TL;DR
This paper tackles the challenge of generating natural-sounding descriptions for multiple objects (plural reference generation). They propose a new algorithm, IApart, that mimics human strategies of partitioning sets of objects and emphasizing similarity, outperforming previous methods in generating descriptions that closely match human language.
What Problem Does This Paper Address?
Generating descriptions for multiple items (e.g., "the red desk and the blue sofa") is a crucial part of natural language generation. Existing algorithms often focus on brevity, but this can lead to descriptions that sound unnatural or don't reflect how humans actually describe things. Specifically, previous work on "plural reference generation" (PRG) has struggled with:
- Empirical Backing: Many algorithms lack validation against real human language use.
- Human-like Strategies: Algorithms haven't fully captured how humans group objects (partitioning) and describe them in similar ways (similarity) to make descriptions more coherent.
- Complexity: Handling disjunctions (using "or") in descriptions can lead to logically complex and computationally expensive outputs.
This paper aims to bridge this gap by proposing a psycholinguistically-motivated algorithm that is grounded in corpus data and generates more human-like plural descriptions.
Methods
The core of this paper is the introduction of a new algorithm called IApart (Incremental Algorithm with Partitioning). It builds upon the existing Incremental Algorithm (IA) but incorporates two key human-like strategies:
- Partitioning: Instead of trying to find a single description for the entire set of objects, IApart first partitions the set based on fundamental categories (like "type" of object). This is akin to humans first thinking, "Okay, I have some desks and some sofas."
- Similarity and Salience: IApart prioritizes attributes that make the objects within a partition similar to each other. It also considers the "salience" or conceptual importance of attributes, learned from corpus data. This means it favors attributes that humans naturally use and find important (like color) and uses them redundantly across similar objects if it enhances similarity, even if it leads to slight overspecification.
The algorithm uses a corpus-derived heuristic to estimate attribute salience, which is crucial for deciding which attributes to propagate across partitions.
Datasets
The study utilizes a subset of the TUNA Corpus, specifically focusing on 900 references to furniture and household items. This corpus was collected through a controlled experiment with 45 participants. The objects in the domain have attributes like TYPE, COLOUR, ORIENTATION, and SIZE. The corpus data was annotated with semantic markup, allowing for the derivation of logical forms from natural language descriptions.
The authors used a stratified random sample of 180 plural descriptions (PL1) for analysis and the remaining 405 plural descriptions (PL2) for evaluation. They also used singular description data (SG) for estimating attribute probabilities.
Key Insights and Ideas
- Humans Partition First: People naturally group objects by their basic-level categories (like TYPE) before describing them, rather than trying to find a single, overarching description.
- Redundancy for Similarity: Humans often repeat attributes across descriptions of similar objects (e.g., "the small red desk and the small blue sofa") to emphasize similarity, even if it means slight overspecification. This is more common than aggregation (e.g., "the small red and blue desk and sofa").
- Attribute Salience Matters: The likelihood of an attribute being used redundantly in a description is strongly predicted by its overall salience or conceptual importance, as learned from singular descriptions.
- Overspecification is Natural: Contrary to the goal of brevity in some algorithms, human descriptions often include more information than strictly necessary for identification, especially when it enhances similarity or reflects conceptual grouping.
Key Results
- IApart Outperforms IAbool: The proposed IApart algorithm achieved significantly better results than van Deemter's IAbool algorithm in generating descriptions that matched human output. This was measured using tree edit distance, with IApart showing a consistently higher "perfect recall percentage" (PRP) on both datasets (with and without locative expressions).
- Human Descriptions are Parallel: The corpus analysis revealed that the vast majority of human plural descriptions exhibit "parallel semantic structure," meaning similar attributes are used to describe similar objects.
- Salience Predicts Propagation: The model successfully predicted the likelihood of attributes being propagated across partitions based on their salience, as estimated from singular descriptions. For example, COLOUR was found to be more likely to be propagated than SIZE.
Limitations and Future Work
- Complexity in Complex Domains: While IApart performs well, the authors acknowledge that in very complex domains, the partitioning strategy might lead to lengthy outputs. They suggest that incorporating negation could help manage this.
- Aggregation vs. Parallelism: The paper notes that while humans often favor parallel structures, aggregation (combining properties) might be desirable in some cases, but its use is limited by syntactic complexity.
- Further Empirical Investigation: The authors suggest that future work could explore how syntactic complexity influences the choice between aggregation and parallelism, and how negation can be used to avoid overspecification in complex scenarios.
Summary updated