📄 100 Things You Always Wanted to Know about Semantics & Pragmatics But Were Afraid to Ask

Summary

This is a one-page tutorial abstract from ACL 2018, not a research paper with experiments. Bender argues that NLP cannot move past narrow, task-specific systems unless it draws on what linguistics knows about how people encode and recover communicative intent. The tutorial aims to give practitioners with little or no linguistics training a curated set of useful facts about semantics and pragmatics. It separates two kinds of meaning. One is meaning carried by the linguistic signal itself and its link to syntax, called sentence meaning. The other is meaning arising from situated use, called speaker meaning. The material is organised into six topics: an introduction to meaning, lexical semantics, compositional phrase semantics including multi-word expressions, discourse and reference, presupposition and implicature, and semantic resources. Examples are mostly English, but cross-linguistic comparison is used to show which phenomena vary between languages and which never appear in English. There is no data, method, or quantitative result. The value is a structured map of the meaning-related phenomena an NLP system should account for.

Key points

• Format: ACL 2018 tutorial abstract (Melbourne, July 15–20, 2018), one page. It contains no experiments or evaluation numbers.
• The six topics are:
  • introduction (speaker vs. sentence meaning)
  • lexical semantics (formal representations, word relations, meaning change)
  • phrase semantics and multi-word expressions
  • meaning beyond the sentence (discourse links, coreference)
  • presupposition and implicature
  • resources for semantic processing
• The target audience is NLP practitioners with minimal or no prior linguistics background.
• English is the main source of examples, with a typological angle added where possible.
• The author is linked to the Grammar Matrix customization system, the DELPH-IN Consortium (HPSG-based deep processing), and NAACL, which she chaired in 2016–2017.

Interesting ideas

• I1. Keep what the words literally encode apart from what the speaker intends in context, because tasks differ in which of the two they actually need.
• I2. Build phrase meaning compositionally from the parts, but treat multi-word expressions as possibly only partially compositional rather than all-or-nothing.
• I3. Presuppositions come from identifiable trigger expressions and behave in specific ways when embedded inside larger structures, so they can be tracked systematically.
• I4. Formal lexical representations should be compact yet still support tractable inference, and word meanings should be modelled as related to one another.
• I5. A cross-linguistic view shows which semantic phenomena are English-specific and which are absent from English altogether.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (frames): Have crate prag produce a literal frame layer from the UD tree and a separate speaker-meaning layer for speech acts, implicature and irony, each with its own explanation trace.
• I2 (frames): Add an MWE lexicon that can override or modify frame composition for specific subtrees while still composing the literal parts where they contribute.
• I3 (world): Record triggered presuppositions as background facts in the world of the text, noting whether they survive embedding, so answers can rely on them explicitly.
• I4 (world): Use this as a design check for the compiled ontology from WordNet/FrameNet, preferring relations that support explainable inference steps over large opaque feature sets.
• I5 (back): Test whether the meta-world used for translation and generation silently bakes in English-only distinctions before relying on it as a language-neutral pivot.