The lexicon in computational linguistics
Computers’ ability to store and process large amounts of information makes them particularly valuable tools for the modelling of lexical knowledge, easily the aspect of human linguistic ability in which the greatest quantity of information has to be manipulated. Apart from their speed, computer lexical databases have a major advantage over dictionaries, their paper competitors. Whereas dictionaries only allow words to be searched for alphabetically, an electronic database allows for as many different search criteria as there are different data-codes in the entries (Fellbaum 1998: 7). This allows a much more efficient retrieval of information.
The most comprehensive attempt to model lexical knowledge on a computer is the WordNet project, which has been running since 1985. WordNet is an online lexical database which sets out to represent and organize lexical semantic information in a psychologically realistic form that facilitates maximally efficient digital manipulation. So far, the creators of WordNet have completed the coding of many thousands of English nouns, verbs, adjectives and adverbs; similar work has also been begun in a range of other languages.
The basic motivation behind WordNet was the ‘patterning assumption’: the assumption that ‘ people could not master and have readily available all the lexical knowledge needed to use a natural language unless they could take advantage of systematic patterns and relations among the meanings that words can be used to express’ (Miller 1998: xv). WordNet shows words’ mutual interrelations within the system (the ‘net’) of lexical relations such as synonymy, antonymy, hyponymy, entailment, and so on (see Chapter 5). The main organizational unit in WordNet is the synset, conventionally marked by curly brackets. Synsets are groupings of near-synonyms, like {beat, crush, trounce, vanquish}, which identify a particular lexicalized concept (word sense): here, the ‘defeat’ sense of beat. Each synset is given a short definition or gloss: ‘come out better in a competition, race, or conflict’ is the WordNet gloss of this syn set, and serves as a definition of all the words in that synset.
Polysemous words are dealt with by being assigned to multiple synsets, one for each sense. Thus beat also belongs to the following synsets (Fellbaum 1990: 279): {beat, flatten}, {beat, throb, pulse}, {beat, flog, punish}, {beat, shape, do metalwork}, {beat, baffle}, {beat, stir, whisk} and so on. Each of these synsets reflects a different sense of the verb – the ‘change’ sense, the ‘pulsate’ sense, and so on. Similarly, the polysemous noun board belongs to the synsets {board, plank} and {board, committee}. There is also the sense found in collocations like full board and room and board, for which a suitable synonym is not available; in cases like these the gloss of the meaning is used to identify the intended sense: {board, (a person’s meals, provided regularly for money)}.
WordNet’s focus on synonyms means that it is just as much a sophisticated thesaurus as a dictionary. In many ways, it privileges a thesaurus style, relational representation of semantic information over a dictionary like, definitional one. Originally, WordNet’s designers believed that the synsets on their own would be adequate to identify the different senses of terms, and that there would be no need to supply glosses. ‘For example,’ explains Miller (Miller et al. 1990: 240), ‘someone who knows that board can signify either a piece of lumber or a group of people assembled for some purpose will be able to pick out the intended sense with no more help than plank or committee. The synonym sets, {board, plank} and {board, committee} can serve as unambiguous designators of these two meanings of board.’ But it soon became obvious that glosses were needed to separate different senses of words, and to identify the meanings involved. That itself, of course, is an interesting result. The amount of synonymy in the lexicon is simply not adequate to differentiate the total number of words’ senses, since many senses exist for which there simply are no synonyms. The entry for the noun key is a good example of this situation. Synsets 1, 2, 3, 5, 6, 7 and 10 have no other members than key itself, and the sense is only identified by a gloss (in round brackets).
