Description
Question 1: Lambda Calculus and Compositional Semantics (30 points)
a) Simplify the following lambda calculus expressions by applying beta-reduction as much as possible.
Show each step of the derivation.
• (λx.x x) (λy. y x) z
• (λuvw.wvu) aa (λpq.q)
• ((λv.v v) (λu.u)) ((λv.v)(λv.w))
b) Augment the grammar given in Lecture 14 to account for the quantifier no, including the corresponding
lexical rule and semantic attachment. Show a derivation of the sentence No student hates COMP-550.
(You’ll also need to add an entry for the verb hates.) Use explicit event variables, and be sure to show
the intermediate lambda expressions at each node of the parse tree.
c) Show how to construct an underspecified representation of the sentence No student wants an exam
using the Cooper storage scheme presented in class. Show how to recover both interpretations of the
sentence (and explain what the interpretations are).
Question 2: Lesk’s Algorithm (40 points)
Implement and apply Lesk’s algorithm to the publicly available data set of SemEval 2013 Shared Task
#12 (Navigli and Jurgens, 2013), using NLTK’s interface to WordNet v3.0 as your lexical resource.
(Be sure you are using WordNet v3.0!) The relevant files are available on the course website. Starter
code is also provided to help you load the data. More information on the data set can be found at
https://www.cs.york.ac.uk/semeval-2013/task12/.
The provided code will load all of the cases that you are to resolve, along with their sentential context.
Apply word tokenization and lemmatization (you have code to do this from A1) as necessary, and remove
stop words.
As a first step, compare the following two methods for WSD:
1. The most frequent sense baseline: this is the sense indicated as #1 in the synset according to
WordNet
1
2. NLTK’s implementation of Lesk’s algorithm (nltk.wsd.lesk)
Use accuracy as the evaluation measure. There is sometimes more than one correct sense annotated in
the key. If that is the case, you may consider an automatic system correct if it resolves the word to any
one of those senses. What do you observe about the results?
Next, develop a third method that combines distributional information about the frequency of word
senses, and the standard Lesk’s algorithm. Make and justify decisions about any other parameters to
the algorithm, such as what exactly to include in the sense and context representations, how to compute
overlap, and how to trade off the distributional and the Lesk signal, with the use of the development
set, which the starter code will load for you. You may use any heuristic, probabilistic model, or other
statistical method that we have discussed in class in order to combine these two sources of information.
It is beyond the scope of the assignment to use external corpora or lexical resources (e.g., thesauri,
or WordNet in other languages, etc.), so do not do so. Given these constraints, feel free to use your
creativity to find ways to improve performance!
Some issues and points to watch out for:
• The gold standard key presents solutions using lemma sense keys, which are distinct from the
synset numbers that we have seen in class. You will need to convert between them to perform the
evaluation. This webpage https://wordnet.princeton.edu/man/senseidx.5WN.html explains
what lemma sense keys are.
• The data set contains multi-word phrases, which should be resolved as one entity (e.g., latin america).
Make sure that you are converting between underscores and spaces correctly, and check that you
are dealing with upper- vs lower-case appropriately.
• We are using instances with id beginning with d001 as the dev set, and the remaining cases as the
test set, for simplicity. This is different from the setting in the original SemEval evaluation, so the
results are not directly comparable.
Discuss the results of your experiments with the three models. Also include a discussion of the successes
and difficulties faced by the models. Include sample output, some analysis, and suggestions for improvements. The entire report, including the description of your model, must be no longer than two pages.
Going beyond this length will result in a deduction of marks.
Your grade will depend on whether you adequately followed the guidelines above, whether you followed
standard model design and experimental procedure during the development of your method, and on the
quality of your report (both linguistic, and content).
Question 3: Reading Assignment — Compositional Distributional Semantics (30 points)
Read the following paper:
Jeff Mitchell and Mirella Lapata. Vector-based Models of Semantic Composition ACL 2008. http:
//aclweb.org/anthology/P/P08/P08-1028.pdf
Write a max. one-page (c. 500 words) discussion on this paper, including the following points:
1. A brief summary of the contents of the paper, including the theoretical framework and the experiments.
2. An evaluation and synthesis of what you learned in the paper. What are the advantages and
limitations of this work? How does it relate to the topics that we have discussed in class on
unsupervised learning and frame semantics?
3. Three questions related to the paper. These can be clarification questions, or questions about
potential extensions of the paper, or its relationship to other work.
Page 2
What To Submit
Electronically: Submit a .pdf containing the written answers to Question 1 and 3 as well as the report
part of Question 2, called ‘a3-written.pdf’. For the programming part of Question 2, you should submit one
zip file called ‘a3-q2.zip’ with your source code to MyCourses under Assignment 3.
Page 3
