TL;DR
This paper introduces a novel method for correcting noisy sentences, particularly those generated by machine translation systems. By leveraging monolingual corpora and synonym dictionaries, the approach efficiently identifies and replaces erroneous phrases with more fluent alternatives, significantly improving sentence quality.
What Problem Does This Paper Address?
Machine translation (MT) systems, while increasingly sophisticated, often produce output sentences containing various forms of "noise." This noise can manifest as spelling mistakes, incorrect word choices, grammatical errors, or improper word order. These errors degrade the readability and accuracy of the translated text, hindering effective communication. Existing methods for correcting such errors can be time-consuming, especially when requiring parallel corpora of noisy and correct sentences. This paper aims to address this by developing an efficient post-processing technique that utilizes readily available monolingual data and synonym resources.
Methods
The core of the proposed method involves a dynamic programming approach to de-noise sentences. The process begins by splitting a noisy sentence into smaller phrases. For each phrase, the system searches a language model (built from a monolingual corpus) for frequent and similar phrases. This search is enhanced by considering variations of the input phrase, including:
- Phonetic variations: Words that sound similar.
- Morphological variations: Stemmed versions of words.
- Lexical variations: Synonyms of words.
A scoring function, inspired by search engine ranking, is used to evaluate the similarity and frequency of candidate phrases. This function considers factors like term frequency, inverse document frequency, and a boost for phrase frequency. The system then combines candidate phrases from sub-phrases to construct potential corrected sentences. A triangular matrix data structure is employed to efficiently store and combine phrase candidates in a bottom-up manner. The best candidate sentence is selected based on these scores.
Datasets
- Monolingual Corpora: Hindi and Bengali monolingual corpora, crawled from the web and cleaned, were used to build language models. The sizes are 300K sentences for Hindi and 430K for Bengali.
- Synonym Dictionaries: Hindi and Bengali synonym dictionaries were used to group words into synonym sets.
- Stemmers: Hindi and Bengali longest suffix strippers were used for morphological analysis.
- Phonetic Mappers: Hindi and Bengali words to phonetic word map tables were used.
Key Insights and Ideas
- Leveraging Monolingual Data: The paper demonstrates that high-quality sentence correction can be achieved without the need for expensive parallel corpora, relying instead on abundant monolingual data.
- Multi-faceted Phrase Similarity: By considering phonetic, morphological, and lexical variations, the system can identify a broader range of suitable replacements for noisy phrases, going beyond exact matches.
- Dynamic Programming for Efficiency: The use of dynamic programming and a triangular matrix structure provides an efficient way to explore and combine phrase candidates, making the de-noising process scalable.
- Scoring Function Adaptability: The scoring function, adapted from search engine principles, effectively balances phrase frequency, term relevance, and phrase length, leading to better candidate selection.
Key Results
The proposed approach was applied as a post-processing step for Hindi-Bengali and Bengali-Hindi machine translation systems. The results showed significant improvements in sentence quality:
- Perplexity Reduction: The modified systems achieved lower perplexity scores, indicating more fluent and natural-sounding sentences. For example, the BHMT system's perplexity dropped from 34.243 to 27.811.
- BLEU Score Improvement: BLEU scores, a common metric for MT evaluation, also saw improvements. The BHMT system's BLEU score increased from 0.232 to 0.243.
- Qualitative Improvements: The paper provides examples of corrected Hindi sentences, showcasing the system's ability to fix spelling errors, insert missing words, and correct word order.
Limitations and Future Work
- Long-Range Re-orderings: The authors acknowledge that the current approach is not adept at handling long-range re-orderings within sentences.
- Generalizability: While the approach is presented as general and usable for any language, further experiments are needed to confirm its effectiveness across diverse linguistic structures.
- Future Work: The authors suggest further experiments, better tuning of the scoring model, and testing on different sources of noisy sentences as directions for future research.
Summary updated