On the semantics of focus phenomena in EUROTRA

In this paper, we discuss issues connected to the phenomenon of linguistic FOCUS or INFORMATION DISTRIBUTION in the sentence in the context of the multi-lingual machine translation project EUROTRA. We shall present some of the arguments why a consideration of FOCUS phenomena is important for the determination of linear order and for semantic interpretation. We shall proceed, in sections 2 and 3 of the paper, to mention the main lines of development in the dicussion of FOCUS phenomena in Computational Linguistics and in Linguistics respectively. Section 4 contains an illustration of a pilot implementation covering some aspects of FOCUS phenomena in EUROTRA-D. 1. Problem description 1.1. The determination of linear order in the clause An initial, but faulty assumption, underlying a possible theory of language is that linear order of syntactic constituents in the clause is fully determined by a combination of the following types of information: basic type of language (SVO,SOV,etc) clause modus (indicative, interrogative,etc.) information on verbs in the lexicon (lexical forms). This paper will present some of the arguments why this assumption is wrong for the clause, in how far this fact is relevant for EUROTRA, and initial suggestions towards a treatment within our framework. We shall not be able here, for reasons of time, to deal with linear order within NPs, ADJPs, or ADVPs (cf. for English on these questions Halliday.1985 : 158ff, and many others). At least the following hierarchy of factors can be said to contribute to linear sequence (cf. Allen.1987 : 51): (1) The familiarity hierarchy (2) The topic < comment, given < new hierarchies (3) The universal sequencing conventions (4) The definitieity and referentiality hierarchies (5) The personal, social status, and role hierarchies (6) The dominant descriptor hierarchies (7) The formal hierarchies Of these, this paper will deal with focus phenomena, covered in (2) above. Different aspects of this area are covered in the linguistic literature under the headings of "topic-comment", "focus-presupposition", "theme-rheme", and "given-new". The difficulty with these pairs of terms is that they reflect a difference in orientation (rhetorical vs. logical vs. psychological); that they cover the related, though different aspects of linear order and intonation. 630 Rather than go into an extended theoretical discussion at this point, we shall in 1.2. discuss a range of examples to show where exactly these phenomena would seem to be relevant for a multi-lingual MT-system like EUROTRA. The system within which the implementation of the ideas suggested here were implemented is the Euro t ra -D system as described in Steiner,1986 and Steiner et al.forthcoming and in other places. 1.2. The relevance of linear order for semantic interpretation 1.2.1. The scope of negation Cf. the following pair of sentences (capitals standing for stress): (1) The eee is not controversial, because it is a multi-natlonal ORGANIZATION. (2) Because it is a multi-national organization the eec is not CONTROVERSAL (1) and (2) are identical, except for the position of the Adverbial Clause, and the assignment of primary information focus realized by main stress and symbolized here by capitals. However, their semantic interpretation differs with respect to the scope of negation: (1) is ambiguous with respect to the queston of whether or not the propostion expressed in the matrix clause is true, cf. (3): • (3) The eee is not controversial, because it is a multi-national ORGANISATION. It is controversial because it is EXPENSIVE. The second sentence in (3) is not possible with (1), whereas it is with (2). In (1), the scope of negation includes the Adverbial Clause, whereas in (2), it does not. In general, it seems to be the case that the scope of negation may include everything in the clause to the right of the negation particle and including the constituent which has a lexical item carrying focus. Therefore, the truth values of (1) and (2) are different, which implies that they are not acceptable translations of each other in a truth value oriented semantics, such as we are having it at the moment in EUROTRA. Yet, in the present framework, (1) and (2) would receive one representation at ERS, which means that ECS-ERS translations are not translationally equivalent. The implications for other pairs of levels of representation are obvious. We shall not go into the question here of how marked intonation in such cases would influence truth values as we ar0 dealing with written texts exclusively, we shall assume unmarked intonation for each syntactic structure, which implies that in the case of thematizing constructions, we shall assume main stress in the constituent functioning as a marked theme, and stress within the rightmost clause constituent carrying a fully lexical item otherwise. All this, for the time being, applies to English. 1.2.2, The scope of quantification