Welcome to the first video lecture in this tutorial on Fluid Construc5on Grammar. Today’s lecture will introduce “feature structures” for represen5ng linguis5c informa5on, and “unifica5on” as the basic opera5on for processing that informa5on. If you are already familiar with these concepts, it is s5ll useful to go through the video because you will find that FCG implements them in a different way than many other grammar formalisms. 1 Morphology has many defini5ons, but the tradi5onal one is that it studies the smallest meaningful elements in language. These meaningful units are called “morphemes”. A word such as “books” is then said to be composed of two morphemes: “book” and the plural suffix “-­‐s”. “Book” is an unbounded morpheme because it can occur in a sentence without any other morphemes. The suffix, however, is a bounded morpheme because it can never occur alone. For the purposes of our tutorial, however, we’ll take a much broader look at morphology to include everything that is related to the FORM in which words appear in a sentence. That is, how can we decide which form a word should have when we produce uPerances, and vice versa, what kind of informa5on can we extract from the forms we observe in parsing? We’ll see that this may comprise phonological, seman5c and syntac5c constraints as well. 2 Here are some small examples of the different kinds of constraints that have an influence on the form in which words appear in the sentence. Construc5on Grammar is ideally suited for handling such phenomena because construc5ons can access all of these different levels simultaneously. 3 Why should we care about morphology? Can’t we just create lexical entries for each possible word form? You can get away with such a strategy if you only care about parsing, but as soon as you also want to produce acceptable uPerances, you cannot avoid morphology any longer. 4 When we introduce regular expressions later, you will see some intricate examples of morphology such as root-­‐and-­‐template morphology that is needed for Arabic. For now, however, we will introduce you to two ways of handling morphology. The first one is the holis5c approach. In this approach, you define a lexical construc5on without immediately adding a string to it. You then define different morphological construc5ons that are compa5ble with that lexical construc5on, one for each possible word form. The advantage of this approach is that complex word forms can be built and iden5fied in a single step, so it is easy to op5mize for large-­‐scale implementa5ons. However, the holis5c approach does not reflect the produc5ve paPerns of the language, so you will need to implement addi5onal procedures for handling unknown words (such as an instance-­‐based classifica5on). The second approach is the composi5onal approach where you define the system of construc5ons that, when combined with each other, will build or analyze a word at processing 5me. The good thing about this approach is that you reflect the produc5ve paPerns of the language so you can exploit the grammar for analyzing unknown words. On the other hand, you need more processing effort because each word needs to be composed from the ground up for every uPerance. Note, however, that the two approaches are fully compa5ble with each other and that you can thus combine their strengths. 5 6 Here is an overview of what we’ll do in produc5on and in parsing. In produc5on, a lexical construc5on will build a unit for a word. Other construc5on will then fill in more informa5on, and depending on the informa5on we get, a morphological construc5on will apply and give the word its surface form. In parsing, that same morphological construc5on first applies, a_er which we get the meaning from the lexical construc5on. 7 Let’s look at what morphology changes to our lexical construc5ons. Here you see the seman5c pole of a lexical construc5on for the word “book”. This should look familiar to you know: there is a meaning in the ROOT that we tag, and we repeat that tag in a J-­‐unit so that when the J-­‐unit builds a new unit for the word “book”, the meaning is encapsulated into that unit. But what do we do with the syntac5c pole? In produc5on, the lexical construc5on should build this unit. However, in parsing, the unit should already exist because a morphological construc5on should have already built it! 8 The solu5on is that we do not use the J-­‐operator for building this new unit. Remember that in produc5on, the syntac5c pole of a construc5on does not need to be matched against the transient structure, but that it will be merged. So by simply wri5ng the unit here as if it already existed will do the trick: if the unit is not in the transient structure yet, FCG will merge it in. We thus only use the J-­‐operator for adding a footprint to that unit in order to prevent the construc5on from applying again. Now take the perspec5ve of parsing: in parsing, we want to match the ?book-­‐
word unit with the transient structure, and that is exactly what will happen here. 9 We now turn to the morphological construc5ons. The special thing about morphological construc5ons is that they consist – at least in this example – of two syntac5c poles. In other words: they will only operate on the syntac5c pole of a transient structure. These construc5ons are applied from “le_ to right” in produc5on, which means that when I produce a sentence, the le_-­‐pole of the construc5on is matched against the syntac5c pole of the transient structure, a_er which the rest of the construc5on is merged. In parsing, we go the other direc5on. If you really insist on all construc5ons being pairings between meaning and form, you could say that the le_ pole of the morphological construc5on is the pole that carries the func5onal meaning of that construc5on, whereas the right pole consists of its form. On this slide you see the le_ pole of a morphological construc5on for the word form “book”. As you can see, it is almost an exact copy of the syntac5c pole of the lexical construc5on, but it is more specific because it states that the number of “book” is singular. 10 The right pole, then, solely deals with merging a string to that unit in produc5on. In parsing, it will match on the string “book”. 11 Here you see the defini5on of a morphological construc5on for “books”, which will only apply if the value of number is plural (produc5on) or if the string “books” is observed (parsing). 12 We can now try out our construc5ons by parsing and producing them. What do you get when you produce all solu5ons given this meaning? You get both “book” and “books” as a solu5on: since the value of the NUMBER feature is underspecified in the lexical construc5on, both morphological construc5ons are compa5ble. We therefore need an addi5onal construc5on to avoid this problem. 13 Here I define a construc5on for the plural. You can see that, even though this is a gramma5cal construc5on, it carries meaning just like lexical construc5ons. 14 So if I now parse “books”, I end up with the meaning predicates “book” and “more-­‐
than-­‐one”. If I produce that meaning, I will always correctly produce the word “books”. However, when I produce ‘((book x)), I s5ll end up with two solu5ons. How is that possible? 15 The answer is that the plural construc5on s5ll matches because the value of NUMBER is a variable. You can solve this problem in mul5ple ways, and it is up to you as a grammar developer to choose which one you prefer. One way is to define the singular-­‐construc5on. Another way is to always give priority to the singular-­‐
morpheme in processing. If it applies, it will prevent the other morphological construc5on from applying because they use the same footprints. You will learn later how to manage processing in more detail. 16 17 The holis5c approach to morphology is o_en difficult for people who are not used to it. The main source of error is the fact that the syntac5c pole of a lexical construc5on must match with the syntac5c pole of all of the morphological construc5ons associated with it. Beginning FCG users o_en forget or misspell a feature, which prevents the construc5ons from combining with each other. It also means that you will have a lot of code duplica5on if you spell out every construc5on in detail. Finally, as a result of all that, whenever you decide to make a change in the lexical construc5on, you have to manually update all of the morphological construc5ons as well. I therefore advise you to use templates! 18 For lexical construc5ons, we can just reuse the templates we saw in the previous lecture. If you do not specify a :string in def-­‐lex-­‐skeleton, the template will know that you are going to combine the lexical construc5on with a morphological one. 19 For morphological construc5ons you can use the def-­‐morph template, in which you specify a string and the syn-­‐cat features. You can define each morpheme separately, as you see here. However, as you can see, we s5ll have the problem of code-­‐
duplica5on and a poten5al source of errors. 20 You can therefore embed the def-­‐morph template in the template def-­‐morphs-­‐for. You pass the name of the lexical construc5on that these construc5ons should be compa5ble with, and then you only need to specify what’s different in the morphological construc5on. The other features are automa5cally copied from the lexical construc5on. 21