Abstract
Previous work on bridging anaphora resolution (Poesio et al., 2004; Hou et al., 2013b) use syntactic preposition patterns to calculate word relatedness. However, such patterns only consider NPs' head nouns and hence do not fully capture the semantics of NPs. Recently, Hou (2018) created word embeddings (embeddings_PP) to capture associative similarity (i.e., relatedness) between nouns by exploring the syntactic structure of noun phrases. But embeddings_PP only contains word representations for nouns. In this paper, we create new word vectors by combining embeddings_PP with GloVe. These new word embeddings (embeddings_bridging) are a more general lexical knowledge resource for bridging and allow us to represent the meaning of an NP beyond its head easily. We therefore develop a deterministic approach for bridging anaphora resolution, which represents the semantics of an NP based on its head noun and modifications. We show that this simple approach achieves competitive results compared to the best system in Hou et al. (2013b) which explores Markov Logic Networks to model the problem. Additionally, we further improve the results for bridging anaphora resolution reported in Hou (2018) by combining our simple deterministic approach with Hou et al. (2013b)'s best system MLN II.
Introduction
In natural language processing, anaphora resolution is the task of identifying what a noun phrase (NP) refers to. While identity anaphora (e.g., "the CEO" and "she") is well-studied, bridging anaphora is more complex. It involves associative links based on relations like part-whole, set-member, or entity-attribute. For example, in the sentence "Delmed will continue to supply dialysis products... Delmed is exploring distribution arrangements," the phrase "distribution arrangements" relates back to "dialysis products."
Resolving these links requires understanding the semantics of modifiers (like "distribution") rather than just the head nouns ("arrangements" vs. "products"). Previous approaches often relied on complex machine learning frameworks or narrow syntactic patterns that ignored these crucial modifiers.
Methodology
This research introduces two primary innovations to address the limitations of head-only resolution:
1. embeddings_bridging
We propose a new word representation that combines general lexical knowledge with specific associative relations. By concatenating embeddings_PP (vectors trained on prepositional and possessive structures) with standard GloVe vectors, we create a 200-dimensional vector. This representation captures both general similarity and the specific "relatedness" required for bridging resolution.
2. A Deterministic Algorithm
Instead of a complex classifier, we employ a rule-based deterministic approach:
- Candidate Selection: Potential antecedents are sought in the current sentence, the two preceding sentences, and the document's first sentence.
- NP Representation: The semantics of an NP are represented by averaging the vectors of the head noun and its modifiers (adjectives, common nouns, and participles).
- Resolution: The algorithm selects the candidate with the highest cosine similarity to the anaphor.
The Role of Modifiers
A key insight of this work is that the head noun of a bridging anaphor is often semantically "thin" (e.g., arrangements, members, parts). The specific relationship is typically defined by the modifiers. Including these modifiers in the vector representation significantly improves the system's ability to distinguish between candidates.
As shown in the example above, using only the head noun might lead to selecting the wrong antecedent. When modifiers are included, the distance score ($dist_{hm}$) correctly identifies "dialysis products" as the primary associate for the anaphor.
Experimental Results
The system was evaluated primarily on the ISNotes corpus, consisting of 50 Wall Street Journal articles. To test generalizability, evaluations were also conducted on the BASHI and ARRAU (RST) datasets.
On the ISNotes dataset, the simple deterministic approach using embeddings_bridging achieved 39.52% accuracy, nearly matching the 41.32% achieved by the more complex MLN model from Hou et al. (2013b). Furthermore, when these deterministic insights were integrated into the MLN framework, the combined system reached a new state-of-the-art benchmark of 46.46% accuracy.
Conclusion
This work demonstrates that a simple deterministic algorithm, when paired with specialized word embeddings that account for NP modifiers, can rival complex machine learning models for bridging anaphora resolution. While challenges remain—particularly with "set-of" relations and broader discourse context—this approach provides a powerful, unsupervised baseline for associative link resolution.
Summary updated