We're Afraid Language Models Aren't Modeling Ambiguity

Introduction

• Present AmbiEnt (Ambiguity in Entailment)
  • English benchmark of 1,645 examples
  • Coverts a variety of lexical, syntactic, and pragmatic ambiguities, and sentences which can be plausibly read as conveying one of multiple different messages
• Characterising ambiguity requires a choice of meaning representation to distinguish between possible interpretations (enumerating can be tricky or impractical)
  • Adopt a functional approach
  • Characterise ambiguity in the premise and/or hypothesis by its effect on entailment relations
• Design a suite of tests based on AmbiEnt to investigate the extent to which understanding of ambiguity is acquired during pretraining of LLMs
  • Investigate whether LMs can be finetuned on existing NLI data for the less demanding task of ambiguity recognition (w/o explicit disambiguation)

AmbiEnt

Curated Examples

• Handwritten or sourced from existing NLI datasets and linguistics textbooks
  • DistNLI, ImpPres, NLI Diagnostics, MNLI, WaNLI

Generated Examples

• Automatically identify groups of premise-hypothesis pairs that share a reasoning pattern
• Use WaNLI as source of examples
  • Each group contains
    • A randomly chosen example on which its two annotators disagreed
    • Its 4 nearest neighbours (according to the final-layer embedding of a WaNLI-trained NLI model)
  • Format each group into a prompt with instruction, sample 5 continuations from InstructGPT

Annotation and Validation

• Examples from InstructGPT consist of unlabelled premise-hypothesis pairs
• Annotation phase
  • 37 university-level linguistics students
  • Select a set of labels for each examples
    • Provide a disambiguating rewrite for each one if more than one label is chosen
  • Discard offensive or low-quality examples
• Validation phasse
  • Performed by subset of the authors
  • Review two sets of annotations to revise
  • Aggregate them into a single coherent annotation

Agreement

• Four validators annotate a subset of 50 examples in common to calculate inter-annotator agreement

Analysis

• Analysis to understand how annotators behave on ambiguous input (under the traditional 3-way annotation scheme for NLI)

Setup

• Crowdworkers review ambiguous examples in AmbiEnt
• Task is split into three steps
  1. Annotation of ambiguous example
     • Follows the traditional NLI labelling setup
  2. Recognition of disambiguities
     • The ambiguous sentence of the example is given for consideration
     • Given 3 candidate interpretations (2 disambiguations and 1 distractor), indicate whether each sentence is a possible interpretation of the ambiguous sentence
  3. Annotation of disambiguated examples
     • 3 new NLI examples are obtained by substituting the ambiguous sentnce of the original excample with each candidate interpretation from step 2
     • Select a single NLI label for each new example

Evaluating Pretrained Language Models

• Evaluate if LMs can:
  1. Directly generate relevant disambiguations
  2. Recognise the validity of plausible interpretations
  3. Model open-ended continuations reflecting different interpretations
• Ambiguous instances from AmbiEnt

Generating Disambiguities

• Whether LMs can learn in-context to directly generate disambiguations and corresponding labels
• Perform both automatic and human evaluation
  • Automatic: EDIT-F1 metric from AmbigQA
  • Human: Same setup as the previous crowdworker experiment (w/o step 1)

Recognising Disambiguities

• Focus on the ambiguous sentences alone
• Model prediction: token with the greater logit between True and False
• Executed zero-shot

Modelling Interpretation-Specific Continuations

• Whether LMs implicitly model different interpretations in their distributions of text continuations
• Obtain continuations for each interpretation, and quantify how surprised the LM is to see them
  • Sample 100 continuations $c\sim p(\cdot |d_i)$ conditioned on each disambiguation $d_i$ as context
  • Compare the likelihood of $c$ under ambiguous sentence $a$ vs. the corresponding disambiguation $d_i$:
  $$\log p(c|d_i)-\log p(c|a)$$
  • This is an unbiased estimate of the KL divergence between $p(\cdot |d_i)$ and $p(\cdot |a)$
• Expect the LM to model continuations from disambiguations $d_i$ better than those from the distractor $\tilde{d}$: $$D_{KL}\left(p(\cdot |\tilde{d})||p(\cdot |a)\right)>D_{KL}\left(p(\cdot |d_i)||p(\cdot |a)\right)$$
  • KL ranking accuracy: The fraction for which this is true

Results

References

• Liu, A., Wu, Z., Michael, J., Suhr, A., West, P., Koller, A., Swayamdipta, S., Smith, N. A., & Choi, Y. (2023). We’re Afraid Language Models Aren’t Modeling Ambiguity. arXiv (Cornell University). https://doi.org/10.48550/arxiv.2304.14399