serhii.net

In the middle of the desert you can say anything you want

03 Dec 2023

Masterarbeit eval task LMentry-static-UA

Context: 220120-1959 taskwarrior renaming work tasks from previous work

First notes

Just tested this: DAMN!

2023-12-03-174516_813x723_scrot.png 2023-12-03-175001_773x692_scrot.png

User
Can you, in English, name one word for each of these tasks:
1. Rhymes with "chair"
2. Is a number larger than eleven
3. Has two letters "a"
4. Ends with the letter "k"
5. In the sentence "the cat had four paws and a good mood" is BEFORE the word "paws"

Also:
6. A sentence that starts with the word "dogs"
7. A sentence that ends with the word "beaver"
8. A sentence that uses the word "metal" twice

https://chat.openai.com/share/3fdfaf05-5c13-44eb-b73f-d66f33b73c59

lmentry/data/all_words_from_category.json at main · aviaefrat/lmentry

Not all of it needs code and regexes! lmentry/data/bigger_number.json at main · aviaefrat/lmentry

I can really do a small lite-lite subset containing only tasks that are evaluatable as a dataset.

LMentry-micro-UA

// minimal, micro, pico

Plan:

  • go methodically through all of those task, divide them into regex and not regex, clone the code translate the prompts generate the dataset

Decision on 231010-1003 Masterarbeit Tagebuch#LMentry-micro-UA: doing a smaller version works!

LMentry-static-UA

Basics

Will contain only a subset of tasks, the ones not needing regex. They are surprisingly many.

The code will generate a json dataset for all tasks.

Implementation

Original task/code/paper analysis

Distractors

  • order of words
  • template content
  • adj tasks VS arg. content

My changes

  • I’d like to have words separated by:
    • frequency
    • length,
    • … and maybe do cool analyses based on that
  • (DONE) in addition to first/last letter/word in word/sentence, add arbitrary “what’s the fourth letter in the word ‘word’?”
  • Longer/shorter words: add same length as option

My code bits

  • I have to write it in a way that I can analyze it for stability wrt morphology etc. later

Ukrainian numerals creation

Problem: ‘1’ -> один/перший/(на) першому (місці)/першою

Existing solutions:

Created my own! TODO document

TODO https://webpen.com.ua/pages/Morphology_and_spelling/numerals_declination.html

More tagsets fun

Parse(word='перша', tag=OpencorporaTag('ADJF,compb femn,nomn'), normal_form='перший', score=1.0, methods_stack=((DictionaryAnalyzer(), 'перша', 76, 9),))

compb

Nothing in docu, found it only in the Ukr dict converter tagsets mapping: LT2OpenCorpora/lt2opencorpora/mapping.csv at master · dchaplinsky/LT2OpenCorpora

I assume it should get converted to comp but doesn’t - yet another future bug report to pymorphy4

Even more tagsets fun

pymorphy2 doesn’t add the sing tag for Ukrainian singular words. Then any inflection that deals with number fails.

Same issue I had in 231024-1704 Master thesis task CBT

Found a way around it:

@staticmethod
def _add_sing_to_parse(parse: Parse) -> Parse:
	"""
	pymorphy sometimes doesn't add singular for ukrainian
	(and fails when needs to inflect it to plural etc.)

	this creates a new Parse with that added.
	"""
	if parse.tag.number is not None:
		return parse

	new_tag_str = str(parse.tag)
	new_tag_str+=",sing"
	new_tag = parse._morph.TagClass(tag=new_tag_str)
	new_best_parse = Parse(word=parse.word, tag=new_tag, normal_form=parse.normal_form, score=parse.score, methods_stack=parse.methods_stack) 
	new_best_parse._morph=parse._morph
	return new_best_parse


# Not needed for LMentry, but I'll need it for CBT anyway...
@staticmethod
def _make_agree_with_number(parse: Parse, n: int)->Parse:
	grams = parse.tag.numeral_agreement_grammemes(n)
	new_parse = Numbers._inflect(parse=parse, new_grammemes=grams)
	return new_parse
  • parse._morph is the Morph.. instance, without one added inflections of that Parse fail.
  • TagClass follows the recommendations of the docu2 that say better it than a new OpencorporaTag, even though both return the same class.

