Abstract
This paper describes a first attempt at a statistical model for simultaneous syntactic parsing and generalized word-sense disambiguation. On a new data set we have constructed for the task, while we were disappointed not to find parsing improvement over a traditional parsing model, our model achieves a recall of 84.0% and a precision of 67.3% of exact synset matches on our test corpus, where the gold standard has a reported inter-annotator agreement of 78.6%.
Introduction
In the field of Natural Language Processing (NLP), syntactic parsing and word-sense disambiguation (WSD) have traditionally been treated as distinct, sequential tasks. However, these processes are inherently codependent. Semantic information can guide syntactic decisions—for instance, identifying "Lotus" as a corporation helps resolve prepositional phrase attachment. Conversely, syntactic structure provides vital context for disambiguating word meanings; the grammatical frame of a verb often dictates its intended sense. This research explores a unified generative statistical model designed to tackle both challenges simultaneously.
Methodology
The proposed model extends the BBN SIFT parser, utilizing a lexicalized Probabilistic Context-Free Grammar (PCFG). The generative process is modified to incorporate WordNet synsets directly into the lexical triples (word, part-of-speech, and features). Specifically, the model generates a word's synset after the part-of-speech tag but before the word itself, allowing semantic categories to influence the structural probabilities of the tree.
To address the "sparse data" problem inherent in modeling specific word-sense combinations, the author implements semantic smoothing via WordNet’s hypernym hierarchy. When the model encounters an unseen relationship between senses, it "backs off" to more general concepts (e.g., from "attack" to "strike"). Decoding is performed using a modified CKY algorithm with a beam search to identify the most probable joint assignment of syntactic and semantic labels.
Dataset Construction
A critical component of this work was the creation of a synchronized dataset. The author merged the Penn Treebank-II (syntactic gold standards) with SemCor (WordNet sense annotations). This required a rigorous three-pass cleaning process to align tokenization and resolve discrepancies, resulting in a 220,000-word corpus suitable for joint training.
Results and Discussion
The model's performance in word-sense disambiguation is highly competitive. It achieved a total recall of 84.0% and a precision of 67.3% for exact synset matches. As shown in the table below, performance varied across parts of speech, with nouns showing the highest recall.
While the integration of semantic data did not yield the expected boost in parsing accuracy—with scores remaining between 80% and 83%—the model successfully maintained competitive performance despite the vastly increased search space introduced by thousands of potential synsets.
Limitations and Future Directions
One identified limitation involves the "head rules" used to determine the primary word in a phrase. Currently tuned for syntax, these rules sometimes prioritize semantically "light" words. For example, in the sentence "Jane will kill Bob," the modal "will" is treated as the head rather than the semantically rich verb "kill." Future iterations will focus on retuning these rules and incorporating broader contextual features to improve precision. Additionally, the author intends to apply these semantic enhancements to more advanced parsing architectures, such as Collins’ Models 2 and 3.
Summary updated