Skip to content

Meaning Beyond Truth Conditions: Evaluating Discourse Level Understanding via Anaphora Accessibility

Conference: ACL 2025
arXiv: 2502.14119
Code: None
Area: NLP Understanding
Keywords: Discourse Semantics, Anaphora Accessibility, Dynamic Semantics, Quantifier Scope, LLM Evaluation

TL;DR

This paper proposes a three-level hierarchical system of natural language understanding capabilities (lexical/sentential/discourse), utilizing anaphora accessibility as a diagnostic task for discourse-level understanding. Through an evaluation dataset inspired by dynamic semantics, it systematically investigates LLMs' discourse understanding capabilities under three linguistic structures: universal quantifiers, negation, and disjunction.

Background & Motivation

  1. Background: Existing NLU evaluations primarily focus on the lexical level (word sense disambiguation, analogical reasoning) and sentential level (NLI, entailment judgment), while assessments of discourse-level semantic understanding remain far from adequate.
  2. Limitations of Prior Work: The few existing discourse-level evaluation works either target simplistic scenarios (such as entity tracking) or only consider the scope of negation, lacking a systematic investigation into the interaction of various logical connectives (e.g., universal quantifiers, conditionals, disjunctions) with discourse referents (entities).
  3. Key Challenge: Discourse understanding requires not only sentential truth-conditional semantics but also the capability to dynamically update discourse states; however, it remains unknown whether LLMs truly possess this structure-based capability for state updates.
  4. Goal: Systematically evaluate whether LLMs understand how the scope of various semantic operators affects the accessibility (referentiality) of discourse referents.
  5. Key Insight: Utilize precise predictions regarding anaphora accessibility from formal semantics (specifically, dynamic semantics) to design minimal-pair evaluation stimuli.
  6. Core Idea: If LLMs possess discourse-level understanding, they should assign higher probabilities to licensing continuations—i.e., allowing anaphoric reference after existential quantifiers, but rendering it inaccessible after universal quantifiers.

Method

Overall Architecture

Three sets of experiments are designed to test the effects of universal quantifiers (every/if/whenever), negation (single/double negation), and disjunction (either...or/and) on the accessibility of discourse referents. Surprisal (negative log-probability) is used to measure LLMs' acceptability of continuations, which is then compared with human experiments using a forced-choice paradigm.

Key Designs

  1. Universal Quantifiers Experiment:

    • Function: Test whether LLMs distinguish the discourse referent accessibility between existential and universal quantifiers
    • Mechanism: Compare pronoun resolution following "A farmer worked..." (which should be accessible) vs. "Every farmer worked..." (which should be inaccessible). A difference-of-difference metric is employed to compare whether the probability difference between inter-sentential and intra-sentential continuation conditions differs under existential vs. universal environments, thereby factoring out simple lexical biases.
    • Design Motivation: Merely comparing absolute probabilities is too weak of a test—even if an LLM prefers the existential condition, it could be due to lexical preferences rather than genuine discourse understating.

  2. Negation/Double Negation Experiment:

    • Function: Test whether LLMs understand that negation blocks accessibility while double negation restores it
    • Mechanism: Compare the probabilities of continuations under three conditions: "The farmer owned a cow" (accessible) vs. "didn't own a cow" (inaccessible) vs. "It was not the case that...didn't own" (double negation = accessible). Furthermore, "in fact" is introduced to probe the influence of lexical cues.
    • Design Motivation: Double negation elimination is a sophisticated semantic inference—two negations canceling each other out should restore anaphoric accessibility. This serves as a litmus test for genuine structural understanding vs. surface lexical pattern matching.

  3. Disjunction Experiment:

    • Function: Test Evans' observation regarding the special behavior of negative quantifiers in disjunctions
    • Mechanism: Compare "Either there was no manuscript, or it was hidden" (licensing/valid) vs. "Either there was a manuscript, or it was hidden" (unlicensing/invalid). SLOR (Syntactic Log-Odds Ratio) scores are used to normalize the effects of sentence length and word frequency.
    • Design Motivation: Anaphora accessibility in disjunctions is one of the most subtle predictions in formal semantics, providing a rigorous test for LLMs' discourse understanding.

