Tense Sense Disambiguation: A New Syntactic Polysemy Task

Roi Reichart, Ari RappoportEMNLP 2010 · OpenReview

Abstract

Polysemy is a major characteristic of natural languages. Like words, syntactic forms can have several meanings. Understanding the correct meaning of a syntactic form is of great importance to many NLP applications. In this paper we address an important type of syntactic polysemy — the multiple possible senses of tense syntactic forms. We make our discussion concrete by introducing the task of Tense Sense Disambiguation (TSD): given a concrete tense syntactic form present in a sentence, select its appropriate sense among a set of possible senses. Using English grammar textbooks, we compiled a syntactic sense dictionary comprising common tense syntactic forms and semantic senses for each. We annotated thousands of BNC sentences using the defined senses. We describe a supervised TSD algorithm trained on these annotations, which outperforms a strong baseline for the task.

Introduction: The Problem of Syntactic Polysemy

In NLP, Word Sense Disambiguation (WSD) is a well-established task—deciding which "bank" is intended in a sentence. However, this paper argues that syntactic forms are also polysemous.

Consider the sentence:

  1. "They are playing chess in the park." (Happening right now)
  2. "They are playing chess next Tuesday." (A future plan)

The words and the syntax ("present progressive") are identical, but the sense of the tense is different. This "syntactic polysemy" is a major hurdle for Machine Translation (where different senses might require different target tenses), event ordering, and textual entailment. Before this research, there was no formalized task or benchmark to help models navigate these temporal nuances.

Methodology

The authors formalize the task by distinguishing between an Abstract Syntactic Form (ASF) (e.g., present progressive) and a Concrete Syntactic Form (CSF) (the actual words in the sentence, like "are playing").

To solve TSD, they developed a supervised learning algorithm using the SNoW (Sparse Network of Winnow) architecture. Their feature set included:

  • Basic Features: The specific words in the tense form, their position, and proximity to pronouns.
  • Lexical Features: Presence of "time words" (e.g., yesterday, soon) or reporting verbs (said, replied).
  • POS Features: Part-of-speech tags of the verb and its immediate neighbors.
  • Structural Constraints: A key innovation where the model restricts its possible outputs to only those senses that are valid for the identified ASF.

Datasets and Resources

The authors introduced two major resources:

  1. Syntactic Sense Dictionary: Compiled from three major English grammar projects (Cambridge, Cobuild, and Longman). It defines 18 ASFs and 103 distinct senses.
  2. TSD Corpus: 3,000 sentences sampled from the British National Corpus (BNC), containing 4,702 annotated tense forms.

The dataset is notable for its fine-grained granularity; for example, the "present simple" alone has 11 different possible senses ranging from general facts to newspaper headlines.

Example senses for the Present Simple and 'be + to + infinitive' from the Syntactic Sense Dictionary.

Key Results

The authors compared their model against a "Most Frequent Sense" (MFS) baseline across different scenarios. The results demonstrate that the supervised approach significantly outperforms the baseline, especially when the underlying tense type is unknown.

Performance comparison between the proposed algorithm and the Most Frequent Sense (MFS) baseline.
  • TypeKnown Scenario: When the model knows the tense type (e.g., it knows it's looking at a present simple form), it achieved 62% accuracy, significantly outperforming the 46.7% baseline.
  • TypeUnknown Scenario: Even without being told the tense type, the model reached 49.7% accuracy, nearly four times better than the MFS baseline of 13.5%.
  • Inter-annotator Agreement: Human annotators agreed on the sense 84.2% of the time, suggesting that while the task is difficult and fine-grained, it is well-defined and learnable.

Insights and Future Directions

  • Tense ≠ Time: The paper reinforces the linguistic theory that grammatical tense is a poor proxy for actual time. A "present" tense can easily refer to the past or the future depending on context.
  • Syntactic Units Matter: Polysemy here belongs to the syntactic construction as a whole, not the individual words. In the "playing chess" examples, the verb "play" doesn't change its meaning; the construction does.
  • The Power of Constraints: Performance improves significantly when the model narrows the search space to senses valid for a specific ASF.

Future work involves incorporating document-level context, as human annotators often required full paragraphs to disambiguate senses. Additionally, integrating TSD into downstream applications like Machine Translation remains a key next step to measure real-world impact.

Summary updated

Abstract

Polysemy is a major characteristic of natural languages. Like words, syntactic forms can have several meanings. Understanding the correct meaning of a syntactic form is of great importance to many NLP applications. In this paper we address an important type of syntactic polysemy --- the multiple possible senses of tense syntactic forms. We make our discussion concrete by introducing the task of Tense Sense Disambiguation (TSD): given a concrete tense syntactic form present in a sentence, select its appropriate sense among a set of possible senses. Using English grammar textbooks, we compiled a syntactic sense dictionary comprising common tense syntactic forms and semantic senses for each. We annotated thousands of BNC sentences using the defined senses. We describe a supervised TSD algorithm trained on these annotations, which outperforms a strong baseline for the task.
BibTeX
@inproceedings{reichart2010tense,
  author = {Roi Reichart and Ari Rappoport},
  title = {{Tense Sense Disambiguation: A New Syntactic Polysemy Task}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2010},
  url = {https://openreview.net/forum?id=H1ZFUGzOWH},
}
Figure