1. {key (metal device shaped in such a way that when it is inserted into the appropriate lock the lock’s mechanism can be rotated)}
2. {key (something crucial for explaining)} ‘the key to development is economic integration’
3. {key (pitch of the voice)} ‘he spoke in a low key’
4. {key, tonality (any of 24 major or minor diatonic scales that provide the tonal framework for a piece of music)}
5. {key (a list of answers to a test)} ‘some students had stolen the key to the final exam’
6. {key (a list of words or phrases that explain symbols or abbreviations)}
7. {key (a generic term for any device whose possession entitles the holder to a means of access)} ‘a safe-deposit box usually requires two keys to open it’
8. {winder, key (mechanical device used to wind another device that is driven by a spring (as a clock))}
9. {keystone, key, headstone (the central building block at the top of an arch or vault)}
10. {key (a lever that actuates a mechanism when depressed)}
This is an interesting finding about the extent of synonymy in English. English is known to be a language rich in synonyms, but even so it frequently turns out that words have numerous senses for which no synonyms exist.
Hyponymy/taxonomy is another crucial relation for the organization of WordNet, which displays the entire hyponymic/taxonomic hierarchy for every noun. For example, the nouns table and furniture would be linked by a hyponym–superordinate relation: the WordNet entry for table would include a pointer that labelled furniture as the superordinate term, called the hypernym (Miller 1998: xvi; in Chapter 5 we call it the hyperonym). These hyponymy relations can be followed in both directions. Thus, the following tree of synsets contains furniture and table:
Several of the hypernyms, such as {instrumentality, instrumentation} and {whole, unit} are probably not typical of natural discourse. Nevertheless, their inclusion in the hierarchy is claimed to reflect the conceptual structure of the vocabulary.
The fact that every noun is linked to its hyponyms and superordinates creates what is normally known in computer science as an inheritance hierarchy. Standard dictionary entries rely on readers’ general intelligence to supplement definitions with necessary information that is only implicit. For instance, a standard definition of tree – say ‘a large, woody, perennial plant with a distinct trunk’ – does not say anything about trees having roots, or consisting of cells having cellulose walls, or even being living organisms: the reader is expected to assume all of that information by virtue of the fact that trees are plants. But the dictionary itself doesn’t make this connection explicit – and, in fact, a reader needs to know which sense of plant is the relevant one to the definition of tree: the bio logical one, or the one that means ‘place where a product is manufactured’. All of this implicit information ‘is omitted on the assumption that the reader is not an idiot, a Martian, or a computer’ (Miller 1990: 246). In an inheritance hierarchy, however, each term ‘inherits’ the information associated with its hypernyms. In WordNet, the user can immediately see the inheritance hierarchy of the term in question, thereby gaining access to all the relevant definitional information. Here, for example, is the inheritance hierarchy for French horn, in which each synset is the hyponym of the one above it:
{musical instrument, instrument (any of various devices or contrivances that can be used to produce musical tones or sounds)}
2. {wind instrument, wind (a musical instrument in which the sound is produced by an enclosed column of air that is moved by the breath)}
3. {brass, brass instrument (a wind instrument that consists of a brass tube (usually of variable length) that is blown by means of a cup shaped or funnel-shaped mouthpiece)}
4. {French horn, horn (a brass musical instrument consisting of a conical tube that is coiled into a spiral and played by means of valves)}
By viewing the complete inheritance hierarchy all at once, the WordNet user has access to far more definitional information than would be available for the single term French horn alone. The user also has immediate access to a range of other information about the definiendum: for instance, simply by clicking on the ‘sister term’ link, the user gains access to a list of other hyponyms of brass instrument, such as trombone, bugle and flugelhorn, which occupy the same level of the hyponymic tree as French horn. As Miller (1990: 247) puts it, ‘lexicographers make an effort to cover all of the factual information about the meanings of each word, but the organization of the conventional dictionary into discrete, alphabetized entries and the economic pressure to minimize redundancy make the reassembly of this scattered information a formidable chore’. The electronic possibilities of an electronic database greatly reduce the difficulty of the task.
This is not the only advantage of the explicit labelling of the relations between words in WordNet: another is that the user is given great freedom of movement. ‘Unlike in a thesaurus, the relations between concepts and words in WordNet are made explicit and labeled; users select the relation that guides them from one concept to the next and choose the direction of their navigation in conceptual space’ (Fellbaum 1998: 8). Finally, not the smallest advantage of an inheritance hierarchy is that it saves space: since each piece of information percolates down the hierarchy, it does not need to be reduplicated at several different points.
