From inside the subdirectory of your home directory where you keep all your code for the NLG course, type the following command into a terminal window:
svn checkout http://opennlg13.googlecode.com/svn/trunk/ass-1 ass-1
This should create a subdirectory called ass-1 containing five XML files and a copy of these instructions. Move into this subdirectory and list them to make sure.
Before starting the assignment questions listed below, you should spend a while familiarising yourself with the OpenCCG grammar in the XML files. In this assignment, you will be gradually extending this grammar, step-by-step, so as to generate a wider range of grammatical sentences about restaurants.
There is a testbed.xml file included with the grammar. YOU WILL NOT NEED TO EDIT THIS FILE IN THE COURSE OF THIS ASSIGNMENT. The testbed contains around 50 tests. At the moment the grammar passes at least the first 13 of these (using ccg-test). Your job is to extend the grammar so that it passes ALL tests by the time you are finished. If you take a look at the testbed file, you will see that the subsequent tests are divided into sections, one for each of the questions in the assignment, so you can easily keep track of your progress and success.
N.B. If your entire grammar is working correctly, all tests in the testbed will have "ok" in the Parse column. "FAILED" means that the test failed, not that the parse failed. For instance, we don't want "Giovanni's serves" to parse, so we have set numOfParses to "0".
<item numOfParses="0" string="Giovanni's serves"/>
In this cases, the output from ccg-test will be
Parse Realize String
----- ------- ------
ok - *Giovanni's serves
Hint: When you start tccg, if you have made a mistake there may be useful error messages, which appear between "Loading grammar from URL: ..." and "Grammar ass-1 loaded". If you get stuck, try consulting the Rough Guide to OpenCCG.
In some questions, we have specified a name for a category, or specific semantics for you to aim for in your answers. You may disagree with our choices, but that's ok, because in OpenCCG there are often many ways to do the same thing. Ours may not even be the best way, but please follow the instructions where specified, and don't spend too much time worrying about alternative category names or semantics.
PLEASE READ THROUGH ALL THE QUESTIONS BEFORE STARTING.
And last but not least, remember that we're only interested in your work, not anyone else's submitted under your name.
Currently, the grammar is capable of generating sentences with conjoined subject NPs, for example "Giovanni's and Dario's serve cheap Italian food". Please extend the grammar so that it can generate sentences involving two conjoined sentences, for example "Giovanni's rocks and Dario's serves cheap food". You should do this by adding a second entry element to the lexical family conjunction.
Your grammar should assign the following semantic representation to the sentence "Giovanni's rocks and Dario's serves food", where each sentence denotes a separate conjunct (you might want to draw the graph out on paper to make sure you understand what is going on):
@w2(and ^
<CONJ>(w1 ^ rock ^
<ARG1>(w0 ^ Giovanni's)) ^
<CONJ>(w4 ^ serve ^
<ARG1>(w3 ^ Dario's) ^
<ARG2>(w5 ^ food)))
Please extend the grammar so that it can generate sentences involving two conjoined verb phrases, for example "Giovanni's rocks and serves cheap food". Make sure you get subject/verb agreement to work out properly - it shouldn't generate sentences like "*Giovanni's rocks and serve cheap food".
The semantics you assign to VP conjunction should be the same as sentential conjunction, but you need to ensure that the subject is "shared" between the two. Thus, the semantics of "Giovanni's rocks and serves food" should be:
@w2(and ^
<CONJ>(w1 ^ rock ^
<ARG1>w0) ^
<CONJ>(w3 ^ serve ^
<ARG1>(w0 ^ Giovanni's) ^
<ARG2>(w4 ^ food)))
We now want to be able to generate sentences using adverbs like "very" or "really", which modify (or "intensify") adjectives. Examples of such sentences are "Giovanni's serves really cheap food" or "Giovanni's and Dario's serve very good Italian food". We want a phrase like "really cheap" to have a semantic structure which makes it clear that the adverb modifies the adjective, like this:
@w1(good ^
<MOD>(v1 ^ very))
Add the words "very" and "really" to the morphology file. Call your new lexical family adverb.
