Skip to content


FrancisBond edited this page Aug 11, 2017 · 4 revisions

A discussion of using delph-in tools for langauge learning, plus a brief intro to the new work Sanghoun is doing.

Discussion at the 2017 Grammar Engineering Meeting led by SanghounSong, scribed by FrancisBond.

3 topics (SSH)

  • using the ERG to help people
  • mal-rules
  • how to learn mal-rules from leaner's corpora
    • from parallel text, from corpora
  • how to group by learner type
    • model for different learner types (e.g. Korean Native vs Chinese Native learning English)
    • do we have enough data?

LMC: there is a nice learner corpus, Lang-8 Learner Corpora, (many pairs of L1 and L2) crawled from the lang8 website

SSH has 320,000 sentences written by Korean Uni Students

  • focusing on verbs
  • some metadata per student (grade, age, gender, experience)

Sanghoun has parsed them all with the ERG and has a nice tool showing the results. ERG 1214; <80%

All sentences are then annotated, correcting if necessary, adding information if necessary. People can add comments and add some set annotation (irrelevant, ...). There are some issues with the low English level of the annotators.

They are trying to document the learner language as though it were a new language.

WDP: similar to trying to learn the long tail of wierd sentences.

LMC: if all you are doing is trying to identify that there is an error, then it may be easier. Hard to learn rules if you need to reconstruct the MRS.

FCB: can we say this is typically learner vs this is good prose? By e.g. comparing n-grams?

WDP+MWG: could lead to changing meaning if you compare too much.

SSH: coverage is higher than expected, because sentences are short vocabulary size is small, ... In the errors, typos are common.

It would be good to try to find typos and parse them: unknown word handling is still a problem -- it guessed noun to often (WDP: and many typos create good words, especially for short (2 character) words).

LMC: We are trying to identify errors, without necessarily correcting them for our tool.

SSH: can I use the educ/ subdirectory in the ERG.

DPF: you should!

There was some discussion of vocabulary restriction --- in general Dan and SH allow people to type, so need an open vocabulary and some unknown word handling.

SSH: is Zhong useful for this? (parsing learners' text) ZZ+FCB: yes

SSH: I would like to do this for Korean.

LMC: can we learn in the Miyao style (from a corpus)

FCB: Not really, there is a lot of work in annotating the corpus and writing the conversion rules.

WDP: ace can now run a pcfg for ungrammatical sentences (c-saw) which can be a bit slower but is very robust

ALL: show us how to use it and train our own models, please.

???: when you ran the robust parser, what constraints were broken? learn from there! You need the unification failure results

FCB: I think Sanghoun should first run the mal-rule enhanced grammar and see if we need to learn many more rules before trying to learn rules automatically.


  • use educ /e-duck/ first. Try the trunk version
  • try using c-saw
    • --- needs some documentation and a model (WDP)

SSH: can we used type-diff

ALL: it is a bit difficult if one does not parse, but we could try to compare results from csaw with typediff

PCFGs for ERG-1214:

command-line option

See AceCsaw for more information.