TL;DR
This paper proposes a novel method for resolving structural ambiguities in natural language parsing by combining psycholinguistic principles with probabilistic modeling. By quantifying principles like Lexical Preference Rule (LPR) and Right Association Principle (RAP) using "three-word" and "length" probabilities, respectively, and then merging them with a "back-off" strategy, the authors achieve a significant 89.2% accuracy in disambiguation.
What Problem Does This Paper Address?
Natural language is inherently ambiguous. A single sentence can often be parsed in multiple ways, leading to different structural interpretations. For instance, in "I ate ice cream with a spoon," the phrase "with a spoon" could modify "ate" (how the ice cream was eaten) or "ice cream" (what kind of ice cream). Resolving these structural ambiguities is a fundamental challenge in natural language processing (NLP) for building robust language understanding systems. While psycholinguistics has identified human-like principles for disambiguation (e.g., preferring simpler structures or semantically consistent interpretations), directly implementing these principles has proven difficult. Conversely, purely statistical methods can be data-hungry and may miss established linguistic insights. This paper aims to bridge this gap by integrating psycholinguistic principles into a probabilistic framework.
Methods
The authors propose a two-pronged probabilistic approach to disambiguation, inspired by psycholinguistic principles:
- Lexical Likelihood (LPR-based): They introduce a "three-word probability" to quantify the Lexical Preference Rule (LPR). This probability captures the likelihood of a specific word (e.g., "spoon") appearing in a particular case slot (e.g., "with") of a verb's (e.g., "eat") case frame. The lexical likelihood of an interpretation is then calculated as the geometric mean of these three-word probabilities.
- Syntactic Likelihood (RAP & ALPP-based): For the Right Association Principle (RAP) and Attach Low and Parallel Principle (ALPP), they define a "length probability." This probability models the likelihood of a syntactic attachment based on the "length" (number of words) of the involved syntactic categories. For RAP, shorter distances are preferred, while ALPP favors parallel attachments of equal length. The syntactic likelihood is the geometric mean of these length probabilities across all attachments in a parse tree.
- Back-Off Method: To combine these two likelihoods, they employ a "back-off" strategy. First, they compare interpretations based on their lexical likelihoods. If the lexical likelihoods are very close (within a threshold), they then use the syntactic likelihood to make the final decision. This reflects the psycholinguistic intuition that lexical preferences often override syntactic ones.
Datasets
- Training Data:
- 181,250 case frames extracted from the Wall Street Journal (WSJ) bracketed corpus (Penn Tree Bank) were used to estimate three-word and two-word probabilities.
- 963 WSJ tagged corpus sentences were analyzed using a parser (SAX), and correct syntactic trees were manually selected to train length probabilities.
- Test Data:
- 249 sentences from the WSJ tagged corpus, not used in training, were used for evaluation.
Key Insights and Ideas
- Unifying Psycholinguistics and Probability: The core idea is that psycholinguistic principles, when quantified probabilistically, can significantly improve disambiguation accuracy. This moves beyond purely statistical or purely rule-based approaches.
- "Three-Word Probability" for Lexical Preference: Moving beyond simple two-word co-occurrences, the "three-word probability" (verb-slot-noun) provides a richer representation of lexical preference, capturing more nuanced semantic relationships.
- "Length Probability" for Syntactic Preference: Quantifying RAP and ALPP using the "length" of syntactic constituents offers a principled way to model structural preferences, aligning with human parsing behavior.
- Back-Off Strategy as a Hierarchy: The back-off method effectively models the hierarchical nature of disambiguation principles, where lexical factors are considered first, followed by syntactic ones.
Key Results
The proposed probabilistic disambiguation method achieved a number 1 accuracy of 89.2% on the test set. This is a significant improvement over methods relying solely on lexical likelihoods (95.7% for Lex3, 87.0% for Lex2) or syntactic likelihoods (66.7% for Syn) when used independently. The combined approach, termed "Lex3+Lex2+Syn," demonstrated the effectiveness of integrating both lexical and syntactic information.
Limitations and Future Work
- Data Sparseness: The paper acknowledges the "data sparseness problem," where estimating probabilities for rare word combinations requires large amounts of data. While their method for estimating probabilities is efficient, the problem cannot be entirely resolved.
- Mistakes by Individual Components: The analysis revealed that "Syn" (syntactic likelihood) made mistakes when lexical likelihoods were zero, sometimes preferring incorrect attachments. Similarly, "Lex2" (two-word lexical likelihood) made errors due to the limitations of relying on only two words.
- Future Work: The authors suggest that increasing the training data for lexical preference is the most effective way to improve disambiguation results. They also note that the impact of syntactic likelihood might become more significant with longer sentences. They also briefly compare their length probability model with PCFG for syntactic preference, finding their method to be superior.
Summary updated