Ling/CSE 472: Assignment 6: English Resource Semantics

Background

In this assignment, you will be reviewing the documentation (which is still under development) of the English Resource Grammar's semantic representations, comparing the 'fingerprints' proposed there to analyses produced by the current version of the grammar, and using the 'fingerprints' to search through the Redwoods treebank for further examples of specific phenomena.

Learning outcomes. After completing this assignment, you should be able to:

• Read the semantic representations produced by a broad-coverage grammar of English (at least in part)
• Relate parts of semantic representations to substrings of sentences
• Describe ways in which a linguistically-motivated computer-readable grammar makes semantic details explicit

Resources:

• Index of semantic phenomena documentation pages
• Live grammar demo (use "ERG 2018 (UW)")
• Treebank search interface

Tasks

Part 1

1. Choose two phenomena from the list here, both of which have reasonably complete documentation (sample sentences, semantic fingerprints, some textual description).
   • Off-limits (no credit!): Apposition (we will use this to demo)
2. Announce which phenomena you're doing on the Canvas discussion for Assignment 6. It's okay if multiple people do the same phenomena, but we're hoping that people will pick different ones at least to a certain extent.
3. Using the grammar demo and specifically "ERG 2018 (UW)", check whether the examples on the inventory page match the fingerprints specified. Specifically:
   • Check at least three examples, if possible one from 'ESD Test Suite Examples', one that is a positive example of the phenomenon from 'Motivating examples', and one that is a negative or contrasting example from 'Motivating examples'
   • Make sure you have MRS checked in the interface and look to see if the MRS for the top parse matches (or appropriately fails to match) the fingerprints. It may help to also use the DMRS to see which part you want to focus on.
   • If this is an example where the MRS should match, but the MRS for top parse doesn't, do any of the top five?

Part 2

1. Choose one phenomenon from this shorter list:
   • Compounding
   • Implicit Locatives
   • Implicit Nominals
   • Instrumental Relatives
   • Measure Phrases
   • Nominalization
   • Non-Adverbial Clausal Modifiers
   • Number Sequences
   • Parentheticals
   • Partitives
   • Quasi-Modal Infinitives
   • Ellipsis
   • Identity Copulae
   • Vocatives
2. Use the WeSearch search interface to the Redwoods treebank (WSJ portion) to find two examples in the treebank that match the fingerprints to your phenomenon.
3. For each example, identify the part of the string that gives rise to the part of the MRS the fingerprints matched.
4. For each example, and using the grammar demo, construct one simpler example based on the ones you found that are much shorter but still display the same fingerprints.

Write up

Your write up should include the following:

Part 1

1. Which phenomenon you picked
2. Which sentences you tested
3. For each sentence that should be a match (sentences from 'testsuite examples' and positive 'motivating examples'), if an MRS among the first five produced by the grammar matched the fingerprints, indicate which part of it matched. (You can show this by typing up the relevant predicate names, argument positions, and variables.) If no MRS among the top five matches, find the one that comes closest and describe how it differs.
4. For any negative (contrasting) 'motivating examples' you tried, did they correctly fail to match?
5. Describe a use-case for this part of a semantic representation; that is, for some NLP application, how would it help to have this semantic information made explicit?
6. In the style of reading questions, react to the prose description of the phenomena and its semantic representation. What was confusing? What does this reading make you ask next?
7. Extra credit: Can you find an example that shows that the fingerprints are too general, i.e. match sentences that don't illustrate the phenomenon? If so, provide the example sentence and a description of the part of its MRS that matches. (Note: We don't actually know whether any of the fingerprints are too general!)

Part 2

1. Which phenomenon you picked
2. The two example sentences you found
3. The substrings of the example sentences that give rise to the part of the MRS that matches the fingerprints
4. The simplified sentences that still match and their respective substrings
5. What use case can you imagine, in linguistic research or otherwise, for his kind of search interface?

General

1. In 2-3 sentences, reflect on what you learned from this assignment about the size and complexity of grammatical systems in general and of the ERG in particular.
2. In 2-3 sentences, reflect on the extent to which these analyses seem portable across languages. Would you expect to find similar phenomena to the ones you investigated in any other language? In all other languages? How might they differ?
3. May we share your write up with the other authors of the ERG Semantic Documentation pages? (NB: Either answer is fine! We expect these write ups to be very helpful, but also don't want to share without express permission.)