Semantics is concerned with the meaning of words and how they combine to form sentence meanings. It is useful to distinguish lexical semantics , and structural semantics --- the former is to do with the meanings of words, the latter to do with the meanings of phrases, including sentences. We will begin with the former.
There are many ways of thinking about and representing word meanings, but one that has proved useful in the field of machine translation involves associating words with semantic features which correspond to their sense components. For example, the words man, woman, boy, and girl might be represented as:
man = (+HUMAN, +MASCULINE and +ADULT) woman = (+HUMAN, -MASCULINE and +ADULT) boy = (+HUMAN, +MASCULINE and -ADULT) girl = (+HUMAN, -MASCULINE and -ADULT)
Associating words with semantic features is useful because some words impose semantic constraints on what other kinds of words they can occur with. For example, the verb eat demands that its AGENT (the eater) is animate and that its PATIENT (that which is eaten) is edible, --- concrete (rather than abstract, like sincerity, or beauty), and solid (rather than liquid, so one cannot `eat' beer, coffee, etc.; soup is a borderline case). We can encode this constraint in our grammar by associating the features HUMAN and EDIBLE with appropriate nouns in our dictionary and describing our entry for eat as something like cat=verb, AGENT=HUMAN, PATIENT=EDIBLE. The grammar will now only accept objects of eat that have the feature EDIBLE. Thus these selectional restrictions , as they are called, act as a filter on our grammar to rule out unwanted analyses. Consider sentence ( ):
The English word game is ambiguous - it can mean several things, including a form of play or sport or a wild animal hunted or fished for food. Using selectional restrictions of the sort described above we can eliminate the `form of play or sport' meaning if the system is able to infer that `food' is EDIBLE, but that forms of play are not.
Selectional restrictions have proved a very useful device and are found in most MT systems to a greater or lesser extent. Unfortunately, however, exceptions to selectional restrictions abound, especially in metaphorical speech . Thus we find sentences like This car eats money, used to mean that the car is expensive to maintain, so, rather than use selectional restrictions to eliminate interpretations, we should use them to state preferences between alternative interpretations.
Notice that stating selectional restrictions in terms of semantic relations is easier than trying to state them in terms of (surface) grammatical relations . Using grammatical relations we would have to say that eat prefers an animate SUBJECT in active sentences, and an animate NP in the by phrase in passive sentences (and an edible OBJECT in actives, and an edible SUBJECT in passives).
We will now look briefly at how semantic relations can help in one of the thorniest problems for machine translation, namely the translation of prepositions.
Take, for example, the translation of the English preposition at into Spanish , and, for the sake of exposition, make the simplifying assumption that it receives only two translations in Spanish , namely a and en, as in the following:
The choice of Spanish preposition depends on the type of noun that follows it. Roughly, where the preposition is followed by a temporal noun, as in the first example, it translates as a, but where the preposition is followed by a locational noun, as in the second example, it translates as en.
We can pick out the correct translation of at by assigning it an appropriate Semantic Relation (SR) during analysis. For example, the feature SR=TIME might be assigned to indicate that at expresses a temporal relation, and the feature SR=PLACE might be used to mean that at expresses a location relation. We could then have translation rules of the following form:
at, SR=TIME a at, SR=PLACE en
These semantic relations are assigned on the basis of the type of noun that follows the preposition. This means that the noun midday must be marked in the dictionary with some temporal feature (e.g. semtype=time), while nouns like school must be marked with some locational feature (e.g. semtype=location).
We are assuming that semantic relations attach to prepositions. More properly, a semantic relation describes the role which the whole prepositional phrase, not just the preposition, plays in relation to its head, but it is convenient to allow the preposition to carry this feature too, in order to formulate the above translation rules. A prepositional phrase marked with the semantic relation TIME, for example, might indicate the time at which the action indicated by the verb takes place, while a phrase marked with the semantic relation PLACE might indicate the location at which it took place.
Although these features would solve many problems in translating prepositions, the semantic relations expressed by PLACE and TIME are not always fine grained enough. We can, for example, distinguish two different types of usage for locational at: `(to be) at school' indicates a position, whereas `(to shoot) at the goal' indicates a movement towards a certain place. We could decompose the semantic relation into two separate relations, say PLACE_POSITION for the first phrase, and PLACE_PATH for the second phrase. Note that the calculation of these new semantic relations will depend not only on the semantic features of the nouns that follow them, but crucially on the type of verb.
Our brief example illustrates some of the problems we face when trying to assign semantic relations to prepositional phrases, or other categories. First, it is difficult to know what a canonical set of semantic relations might look like, since the refinement or granularity required (that is, the number of distinctions we want to make) depends to some extent on the type of translation problem encountered. Secondly, the finer the granularity, the more elaborate the feature system will have to be, in order to differentiate nouns, for example. Finally, the calculation of semantic relations depends on a number of factors, including as we have seen the type of verb and the type of the following noun.
