How can large language models become more human?

Daphne Wang, Mehrnoosh Sadrzadeh, Miloš Stanojević, Wing-Yee Chow, Richard BrehenyACL 2024

TL;DR

This paper proposes a novel framework that combines large language model (LLM) predictions with syntactic dependency structures to better model human language comprehension. This "Incompatibility Fraction" (IF) approach significantly outperforms traditional surprisal measures in predicting human reading times, especially for challenging "garden path" sentences.

What Problem Does This Paper Address?

Understanding how humans process language is a long-standing goal in psycholinguistics. A key aspect of this is how we predict upcoming words and syntactic structures. While computational models, particularly those leveraging LLMs, have shown promise, they often struggle with complex sentences, like "garden path" sentences, which can lead to processing difficulty for humans. Traditional measures like "surprisal" (how unexpected a word is) correlate with human reading times for naturalistic text but fail to accurately predict difficulty in these challenging cases. The paper aims to bridge this gap by developing a model that integrates both lexical predictions (from LLMs) and syntactic structure, more closely mirroring human predictive capabilities.

Methods

The core of the paper's contribution is a novel computational framework based on presheaves and sheaves. This mathematical framework allows for modeling data with hierarchical structures and relationships.

Here's a breakdown of their approach:

  • Topological Space as Base: They use a pre-ordered set of sub-phrases of a sentence, ordered by the prefix relation, to represent the incremental nature of reading.
  • Data as Probabilities: The data associated with each sub-phrase consists of probabilities of different syntactic structures for sentence completions.
  • LLM for Lexical Statistics: They use the GPT-2 model to generate completions of sub-phrases and their associated probabilities.
  • Dependency Parser for Syntax: The spaCy dependency parser is used to extract syntactic structures from these completions.
  • Incompatibility Fraction (IF): This is the key novel measure. It quantifies how "incompatible" the syntactic probability of a sub-phrase is with the probabilities of its completions. Essentially, it measures the mismatch between what the parser predicts syntactically and what the LLM predicts lexically.
  • Distance Measures: They employ established measures like Kullback-Leibler (KL) divergence, Jensen-Shannon (JS) divergence, and a variant of Earth Mover's distance to compute this incompatibility.

Datasets

The authors evaluated their framework using two datasets of "garden path" (GP) sentences, previously introduced by Pickering and Traxler (1998):

  • Dataset 1 (NP/S): Contains "easy" GP sentences with complement clause ambiguities.
  • Dataset 2 (NP/Z): Contains "hard" GP sentences with subordinate-clause ambiguities.

Each dataset includes four types of sentences:

  1. The original GP sentence.
  2. A disambiguated control sentence.
  3. A semantically implausible variant of the GP sentence.
  4. A disambiguated control for the implausible variant.

Key Insights and Ideas

  • Beyond Surprisal: The paper strongly suggests that relying solely on lexical surprisal (from LLMs) is insufficient for modeling human language comprehension, especially for complex syntactic structures.
  • Syntax-Lexicon Synergy: Combining syntactic information (from dependency parsing) with lexical predictions (from LLMs) is crucial for a more human-like model. The "Incompatibility Fraction" captures this interaction.
  • Predicting Processing Difficulty: The IF measures show a strong correlation with human reading times, outperforming surprisal in predicting difficulty, particularly for garden path sentences.
  • Distinguishing Sentence Types: The IF measures are effective at distinguishing between easy and hard garden path sentences, a capability that surprisal lacks.
  • Semantic Plausibility Challenges: While the IF measures correlate well with reading times, they struggle to fully capture the effects of semantic implausibility on processing difficulty. This suggests that explicitly modeling semantic structure is a necessary next step.

Key Results

  • Strong Correlation with Reading Times: All IF measures (IF-min, IF-JS, IF-KL) showed a high correlation with human first-pass and total reading times.
  • Outperforming Surprisal: The IF measures consistently outperformed surprisal in predicting garden path effects (GPEs). For example, IF-min achieved the best GPE with 39.47 ms and a low standard error of 0.17, while surprisal predicted very low GPEs.
  • Distinguishing Easy vs. Hard GP: IF-JS and IF-KL were particularly good at distinguishing between easy (NP/S) and hard (NP/Z) GP sentences, with lower errors than surprisal.
  • Semantic Implausibility: While IF measures predicted longer reading times for implausible sentences (aligning with human data), they did not predict the diminishment of GP effects in implausible sentences as hypothesized. Surprisal also failed to capture this nuance.

Limitations and Future Work

  • Semantic Structure: The authors acknowledge that their current model does not explicitly encode semantic structure, which is a limitation for fully understanding the impact of semantic implausibility on processing. They believe presheaves can be extended to address this.
  • Parser Dominance: In cases of semantic implausibility, the parser's predictions are overshadowed by the LLM's probabilities, leading to high incompatibility and surprisal.
  • Generalization: The current work focuses on garden path sentences. Future work will involve testing the framework on a wider range of naturally occurring sentences.
  • Dataset Representation: The plausibility element of the current datasets might not be fully representative, necessitating testing on different datasets.
  • Forward-Looking Model: The current framework is primarily forward-looking. Future research will explore incorporating backward-looking mechanisms for modeling repair and recovery in sentence processing.
  • Relationship with Surprisal: Formalizing the relationship between incompatibility and surprisal is an ongoing area of research.

Summary updated

BibTeX
@inproceedings{wang2024how,
  author = {Daphne Wang and Mehrnoosh Sadrzadeh and Miloš Stanojević and Wing-Yee Chow and Richard Breheny},
  title = {{How can large language models become more human?}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2024},
}

Loading related content...

Figure