N-gram scoring for English translation selection

[Copilot]

LTL-to-English translations were using random.choice() among candidate phrasings, producing inconsistent and sometimes awkward output. For example, F(n → Gz) could produce "Eventually, 'n' implies It is always the case that 'z'" with incorrect capitalization in nested contexts.

Changes

• New ngram_scorer.py module: Scores candidate translations using bigram/trigram frequencies for common English temporal patterns. Selects deterministically based on naturalness.

• ltltoeng.py updates:

  • select_from_patterns() replaces random.choice() with n-gram-based selection
  • uncapitalize_sentence() and get_nested_english() for proper case handling in nested translations

• ltlnode.py updates: All node classes now use get_nested_english() for operand translations, ensuring nested content is lowercase.

Example

# Before: random, inconsistent capitalization
F(n -> G z) => "Eventually, 'n' implies It is always the case that 'z'"

# After: deterministic, proper nesting
F(n -> G z) => "Eventually, 'n' implies it is always the case that 'z'"

# N-gram scorer prefers natural patterns
G p         => "It is always the case that 'p'"  # highest scored
p -> q      => "If 'p', then 'q'"                 # beats "Whenever 'p', then 'q'"
p <-> q     => "'p' if and only if 'q'"           # common phrase scores high

Tests

• 24 new tests covering n-gram scoring logic and translation consistency
• All 76 tests pass

Original prompt

---

This section details on the original issue you should resolve

<issue_title>N-Gram approach for better English translation</issue_title>
<issue_description>Currently our English to LTL translation is lacking. For instance, the
formula F(n → Gz), translates to “Eventually, globally, z holds is necessary for n holds”. We readily acknowledge that this recursively-generated phrasing is both confusing and unnatural.

---

I wonder if (given the small amount of randomness we already have in generating English) we could generate a few candidate English translations and then use a reasonable metric (maybe n-grams? smoothing?) to choose the best / most natural one?
I think relatively little work has been done in the world of LTL to English. Issues include :

• Compositionality gap: English is not fully compositional for temporal semantics
• English handles time through implicit quantifiers, whereas LTL requires explicit ones.
• Human language uses event schemas and narrative structure; LTL uses positions on a trace.
• English negation is structure-sensitive.
  </issue_description>

Comments on the Issue (you are @copilot in this section)

• Fixes N-Gram approach for better English translation #106

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.