We have described semantic features as more or less optional additions to representations --- the addition of a semantic feature may serve to disambiguate a representation, by indicating which sense of a word is involved, but the representation is still conceived of as a structure consisting of lexical items (words). A more radical idea is to take the semantic features as exhausting the meaning of words, and to replace the lexical items by the appropriate set of features. Thus, one would have representations with (+HUMAN, +MASCULINE, +ADULT, ...) in place of the lexical item man. The idea is that the meanings of lexical items can be decomposed into sets of semantic primitive s. Since such sets of semantic primitive s might well be universal, one can in this way approach the goal of an interlingua . Here one cannot manage satisfactorily simply with sets of features, however. Instead, one needs to produce structures in which the predicates are semantic primitive s. For example, the representation of kill might be along the following lines:
As we have already noted in Chapter there are some doubts in general about the feasibility and advisability of this process of lexical decomposition . For example, there is a small but significant difference in meaning between kill and cause to become not alive --- in particular, where a `killing' is a single event, a `causing to become not alive' involves at least two events (a `causing', and a `dying'), and if the causal chain that links a particular event to dying is long enough, one may admit that the event caused the dying, but not want to say there has been a `killing'. Of course, these doubts depend on what one thinks the relation is between the semantic primitive s like CAUSE, BECOME, etc., and English words like cause, become, etc., and also on the assumption that there is no semantic primitive KILL. Notice that, while a collection of semantic primitives that includes KILL is going to be quite large (perhaps in the order of a thousand primitives), this is still far less than the vocabulary one finds in normal use --- so there may still be some value in semantic decomposition , even if the number of primitive s that words decompose into is quite large.
So far we have concentrated our discussion of semantics on the meaning of words, but semantics is also concerned with linguistic `systems' such as tense and aspect and determination, all of which are of considerable importance in translation. Consider the problem of how to translate the present tense in French into English, where there are at least three possibilities, exemplified in the following:
Of course, one could try to formulate rules which describe the conditions under which French present tense is realized as English present, English present perfect, or present progressive, but such rules would be very complex. A more attractive possibility is to try to find some more abstract representation which directly describes the temporal and aspectual relations that these sentences involve. Here we will outline one type of approach.
The English tense system is used to convey two different types of information. One is the time of the event --- both the present simple I sing and the present progressive I am singing describe an event in the present. The other is the nature of the event --- e.g. the progressive stresses that the event is `in progress'. Henceforth we shall reserve the word tense to mean the time of an event and use the word aspect to refer to the way the event is viewed (as an on-going or completed process, a state, or a simple event, etc.). We will use the term time reference to cover both tense and aspect.
We can think of tense as expressing a relation between the time of the event and the time of speech. Thus, with the present ( I sing), the time of the event (which we could call E) overlaps with the time of speech (which we could call S). Contrast the future ( I shall sing) where the time of the event follows the time of speech ( E follows S), or the past, where E precedes S. However, this is not sufficient to distinguish all the different temporal forms of the English verb. There is a problem with the past, where our definition of tense does not allow us to differentiate between the simple past ( I sang) and the pluperfect (or past-perfect --- I had sung), since in both cases the time of the event is prior to the time of speech. One solution is to define an additional point of time, called the reference time (R). Consider, for example, the sentence:
At two o'clock specifies a moment in time which precedes the time of speech, but which is not the time of event. Two o'clock is not the time at which John ate, but the time by which he had already eaten. The temporal relations of this sentence can be expressed as follows, where < means `precedes':
E < R, R < S
This indicates that the time of the event ( E) precedes the reference time ( R), and R precedes the time of speech ( S).
We can now distinguish the pluperfect from the simple past by stipulating that in both cases the time of the event precedes the time of speech ( E < S), but while in the pluperfect the time of the event precedes the reference time (E < R), in the simple past the time of event and the reference time coincide ( E = R).
We can do something similar to distinguish the present perfect ( ) from the other tenses. Here too the event described precedes the speech time, but there is a sense in which sentences in the present perfect are `about' the present (for example, ( ) would be appropriate only if Sam's previous eating habits are still of current relevance). We can capture this by making reference time and speech time coincide ( R=S).
This gives the following picture:
We now have the apparatus to represent the difference in tense and aspect between the examples above. Of course, having a way of representing tense and aspect values as above is one thing, calculating the representations for particular inputs is another. This is no trivial task, since the tense and aspect values of the verb will in general depend on many factors, including the form of the verb, and whether it is modified by any time adverbials such as yesterday and tomorrow.
However, let us assume that we have calculated the tense and aspect values of the following sentence, and see how this helps translation.
This sentence might receive a semantic representation along the lines of Figure . The feature time-ref encodes the information about tense and aspect, in particular, the fact that the reference time coincides with the time of speech, and the event time precedes the reference time (and hence also the time of speech).
Figure: Representation Indicating Time Values after French Analysis
Since the information encoded by the time-ref feature is presumed to be preserved in translation, this feature can treated as an interlingual feature, and thus can be mapped unchanged onto the target language (in this case English), giving the representation in Figure .
Figure: Representation after Transfer but before English Synthesis
The verb form has lived can then be generated from this representation by English synthesis, giving the translation ( ). Other time-ref values would be realized differently --- in principle, the correct translations of the examples above can be obtained.
This treatment of tense and aspect involves a lot of complicated machinery, and is not entirely unproblematic. Nevertheless it gives some indication of how one might attempt to handle the difficult problem of tense and aspect in MT.