TL;DR
This paper introduces a novel system for automatically correcting agreement errors in Russian text, a common issue in this highly inflected language. By employing a bottom-up parsing approach on an extended morphological structure, the system identifies and proposes minimal changes to fix these errors, demonstrating promising results on real-world Russian texts.
What Problem Does This Paper Address?
Russian, being a highly inflected language, possesses a complex system of word morphology. Verbs, nouns, and adjectives have numerous forms (e.g., a typical verb can have 30 finite forms and 190 participles). This richness, while powerful, leads to a significant proportion of grammatical errors in texts being agreement errors – instances where words are used in incorrect forms. Correcting these errors is crucial for improving the quality of Russian texts, especially in applications like automated text processing or editing. Existing approaches often rely on identifying sentences that cannot be assigned a valid syntactic structure and then proposing minimal changes to achieve one. This paper focuses on a specific type of error (agreement) within this broader framework.
Methods
The core of the system is an "agreement corrector" that operates on dependency structures, which are directed trees representing syntactic relationships between words. The process involves:
- Morphological Analysis: For each word in the input sentence, all possible morphological interpretations (homonyms) are generated. This creates a "morphological structure" (MorphS) for the sentence.
- Extended Morphological Structure: To handle agreement errors, an "extended" MorphS is created. This involves generating variants for each homonym by changing semantically "empty" morphological features (like case, number, gender for nouns/adjectives, or tense/voice for verbs). This significantly increases the number of possible word forms to consider.
- Bottom-Up Parsing: The system uses a bottom-up parsing algorithm, similar to those proposed by Lejkina and Tsejtin (1975). This algorithm builds "fragments" (dependency trees on segments of the sentence) by combining smaller fragments.
- Correction Strategy: The parser operates on the extended MorphS for increasing values of a parameter
R, which represents the maximum number of word forms allowed to be changed from the original sentence. The goal is to find the smallestRfor which a syntactically connected sentence (a single dependency tree, C=1) can be constructed. If C > 1 for the initial sentence (R=0), the system attempts to find corrections by increasingR. - Syntactic Preference: The parsing algorithm employs a "syntactic preference" strategy to prioritize more likely syntactic links, reducing the combinatorial explosion of possible structures.
- Pruning: A pruning step is introduced to filter out redundant hypothetical corrections by retaining only those SyntSs with the maximum weight (sum of arc weights).
Datasets
- Training/Evaluation: 100 sentences were chosen at random from the Russian journal Computer Science Abstracts. These sentences were filtered to contain no formulas or words in the Latin alphabet and to have a maximum of 50 words.
- Morphological Dictionary: A large morphological dictionary of about 90 thousand words, compiled by Vladimir Sannikov, was used. This dictionary is described as being close in lexical content to the Grammatical Dictionary by Zaliznjak (1980) but based on the linguistic processor's morphology model.
- Syntactic Dictionary: The system utilizes a syntactic dictionary covering about 15 thousand words. For words absent from this dictionary but present in the morphological one, "standard entries" are assigned, representing generalized syntactic properties.
Key Insights and Ideas
- Extended Morphological Structure for Error Correction: The core innovation is the use of an "extended" morphological structure, which systematically explores variations of word forms. This allows the parser to find valid syntactic structures by changing only a limited number of word forms, directly addressing agreement errors.
- Bottom-Up Parsing with Syntactic Preference: The application of a bottom-up parsing strategy, combined with syntactic preference rules, is crucial for efficiently navigating the vast search space created by the extended morphological structure. This balances thoroughness with computational feasibility.
- "Degree of Syntactic Disconnectedness" (C): The paper formalizes the concept of sentence connectivity using the "degree of syntactic disconnectedness" (C). A sentence with C=1 is considered syntactically connected (parsable into a single structure). The correction process aims to reduce C from its initial value to 1.
- Handling Incompleteness: The system acknowledges the incompleteness of its linguistic resources (syntactic dictionary, grammar). It addresses this by using "standard entries" for unknown words and by considering sentences with C > 1 but less than the initial C as "quasi-correct."
Key Results
- Initial Sentences: For 100 initial sentences, the corrector evaluated 95 as correct or quasi-correct. 3 sentences received "false alarms" (wrong corrections), and 2 cases resulted in system failure (time/memory limits).
- Distorted Sentences (Single Distortion):
- 1st Series: Out of 100 distorted sentences, 14 were well-formed and correctly identified. 79 initial sentences were reconstructed. 5 received false alarms, and 2 failed.
- 2nd Series: Similar results were observed, with 10 well-formed sentences correctly identified, 84 initial sentences reconstructed, 5 false alarms, and 1 failure.
- Performance Difference: The paper claims a performance difference of 5% for correct sentences and 67% for sentences with single distortions compared to an "ideal" corrector.
- Correction Success Rate: For over 90% of ill-formed sentences, the right corrections were found.
- Processing Time: Mean processing time was 11.2 seconds for an initial sentence (0.64 sec/word) and 11.4 seconds for a distorted one on a MicroVAX 3100. For sentences that were initially parsable (C=1), processing time was significantly lower (2.6 seconds mean).
Limitations and Future Work
- Linguistic Resource Incompleteness: The paper acknowledges that the morphological and syntactic dictionaries are not exhaustive, and the grammar does not cover all syntactic constructions. This can lead to well-formed sentences being incorrectly flagged as ill-formed (C > 1).
- Performance Degradation with High Error Density: The authors note that for "high noise" problems, such as reading handwritten texts or speech recognition, where error density increases, the quality of correction becomes "rather low."
- Massively Distorted Sentences: No regular experiments were conducted for sentences with more than one distortion. Experience suggests that performance degrades significantly with massively distorted sentences.
- Future Work:
- Incorporation of a larger morphological dictionary.
- Extension of the syntactic dictionary and grammar.
- Further debugging on real texts.
- The paper implies that improving the linguistic resources would lead to faster performance for sentences that are initially parsable.
Summary updated