Adjectives in English divide into two groups - "gradable" adjectives and "non-gradable" adjectives. Ideally, we want words like "very" and "really" only to intensify the gradable ones. For purposes of this assignment, assume that adjectives of nationality are non-gradable, and all the others are gradable.
Your job now is to modify your adverb lexical family to ensure that the grammar CANNOT generate noun phrases like "*very Italian food" or "*cheap very Indian food".
Define adjective macros @gradable and @non-gradable which introduce one of the following features to every atomic category of type 'A': (a) gradable: yes; or (b) gradable: no. Use the @singular and @non-singular macros on lexical nouns for inspiration. Then modify your adverb family to restrict the set of adjectives that "very" and "really" can combine with.
In English, some adjective orderings are acceptable whereas others are not. In our example domain, adjectives of nationality must always come after other adjectives. In other words, assuming the categories we have just defined, non-gradable adjectives must follow gradable ones. For example "cheap Italian food" is acceptable, while "*Italian cheap food" is not.
Your task in this question is to edit the OpenCCG grammar so that correct adjective ordering is preserved. You need to use the 'gradable' feature on atomic categories of type 'A' that you defined in question 4 to do this.
Here is a tip: In the rules.xml file you will find a typechanging rule with the name adjective, which converts 'predicative' adjecties into 'attributive' (i.e. noun-modifying) ones. You should replace this rule with two copies, one called adj-gradable (which converts gradable predicative adjectives into attributive ones), and the other called adj-non-gradable (which converts non-gradable predicative adjectives into attributive ones). Then edit these two rules so that it is only possible for adjectives to occur in the order specified above.
The lexicon as it stands at the moment contains three 'mass' nouns - 'food', 'music' and 'decor'. Recall that the main syntactic property of a mass noun is that it can form an NP without needing to be prefixed by an article or determiner (e.g. "Giovanni's serves food"). In fact, mass nouns CANNOT be prefixed with the indefinite article - "*Giovanni's serves a food". (There *are* other senses of 'food' which can be used with the indefinite article, but we ignore these in this exercise.)
Note that we have chosen to encode the concept of 'mass' noun in the current lexicon using the feature singular: no.
This question involves adding the noun "restaurant" to the lexicon. It is an example of a "count" noun rather than a "mass" noun. Count nouns CAN and indeed MUST occur with a determiner.
Add the word "restaurant", and give it the POS common-noun. Add the word "a" and give it the POS determiner.
Add a new lexical family for "determiner", and use features to ensure that a determiner can ONLY be used with a count noun. You will probably need to edit the adjective rules as well.
Hint: in order to get this to work, you may need to define the lexical family for 'determiner' as a 'closed' family. There are two possible reasons for this, both of which have to do with allowing the realiser to identify the relevant 'indexing' EP, as discussed in the lectures. In order for successful realisation to take place, the semantic representation for every lexical entry must contain exactly one EP of the form @x p, where p is a node label. In cases where the word is semantically vacuous, we therefore need to identify the lexical family as indexRel="*NoSem*" (this was discussed in the lectures). In other cases, we may want the indexing EP to be an edge label, so we need to identify the family as indexRel="...", where the value of the indexRel is the relevant edge label itself. Finally, note that every family with a specified indexRel MUST be 'closed', and have its members explicitly listed.
Add the plural count noun "restaurants" and the plural verb "are" to the lexicon. Make sure you can parse and generate sentences like "Giovanni's and Kalpna are restaurants". Make sure that subject-verb agreement works, so that you can't generate sentences like "*Giovanni's and Kalpna is a restaurant" or "*Giovanni's are restaurants".
This assignment is due in at 16.00 on Thursday 28 February. It should be submitted using the submit program.
If you have any technical problems with this assignment, please email amyi@inf.ed.ac.uk Don't ask me how to answer the questions :)