Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

We're Afraid Language Models Aren't Modeling Ambiguity

Alisa Liu, Zhaofeng Wu, Julian Michael, Alane Suhr, Peter West, Alexander Koller, Swabha Swayamdipta, Noah A. Smith, Yejin Choi

TL;DR

AmbiEnt introduces the first linguist-annotated benchmark of linguistic ambiguity in entailment (1,645 examples) to evaluate pretrained LMs on recognizing and disentangling multiple readings. The paper combines curated and generated data with a rigorous annotation/validation pipeline, demonstrates that ambiguity explains much of NLI disagreement, and shows that current LMs struggle to generate or recognize disambiguations, even GPT-4, though multilabel NLI can flag ambiguous political claims. It provides a case study and discusses implications for real-world communication and NLP tooling. Overall, AmbiEnt highlights a critical gap in current models and offers concrete benchmarks and methods to advance ambiguity-sensitive NLP.

Abstract

Ambiguity is an intrinsic feature of natural language. Managing ambiguity is a key part of human language understanding, allowing us to anticipate misunderstanding as communicators and revise our interpretations as listeners. As language models (LMs) are increasingly employed as dialogue interfaces and writing aids, handling ambiguous language is critical to their success. We characterize ambiguity in a sentence by its effect on entailment relations with another sentence, and collect AmbiEnt, a linguist-annotated benchmark of 1,645 examples with diverse kinds of ambiguity. We design a suite of tests based on AmbiEnt, presenting the first evaluation of pretrained LMs to recognize ambiguity and disentangle possible meanings. We find that the task remains extremely challenging, including for GPT-4, whose generated disambiguations are considered correct only 32% of the time in human evaluation, compared to 90% for disambiguations in our dataset. Finally, to illustrate the value of ambiguity-sensitive tools, we show that a multilabel NLI model can flag political claims in the wild that are misleading due to ambiguity. We encourage the field to rediscover the importance of ambiguity for NLP.

We're Afraid Language Models Aren't Modeling Ambiguity

TL;DR

AmbiEnt introduces the first linguist-annotated benchmark of linguistic ambiguity in entailment (1,645 examples) to evaluate pretrained LMs on recognizing and disentangling multiple readings. The paper combines curated and generated data with a rigorous annotation/validation pipeline, demonstrates that ambiguity explains much of NLI disagreement, and shows that current LMs struggle to generate or recognize disambiguations, even GPT-4, though multilabel NLI can flag ambiguous political claims. It provides a case study and discusses implications for real-world communication and NLP tooling. Overall, AmbiEnt highlights a critical gap in current models and offers concrete benchmarks and methods to advance ambiguity-sensitive NLP.

Abstract

Ambiguity is an intrinsic feature of natural language. Managing ambiguity is a key part of human language understanding, allowing us to anticipate misunderstanding as communicators and revise our interpretations as listeners. As language models (LMs) are increasingly employed as dialogue interfaces and writing aids, handling ambiguous language is critical to their success. We characterize ambiguity in a sentence by its effect on entailment relations with another sentence, and collect AmbiEnt, a linguist-annotated benchmark of 1,645 examples with diverse kinds of ambiguity. We design a suite of tests based on AmbiEnt, presenting the first evaluation of pretrained LMs to recognize ambiguity and disentangle possible meanings. We find that the task remains extremely challenging, including for GPT-4, whose generated disambiguations are considered correct only 32% of the time in human evaluation, compared to 90% for disambiguations in our dataset. Finally, to illustrate the value of ambiguity-sensitive tools, we show that a multilabel NLI model can flag political claims in the wild that are misleading due to ambiguity. We encourage the field to rediscover the importance of ambiguity for NLP.
Paper Structure (58 sections, 6 equations, 3 figures, 11 tables)

This paper contains 58 sections, 6 equations, 3 figures, 11 tables.

Table of Contents

  1. Introduction
  2. AmbiEnt
  3. Curated Examples
  4. Generated Examples
  5. Annotation and Validation
  6. Agreement
  7. AmbiEnt Statistics
  8. Does Ambiguity Explain Disagreement?
  9. Setup
  10. (i) Annotation of ambiguous example
  11. (ii) Recognition of disambiguations
  12. (iii) Annotation of disambiguated examples
  13. Results
  14. Evaluating Pretrained Language Models
  15. Generating Disambiguations
  16. ...and 43 more sections

Figures (3)

  • Figure 1: Ambiguity can be the result of innocent miscommunication (top), or deliberately used to mislead one's listeners (bottom). For instance, a cat may be lost in the sense of being confused about its whereabouts (entailment edge), or lost in the sense that others cannot find it (neutral edge). Each example in AmbiEnt contains a set of labels corresponding to plausible readings, along with a disambiguating rewrite for each reading.
  • Figure 2: Pipeline for the annotation of generated examples in AmbiEnt. Unlabeled examples are created by InstructGPT, then annotated independently by two linguists, whose annotations are consolidated by an author.
  • Figure 3: Distribution of set labels in AmbiEnt.