported in Mitchell and Lapata ( 2008, 2009, 2010), consists of combining vectors of two related words with arithmetic operations: component-wise addition and multiplication. In the last decade, different compositional models have been proposed and most of them use bag-of-words as basic representations of word contexts in the vector space. The key element of such a syntax-sensitive vector space, which is high dimensional and sparse, is the concept of selectional preferences (defined later in Section 3). This compositional strategy is useful to identify paraphrases, that is, similar composite expressions in the same language. In this approach, the vector space is structured with syntactic dependencies, and word senses are contextualized as words are combined with each other through their dependencies.
So, when two words are syntactically dependent, the compositional process builds two new vectors: one for the head expression and another one for the dependent one. This is the contextualized sense of ball given catch in the same relation dobj, which should denote a spherical object and not a dancing event. On the other hand, the vector of the dependent word, ball, is combined with the selectional preferences, noted catch h↑, imposed by the head catch on ball, so as to obtain a new compositional vector: ball dobj↓. This is the contextualized sense of the head catch given ball in relation dobj, which would be close to the meaning of grab and not to that of contract. When two words, catch and ball, are related by a syntactic dependency, for instance dobj (direct object), we actually perform two different combinations: on the one hand, we combine the vector of the head word, noted catch, with the selectional preferences, noted ball d↓, imposed by the dependent word ball on the head catch, in order to obtain a new compositional vector: catch dobj↑. In a monolingual vector space, we propose a compositional model based on that described in Erk and Padó ( 2008) and Erk et al. This task was known in machine translation as target word selection, namely, the task of deciding which target language word is the most appropriate equivalent of a source language word in context (Dagan 1991). The target expression with the most similar compositional vector to the vector of the source expression will be considered as its most likely (contextualized) translation. In a bilingual distributional framework, we call “contextualized translation” the construction of compositional vectors for the expressions in the target language that are similar to the compositional vectors of the expressions in the source language. Both sense disambiguation and language translation are sensitive to the compositional construction of new meanings (Brown et al. However, the meaning of ball refers to a dancing event when it is combined with attend in attend a ball, being translated now into Spanish by baile. On the other hand, the sense of ball when combined with catch designates a spherical object and its translation into Spanish is pelota. By contrast, this verb has a similar meaning to contract when combined with disease in the expression catch a disease, and its more appropriate translation into Spanish is now contraer. Given the expression catch a ball, the sense of catch combined with ball is similar to grab, and can be translated into Spanish by coger. A new bilingual data set to evaluate strategies aimed at translating phrasal verbs in restricted syntactic domains has been created and released. Experiments were performed on English and Spanish monolingual corpora in order to translate phrasal verbs in context. This process is expanded to larger phrases by means of incremental composition. A phrase in the source language, consisting of a head and a dependent, is translated into the target language by selecting both the nearest neighbor of the head given the dependent, and the nearest neighbor of the dependent given the head. The contextualization of meaning is carried out by means of distributional composition within a structured vector space with syntactic dependencies, and the bilingual space is created by means of transfer rules and a bilingual dictionary. Word translation is modeled in the same way as contextualization of word meaning, but in a bilingual vector space. This article describes a compositional distributional method to generate contextualized senses of words and identify their appropriate translations in the target language using monolingual corpora.