TL;DR
This paper introduces a novel parser for French syntax that operates without a traditional dictionary, relying instead on the inherent redundancy of natural language. By treating sentences as stacks of noun phrases (NPs) and employing a "plucking off" algorithm, it achieves efficient syntactic analysis, demonstrating that rich linguistic information can be extracted from structure alone.
What Problem Does This Paper Address?
Parsing natural language, especially for research into syntax, traditionally relies heavily on extensive dictionaries to identify words and their grammatical properties. This approach faces challenges: building and maintaining comprehensive dictionaries is labor-intensive, and they may not capture the full spectrum of linguistic variation or novel terminology. The authors tackle the problem of syntactic parsing by questioning the necessity of a dictionary, proposing that the inherent "high formal redundancy" within natural language itself—the multitude of morphosyntactic clues—can be leveraged to infer grammatical structure. This is particularly relevant for research into French syntax, where understanding complex relationships between words and phrases is crucial.
Methods
The core of the proposed parser is its NP stack grammar (NPSG). The fundamental idea is to view a sentence as a "stack of NPs," where an initial NP is determined and refined by subsequent NPs through various "links" (prepositions, conjunctions, verbs).
Key methodological components include:
- NP-centric analysis: The sentence is broken down into NPs, which are considered the basic building blocks. The parser distinguishes between the internal grammar of an NP (centered on a nominalized word) and the external grammar of the sentence (treating NPs as closed entities).
- "Plucking off" parser (POP): This algorithm validates a sentence pattern (a sequence of NPs and external words) by progressively "plucking off" leaves (words or phrases) until the "heart" (the core NP) is reached. This process is used to identify and compute relations.
- Valuation functions: For relations that are not easily captured by algorithmic rules (e.g., prepositional phrase attachment, coordinations), heuristic valuation functions are employed. These functions use formal criteria to make choices when algorithms are too fuzzy, based on the hypothesis of high formal redundancy.
- "Reigner" concept: The authors propose "reigner" as a replacement for Chomsky's "governer," referring to the word that "reigns" over its dependents, analogous to Tesnière's "régissant."
Datasets
Not discussed in the paper. The paper mentions using "large text corpora" for testing and confirming hypotheses and a corpus of "about 10 000 words" for performance evaluation.
Key Insights and Ideas
- Dictionary-free parsing is feasible: The most significant insight is that a substantial amount of syntactic information can be extracted without relying on a lexicon, by exploiting the structural redundancy of language.
- NPs as fundamental units: Elevating the noun phrase to the primary structural unit simplifies the parsing process by providing a clear hierarchical framework.
- "Plucking off" as an intuitive parsing metaphor: The artichoke analogy and the "plucking off" process offer a novel and understandable way to conceptualize syntactic analysis.
- Heuristics for ambiguity resolution: Valuation functions provide a principled way to handle linguistic ambiguities that resist purely algorithmic solutions, by leveraging formal criteria.
- Unified treatment of verbs: Verbs, regardless of their form or clause type, are treated as adjectives that determine an NP and act as links in the NP chain, simplifying their integration into the syntactic structure.
Key Results
- High accuracy with minimal data: The parser achieves high accuracy with very few data points: 300 ending rules and a lexicon of only 4 kB for grammatical words.
- Efficient processing speed: On a Macintosh II, the parser processes at approximately 3 words per second.
- Low error rates: Category errors are reported at less than 1% of words, and relation errors at less than 3% of relations.
- Demonstrated syntactic analysis: The parser successfully identifies and computes various types of relations, including internal NP relations, sentence-level relations (like subject-verb), and looser external relations.
Limitations and Future Work
- Limitations:
- The paper acknowledges that valuation functions are heuristic and may not always yield the optimal result, though they are presented as an efficient way to exploit formal criteria.
- The "plucking off" process, while intuitive, might be computationally intensive for very deep or complex sentence structures.
- Future Work:
- The authors suggest that valuation functions can be further developed to compute "looser relations" such as prepositional phrase attachment and coordinations, indicating ongoing research in this area.
- The parser is presented as an "experimental device" for testing linguistic hypotheses, implying its potential for broader applications in linguistic research.
Summary updated