So far, we have mainly considered the representation of nouns in WordNet. Other parts of speech pose different representational problems. Here we only have space to briefly consider verbs. There are fewer genuine synonyms for verbs than for nouns in English (Fellbaum 1990: 280), a situation which reduces the usefulness of the synset as a means of identifying verb meaning. An inheritance hierarchy arrangement also organizes the representation of verbs. The verbal equivalent of hyponymy/taxonomy is the relation of troponymy. A verb x is the troponym of a second verb y if to x is to y in some particular way. Thus, stroll is a troponym of walk, gobble a troponym of eat, and punch a troponym of hit. Verb hierarchies are much shallower than noun hierarchies, typically showing no more than four levels (Fellbaum 1990: 287). Furthermore, antonymy is another significant structuring principle.
QUESTION Explore WordNet’s representation of verbs further by starting with the entry for walk. What are the differences between the information supplied in the walk (n.) entry and the walk (v.) entry?
We will end this discussion by considering some of the problems of WordNet. First, WordNet does not have a principled means for determining the number of polysemous senses of a word. As a result, its distinction between different senses can often seem arbitrary. For instance, consider the WordNet entry for trouble:
1. {trouble, problem (a source of difficulty)} ‘one trouble after another delayed the job’; ‘what’s the problem?’
2. {fuss, trouble, bother, hassle (an angry disturbance)} ‘he didn’t want to make a fuss’; ‘they had labor trouble’; ‘a spot of bother’
3. {trouble (an event causing distress or pain)} ‘what is the trouble?’; ‘heart trouble’
4. {trouble, difficulty (an effort that is inconvenient)} ‘I went to a lot of trouble’; ‘he won without any trouble’; ‘had difficulty walking’; ‘finished the test only with great difficulty’
5. {worry, trouble (a strong feeling of anxiety)} ‘his worry over the prospect of being fired’; ‘it is not work but worry that kills’; ‘he wanted to die and end his troubles’
6. {trouble (an unwanted pregnancy)} ‘he got several girls in trouble’
This arrangement of senses could be accused of masking the fundamental similarity between these different senses. Indeed, an argument could be mounted that trouble only has a single sense in English.
QUESTION Can you formulate a definition of trouble that accounts for all of the sentences quoted in the WordNet entry?
Relatedly, the format of WordNet entries obscures the fundamental difference between homonymy and polysemy. The user who consults the entry for bank, for instance, finds no indication that the two senses ‘financial institution’ and ‘sloping land, especially by water’ are semantically unrelated. These two meanings simply appear in the list of senses, in exactly the same way as the ‘problem’ and ‘unwanted pregnancy’ senses of trouble, between which there is a clear relation. Most theories of lexical representation would assume, however, that there is a fundamental dis tinction here that needs to be recognized, expressed by saying that trouble exemplifies polysemy, and bank homonymy.
Both these problems can be summarized by questioning the extent to which the principles underlying WordNet live up to the psychological realism for which its creators were aiming. WordNet is motivated by the patterning assumption, but there has been very little attempt to reflect the results of psycholinguistic studies of meaning. Indeed, the restriction of the number of semantic relations which WordNet recognizes constitutes a serious limitation. There is no formal way in WordNet to recognize contextual relations between words like ball, racquet and net. The fact that these all belong to a single semantic domain – that of ball-games – is presumably a fact of some psychological importance, but WordNet has no way of representing it, since the relation does not correspond to any of the semantic relations which the database recognizes. This would seem to significantly compromise the seriousness with which WordNet can be viewed as a complete model of human lexical memory. Murphy (2003: 113) sums this criticism up as follows: ‘while WordNet has been claimed to be a psycholinguistic model, its architects have used psycholinguistic evidence selectively and its architecture limits it severely’.
