This repo offers several python (3.x) modules for grammatical analysis:

1. Extracting grammar patterns from sentences. For example, the grammar pattern for "discuss" in the sentence "He likes to discuss the issues ." would be "V n".
2. Aligning grammar patterns from parallel sentences. For example, grammatically erroneous source sentence "He likes to discuss about the issues ." → grammatically correct target sentence "He likes to discuss the issues", the aligned grammar pattern for "discuss" would be "V about n" → "V n".

We currently support grammar patterns for verb, noun and adjective headwords. See what grammar pattern is in Wikipedia.

Before starting to use modules, please install the python dependencies (mainly spaCy and NLTK):

$ pip install -r requirements.txt

$ python -m spacy download en_core_web_lg 

You can simply run test.py to check if we miss some required modules or data.

$ python test.py

Here we demonstrate how to test our shallow parser, extract grammar patterns for a sentence or align grammar patterns for parallel sentences.

Run an existing constituency tree parser to get linearized constituency tree string for every sentence as a pre-processing step. The constituency tree parser we use is AllenNLP. They have also an online demo.

from modules.shallow_parser import shallow_parse
from modules.grampat import sent_to_pats, align_parallel_pats

# source sentence: "He liked to discuss about the issues ."
# target sentence: "He likes to discuss the issues ."
# Note that we parse sentences in advance using AllenNLP's constituency tree parser.

src_parsed = shallow_parse("(S (NP (PRP He)) (VP (VBD liked) (S (VP (TO to) (VP (VB discuss) (PP (IN about) (NP (DT the) (NNS issues))))))) (. .))")
tgt_parsed = shallow_parse("(S (NP (PRP He)) (VP (VBZ likes) (S (VP (TO to) (VP (VB discuss) (NP (DT the) (NNS issues)))))) (. .))")

print(src_parsed)

[[['He'], ['liked'], ['to'], ['discuss'], ['about'], ['the', 'issues'], ['.']],
 [['he'], ['like'], ['to'], ['discuss'], ['about'], ['the', 'issue'], ['.']],
 [['PRP'], ['VBD'], ['TO'], ['VB'], ['IN'], ['DT', 'NNS'], ['.']],
 [['H-NP'], ['H-VP'], ['H-VP'], ['H-VP'], ['H-PP'], ['I-NP', 'H-NP'], ['O']]]

print(tgt_parsed)

[[['He'], ['likes'], ['to'], ['discuss'], ['the', 'issues'], ['.']],
 [['he'], ['like'], ['to'], ['discuss'], ['the', 'issue'], ['.']],
 [['PRP'], ['VBZ'], ['TO'], ['VB'], ['DT', 'NNS'], ['.']],
 [['H-NP'], ['H-VP'], ['H-VP'], ['H-VP'], ['I-NP', 'H-NP'], ['O']]]

shallow_parse() returns a list of chunked elements:

• Original words
• Base form of original words (lemmas)
• POS tag from constituency tree string
• Chunk tags

Note that the prefix HIO of chunk tags represents:

• H: Headword of a chunk. This is the headword of a grammar pattern we're interested in. We simply select the last word of a chunk as our headword.
• I: Non-headword of a chunk.
• O: Outside of a chunk. This is often a punctuation word and not important in our case.

src_pats = sent_to_pats(src_parsed)
tgt_pats = sent_to_pats(tgt_parsed)

print(src_pats)

[('LIKE', 'V to v', 'liked to discuss', (1, 3)),
 ('DISCUSS', 'V about n', 'discuss about the issues', (3, 5))]

print(tgt_pats)

[('LIKE', 'V to v', 'likes to discuss', (1, 3)),
 ('DISCUSS', 'V n', 'discuss the issues', (3, 4))]

sent_to_pats() returns a list of tuples, each tuple contains:

• Headword
• Grammar pattern (POS tag in uppercase corresponds to the headword).
• N-gram that matches grammar pattern
• Start and end positions of n-gram in chunked sentence.

How does sent_to_pats() works:

• Generate a list of n-grams of parsed results.
• For every n-gram, identify if hand-selected grammar patterns (listed in grampat.py) exist in an n-gram.
• The grammar patterns are selected from Collins COBUILD Grammar Patterns I: Verb and Grammar Patterns II: Nouns and Adjectives in advance, which are annotated from experts. We believe those grammar patterns are generally good and able to cover most grammar patterns we used in English.
• Note that it is possible to automatically find good grammar patterns from large monolingual corpora by counting frequencies of various n-grams of POS tag, and select good n-grams of POS tag by frequency. We can roughly interpret grammar pattern as simplied n-gram of POS tag.

parallel_pats = align_parallel_pats(src_pats, tgt_pats)

print(parallel_pats)

[[('LIKE', 'V to v', 'liked to discuss', (1, 3)),
  ('LIKE', 'V to v', 'likes to discuss', (1, 3))],
 [('DISCUSS', 'V about n', 'discuss about the issues', (3, 5)),
  ('DISCUSS', 'V n', 'discuss the issues', (3, 4))]]

align_parallel_pats() returns a list of aligned grammar patterns.

Now that you've completed the Example Usages guide, we can use these modules to count grammar patterns for large English monolingual corpora (BNC) and parallel grammatical error correction corpora (EFCAMDAT, LANG-8, CLC-FCE). We released a python script for doing this (support multi-processing):

$ python compute_grampat.py \
-in_src_path data/src.tree.txt \
-in_tgt_path data/tgt.tree.txt \
-out_path data \
-out_prefix dataset_name \
-n_jobs 4 \
-batch_size 1024

The data structure of the output file data/dataset_name.grampat.dill is a Python Dictionary containing two keys:

• "count_dict" (3-nested dict):
  • key1: source grammar pattern (str)
  • key2: target grammar pattern (str)
  • key3: headword in uppercase (str)
  • value: count
  • Note: We also save the instances that source grammar pattern is same as target grammar pattern.
• "ngram_dict" (4-nested dict):
  • key1: source grammar pattern (str)
  • key2: target grammar pattern (str)
  • key3: headword in uppercase (str)
  • key4: (source ngram, target ngram) (tuple)
  • value: count

We released grammar pattern results for BNC, EFCAMDAT, LANG-8 and CLC-FCE. It can be used for grammatical analysis (See query_grampat.py for example usage).

If you find the repo helpful for your research, you can cite it with the following BibTeX:

@software{yi_chen_howard_lo_2020_3611412,
  author       = {Yi-Chen Howard Lo},
  title        = {howardyclo/grammar-pattern},
  month        = jan,
  year         = 2020,
  publisher    = {Zenodo},
  version      = {v1.0.0},
  doi          = {10.5281/zenodo.3611412},
  url          = {https://doi.org/10.5281/zenodo.3611412}
}

or clicking this badge to export any format you like (on the right hand side of the website).