Loss & Training

This paper is an evaluation study with no training. It evaluates the Llama 3 series (1B/3B/8B/8B-Instruct) and GPT babbage-002/davinci-002. For the human baseline, 104 participants were recruited on the Prolific platform to conduct a forced-choice experiment.

Key Experimental Results

Main Results

Experiment LLM Performance Human Performance Description
exi > every ~75% Close to 100% LLMs succeed but underperform humans
exi > if ~100% ~70% LLMs outperform humans (telescoping effect)
exi > neg Above chance Above chance Both succeed
DN > neg Reversal in some models Above chance LLMs struggle with double negation
EitherOr > And ~100% Above chance LLMs succeed
or > EitherPosOr Reversal! No distinct preference LLMs over-rely on the lexical cue "either"

Ablation Study (Adding "in fact")

Configuration Effect Description
DN>Neg + "in fact" Accuracy increases The co-occurrence pattern of "in fact" and double negation helps LLMs
Exi>Neg + "in fact" Direction reversal! "in fact" is less likely to appear after existential sentences, lowering the probability
Human + "in fact" No change Humans are not distracted by lexical cues

Key Findings

  • LLMs perform well on simple existential/universal contrasts, but their success partially relies on lexical cues rather than structural understanding.
  • Double negation is a weak spot for LLMs—most models fail to correctly handle double negation elimination.
  • Key finding in disjunctions: LLMs rely excessively on the word "either"—exhibiting a preference when "either" is present and none when it is absent, even when the sentences are semantically equivalent.
  • Humans and LLMs exhibit opposite behaviors in conditionals containing "he"-continuations (vs. "it"-continuations), likely due to the human tendency for telescoping.

Highlights & Insights

  • The three-level semantic understanding hierarchical system (lexical \(\to\) sentential \(\to\) discourse) is a valuable conceptual framework that provides systematic guidance for evaluating NLU.
  • Ingenious experimental design: It utilizes precise predictions from formal semantics to design minimal pairs and incorporates human subject experiments, enabling a direct comparison between LLMs and humans.
  • The finding that LLMs rely on lexical cues rather than structural abstraction is highly significant, suggesting that current "understanding" of discourse by LLMs might be mostly surface pattern matching.

Limitations & Future Work

  • Evaluated only relatively small models from the Llama 3 and GPT series; SOTA models like GPT-4o could not be tested because their APIs do not support logprobs.
  • The evaluation construction is relatively simple and does not cover more complex linguistic structures (such as modal subordination).
  • Behavioral-level evaluation cannot reveal the internal mechanisms of the models, requiring mechanistic interpretability methods as a complement.
  • vs. Schuster & Linzen (2022): Tested only the effect of negation on discourse referents, whereas this study expands to universal quantifiers and disjunctions.
  • vs. Kim & Schuster (2023): Evaluated state tracking using simple scenarios (such as moving objects within boxes), whereas this work focuses on the semantic structures of natural language.
  • vs. Li et al. (2021): Probed entity tracking representations in internal states of transformers, whereas this work provides a complementary evaluation at the behavioral level.

Rating

  • Novelty: ⭐⭐⭐⭐ Systematically applies formal semantics predictions to LLM evaluation, offering a unique perspective.
  • Experimental Thoroughness: ⭐⭐⭐⭐ Human baseline experiments + multiple LLMs + various linguistic structures, though the range of models is somewhat limited.
  • Writing Quality: ⭐⭐⭐⭐⭐ Well-established theoretical background and a clear conceptual hierarchy.
  • Value: ⭐⭐⭐⭐ Highlights the limitations of LLMs' discourse understanding, providing valuable insights for NLU research.