Notes by task

Taxonomy

2023-12-17-161653_1515x951_scrot.png +

Comparing two things

Words of different lengths, alphabet order of words, etc.

Main relationship is kind=less|more, where less means “word closer to beginning of the alphabet”, “smaller number”, “word with fewer letters” etc., more is the opposite.

Alphabet order of words

  • (DONE) which word is closer to beginning of alphabet
  • Are these words in alphabet order?

Which word is longer

TODO Which number is bigger

  • use the one-million bits and add to the text that this is why I needed to care about agreemnet
  • do comparisons of entities! one box has a million pencils, the other has five hundred thousand. Which has more pencils?
GPT4 agreement issues

https://chat.openai.com/share/b52baed7-5d56-4823-af3e-75a4ea8d5b8c: 1.5 errors, but I’m not sure myself about the fourth one.

    LIST = [
        "Яке слово стоїть ближче до початку алфавіту: '{t1}' чи '{t2}'?",
        "Що є далі в алфавіті: '{t1}' чи '{t2}'?",
        "Між '{t1}' та '{t2}', яке слово розташоване ближче до кінця алфавіту?",
        # TODO - в алфавіті?
        "У порівнянні '{t1}' і '{t2}', яке слово знаходиться ближче до A в алфавіті?",
        # ChatGPT used wrong відмінок внизу:
        #  "Визначте, яке з цих слів '{t1}' або '{t2}' знаходиться далі по алфавіті?",
    ]

HF Dataset

I want a ds with multiple configs.

Base patterns

starts = "(starts|begins)"

base_patterns = [
rf"The first letter is {answer}",
rf"The first letter {of} {word} is {answer}",
rf"{answer} is the first letter {of} {word}",
rf"{word} {starts} with {answer}",
rf"The letter that {word} {starts} with is {answer}",
rf"{answer} is the starting letter {of} {word}",
rf"{word}: {answer}",
rf"First letter: {answer}",
]

For more: lmentry/lmentry/scorers/more_letters_scorer.py at main · aviaefrat/lmentry

Looking for example sentences

  • spacy example sentences
  • political ones from UP!
  • implemented
  • for words, I really should use some normal dictionary.

Assoc. words and resources

Another dictionary I found: slavkaa/ukraine_dictionary: Словник слів українською (слова, словоформи, синтаксичні данні, літературні джерела)

  • Excel_word_v10.xslx 2024-02-08-035111_1377x790_scrot.png
  • sql as well
  • a lot of columns

Next tasks

  • List
    • Most associated word
    • Least associated word
    • Any words from category
    • All words from category

All basically need words and their categories. E.g. Animals: dog/cat/racoon

I wonder how many different categories I’d need

Ah, the O.G. benchmark has 5 categories: lmentry/resources/nouns-by-category.json at main · aviaefrat/lmentry

Anyway - I can find no easy dictionary about this.

options:

Wordnet

for all-in-one:

> grep -o "_\(.*\)(" all-in-one-file.txt | sort | uniq -c
     49 _action(
      8 _action-and-condition(
     58 _holonym(
    177 _hyponym(
     43 _meronym(
     12 _related(
     51 _sister(
    102 _synonym(

looking through it it’s sadly prolly too small

2009’s hyponym.txt is nice and much more easy to parse.

GPT

Ideas: WordNet Search - 3.1 Ask it to give me a list of:

  • emotions
  • professions
  • sciences
  • body parts
  • animals
  • times (dow, months, evening, etc.)
  • sports It suggests also
  • musical instruments
  • dishes
  • clothing

  1. <_(@bm_lmentry) “LMentry: A language model benchmark of elementary language tasks” (2022) / Avia Efrat, Or Honovich, Omer Levy: z / / 10.48550/ARXIV.2211.02069 _> ↩︎

  2. API Reference (auto-generated) — Морфологический анализатор pymorphy2 ↩︎

Nel mezzo del deserto posso dire tutto quello che voglio.
comments powered by Disqus