[Transcriber: I apologize for just getting a rather 'partial' picture of what every one said in this session. This is a paraphrased gist of the original conversation based on my understanding. Please feel free to modify if you find anything different from what you actually meant. ]

GCS: When I worked on my Thai grammar, I made a tool to increase the grammar coverage by putting in lexicon. But Monday's discussion suggests that we can use controlled vocabulary for testsuits, which allows the handling of lots of phenomena without worrying about other words.

FCB: MRS testsuites are designed like this for handling specific phenomena. Every time you make a change to the grammar, you need to run through the testsuites to ensure no other things are broken.

GCS: I tried with a toy grammar and brought in lots of lexicons using various resources. The efforts seem to be futile in increasing the grammar coverage.

FCB: Zhong and Indra started by building in syntax based on testsuites, but they can only reach medium size without a extensive lexicon.

DPF: MRS is designed to test semantics, but a lot of them are related to syntax things. For example, "I make Mary do something", it's dependent on the word "make". I have 1300 sentences in my syntactic testsuites. They helped to catch a lot of my mistakes. I wish I had constantly increased that testsuite along the way. Testsuites are critical as something suprising might happen when phenomena interact with each other. And it's tricky to answer how many words should be in your testsuite.

GCS: Big test suite may distract you from your analysis.

DPF: Yes. You might be tempted to fix bugs you find immediately. In CSLI testsuite, we control the vocab, but also try to make sentences sound as natural as possible.

GCS: Maybe I should systemize a procedure to incorporate things into testsuites.

DPF: One way is to have one sentence illustrating one interesting phenomenon. The competing goal is to put as many phenomna into a sentence as possible, since regression test is expensive. But that requires careful debugging when that sentence (containing multiple phenomena) fails.

FCB: What about our effort on using lexical database to add example and comment to lexical types?

GCS: We may create a tree for testing. Each leaf is a sentence with one phenomenon, and the nodes are sentences with multiple phenomena, created from leaf or nodes below. Then we can just run test at the root.

FCB: For a sentence with 8 phenomena, we can run test on its sub sentences.

XPG: It would be hard to make a sentence with just one phenomenon.

GCS: There is a useful book "Dictionaries: The Art and Craft of Lexicography", by Sidney I. Landau.

WDP: My suspicion is that it might be hard to keep the lenth of the sentences low when you go up the tree. It might be worth doing. You can replay the derivation to blow through the 100k quickly.

DPF: There's a way to speed up. We can store the derivation and replay those, which requires much less time. If I have broken something, it will tell me, then I can focus on those sentences. But for negative sentences, you don't have derivations to replay. You need to manually curate your testsuite. You may make sure the negatives don't parse, or be aware that they parse due to other reasons and you are fine with them.

GCS: We need to avoid a grammar that allows everything. That's why we need negative examples.

DPF: Testing is to see if things are reasonably healthy or there are obvious problems. Sometimes we also need to see charts to see what edges go in there. Root conditions save you from embarassment, but not from unnecessary computation.

GCS: Can the mechanism be made to check the failure of the negative examples?

DPF: Maybe we can store the previously happened failure (generated ones), which have been fixed later and therefore disappeared. If they appear again, something is wrong.

FCB: Question about the CSLI testsuite: how often do you treebank them when you are frequently changing the grammar?

DPF: Everytime. At least once a day.

GCS: With traditional packing, you know the number of derivations. We can check that number, too.

FCB: I think we need to do more regression test. We used to do that regularly but not much treebanking.

DPF: It's still useful to test even without treebanking.

FCB: The small test suites need to be checked regularly. For larger test sets, we need more vocabulary. I want to discuss different approaches of entering lexical entries. When Dan adds a new word into ERG, he makes sure all its lexical types are entered. Whereas since we wanted to get more coverage quickly, we found it useful to have offline unknown word handling by importing from resources. Then we go in to fill in the gaps, finding a missing lexical type, then bring in more words of that type. What's your current practice?

DPF: If you have a rich resource, you can use the second approach by all means. We didn't have that in early 1980s. It's still not an option for English (due to copy right issues). We then built our lexicon by hand. The other thing is, I put in the ones I'm interested in by hand. Then I let the tagger help me in dealing with the rest. As soon as one lexical entry exists for that stem, I block the generic entries. It might be better for you to allow the tagger to give more information about a stem (as additional to its existing lexical entry), as this might be real ambiguity.

GCS: You could fall back on the tagger based on its confidence value.

DPF: A coarse tagger can still give useful information for an incomplete entry, which can be good enough for parsing, with slight over-generation. You can be a little permissive in the properties of the generic entries.

GCS: I use a hybrid approach to get interesting words and their concordance, then eyeball to see interesting patterns.

DPF: I doubt it would be helpful to have a predictor of fine-grained subcategory of lexical types.

FCB: How is generation for unknown words?

DPF: we made it work in 2014.

WDP: There are two categoreis of unknowns. One type is names with CARG as unknown. ACE can generate this. The other type is real unknown words with PRED containing an "unknown". ACE can't generate that. CARG and PRED values are treated differently. We need a solution that will work across grammars.

DPF: Maybe we should stop packing things into predicate names, so unknown word mechanism can drop the orthography into the right spot.

WDP: You'll have to handle the semantic indexing for unknown words.

DPF: We don't have a way to say surface string is "sought", which is the irregular past tense for "seek". We need morphological analyzing module to put that in predicate.

FCB: An external tagger can give lemma, then we pass this information to the system.

DPF: There may be ambiguity like "revered" as "rever-ed" or "revere-d". This is orthographic ambiguity. May be Francis side can come out with a specification of what things should be.

FCB: Your view will be different because of the language you are working on. Irregular things can be listed. Let's assume unknown words are regular, then we can rely on normal morphology.

GCS: Delphin has modularized pipeline.

DPF: It might not be easy to list all irregular things. The tail is long.

FCB: If we do this, we'll get to the words that are unknown and irregular and rare.

WDP: We have general morphological rules, in most cases. The morphology make guesses. We are computing that morphology already. It seems silly to throw them away. We can make this information accessible to the generic entries. The issue is we create PRED names through regex, which happens before morphology.

FCB: Great. With the information from both the tagger and the mophology, the problem can be solved.

DPF: We want to take orthography of stem as a first class object instead of using regex. It requires radical change of codes. I originally want unknowns to have fancy predicate name, then postprocessing can fix things.

FCB: For a langguage that has external morph analyzer, we can slightly change the interpretation of YY mode to give the STEM for inflection rules.

DPF: What do you do with multi-word proper names?

FCB: It's quite inconsistent. We may have multi tokens.

GCS: I tried to extract Thai names. I can take some surrounding context and glue them (the tokenized ones) together. There are different word-breaking philosophies. One is maximum agressive breaking, to make the tokens as small as possible. I take that approach. The other is to favor longer token. I have multiple hypothesis going into chart mapping.

WDP: Uber-tagging could be useful for that. It does this without pre-committing what tokenization is.

FCB: Both character-based and word-based tagging have been used before. And it (character-based) did lead to exponential chart building.

CapitolHillLexicon (last edited 2017-01-10 22:50:00 by ZhenzhenFan)

(The DELPH-IN infrastructure is hosted at the University of Oslo)