Preliminary Recommendations

Practical NLP lexicons

In this section an overview of the following practical lexicons for NLP is provided:

• ACQUILEX: a typed feature structure, multilingual lexicon developed within the Acquilex project for four different languages (Dutch, English, Italian and Spanish) following a sign-based approach to lexical specification (Sanfilippo, 1993b).
• COMLEX: the American English lexicon prepared for the Linguistic Data Consortium by the University of Pennsylvania, which uses a typed feature structure formalism (Rohen Wolff et al., 1994).
• EUROTRA: the lexicon for generation and analysis in the Eurotra Machine Translation system (ten Hacken et al., 1991).
• GENELEX: a full-scale application- and theory-independent lexicon developed within the Genelex project (GENELEX, 1993). It is assumed that from this general repository of lexical information theory-dependent lexical representations and/or application-specific dictionaries can easily be derived. The general architecture of the lexicon is conceptually based on the entity-relationship model.
• ILCLEX: an Italian lexicon, originally designed for integration with a robust wide-coverage corpus grammar, which collects information extracted from machine readable sources such as dictionaries and corpora (Vanocchi et al., 1994).
• LDOCE: the Longman Dictionary of Contemporary English (Procter, 1987), which was taken as a basis for the construction of several computational lexicons (see, for instance, Boguraev & Briscoe (1989).
• PLNLP: the lexicon used by the PLNLP Italian grammar, originally developed for a style-checking application, and later used for other different tasks, e.g. the acquisition of lexico-semantic information from machine readable dictionaries (Jensen et al., 1993).

The lexicons differ largely as to the number of entries they contain and to the amount of linguistic information each entry is provided with: while some lexicons attach more importance to coverage, others appear to concentrate on granularity and depth of lexical representations often to the detriment of breadth. This is mainly due to the fact that most such lexicons were aimed at a wide variety of different applications (ranging from machine translation to style-checking). Moreover, some are strongly biased towards specific linguistic theories. For the sake of concreteness, a specific case study was chosen as a basis for comparison between these lexicons: the encoding of verb entries and, particularly, of their subcategorisation frame. Although this area has been the focus of intense research over the last fifteen years, it still rests on fairly shaky ground.

In what follows, we will list all aspects which appear to be relevant for the encoding of subcategorisation and consider them in more detail. It should be noted that not all of the following are pertinent to all lexicons, as will be indicated below:

• Number of arguments;
• Syntactic category of arguments;
• Functional role of arguments;
• Control and raising phenomena;
• Lexical selection;
• Morphosyntactic constraints;
• Frame alternation and argument optionality;
• structure and deep structure;
• Other properties.

The division between the top of this list and the bottom is to be interpreted as indicative of a scale from the least to the most uniformly-distributed properties: in other words, the properties in the topmost part of the list are shared by all lexicons, the others are shared only by some and with considerable differences in the way these properties are represented therein. In what follows, we will first sketchily characterise the approach to verb subcategorisation that each lexicon takes, then move on to a more detailed consideration of the above listed aspects.

 
• Approaches to verb subcategorisation
• A comparative overview
• Summary