Enhanced Integrated Scoring for Cleaning Dirty Texts

An increasing number of approaches for ontology engineering from text are gearing towards the use of online sources such as company intranet and the World Wide Web. Despite such rise, not much work can be found in aspects of preprocessing and cleaning dirty texts from online sources. This paper presents an enhancement of an Integrated Scoring for Spelling error correction, Abbreviation expansion and Case restoration (ISSAC). ISSAC is implemented as part of a text preprocessing phase in an ontology engineering system. New evaluations performed on the enhanced ISSAC using 700 chat records reveal an improved accuracy of 98% as compared to 96.5% and 71% based on the use of only basic ISSAC and of Aspell, respectively.

💡 Research Summary

The paper addresses the problem of cleaning “dirty” textual data—texts that contain spelling errors, ad‑hoc abbreviations, and improper casing—particularly as they appear in online sources such as corporate intranets, web pages, and chat logs. While many ontology‑engineering pipelines rely on clean, well‑structured corpora, the authors note that the quality of input text has a direct impact on the quality of the resulting ontologies, yet systematic preprocessing for noisy text has received little attention.
To fill this gap, the authors build upon their earlier Integrated Scoring for Spelling error correction, Abbreviation expansion and Case restoration (ISSAC) framework. The original ISSAC works by feeding each token flagged as erroneous by the Aspell spell‑checker into a candidate list that includes (1) Aspell’s ranked suggestions, (2) possible expansions from an abbreviation dictionary, and (3) the original erroneous token itself. Six weighting factors are then applied to each candidate: the original Aspell rank, a normalized edit‑distance score, a reuse factor (whether the candidate has been used before for the same error), an abbreviation factor (whether the candidate appears in the abbreviation dictionary), a domain‑significance score, and a general‑significance score. The original system achieved 96.5 % accuracy on a set of 700 chat records, but the authors observed systematic failures when multiple error types co‑occurred.
The enhanced ISSAC introduces three major improvements. First, the domain‑significance (DS) and general‑significance (GS) scores are refined by incorporating contextual information from the immediate left and right neighboring words (l and r). Instead of relying solely on raw frequency ratios, the new scores weigh how often a candidate appears in the specific domain corpus versus a general corpus in the given context, thereby reducing false positives for domain‑specific jargon. Second, the reuse factor (RF) is expanded into a persistent correction history table. Whenever a particular error‑candidate pair is corrected, the pair is stored; subsequent occurrences of the same error automatically receive a higher RF value, promoting consistency and reducing processing time. Third, the abbreviation expansion step is linked to an external web service (www.stands4.com). When a candidate is identified as an abbreviation, the system queries the service for the most up‑to‑date long form, caches the result locally, and adds it to the abbreviation dictionary. This dynamic approach mitigates the staleness of static abbreviation lists and helps disambiguate abbreviations that also exist as ordinary words.
The evaluation methodology consists of splitting the 700 chat logs into seven distinct subsets and performing cross‑validation. For each subset, three configurations are compared: (a) basic ISSAC, (b) the enhanced ISSAC, and (c) Aspell alone. The enhanced system consistently outperforms the baselines, achieving an average accuracy of 98 %—an improvement of 1.5 percentage points over basic ISSAC and a striking 27 percentage points over Aspell alone. Detailed analysis shows that the gains are most pronounced in abbreviation expansion and case restoration, where contextual weighting and dynamic lookup reduce ambiguous corrections.
The authors acknowledge several limitations. The computation of DS and GS requires a sufficiently large domain‑specific corpus; in domains with scarce data the scores may become unstable. Dependence on an external web service introduces a potential point of failure and may raise latency concerns for real‑time applications. Moreover, the current implementation is English‑centric; extending the approach to multilingual settings would require language‑specific spell‑checkers, abbreviation resources, and case‑handling rules.
In conclusion, the enhanced ISSAC demonstrates that an integrated, context‑aware scoring mechanism can substantially improve the cleanliness of noisy texts, thereby supporting higher‑quality ontology construction and downstream text‑mining tasks. Future work is proposed in three directions: (1) adapting the framework to multiple languages, (2) applying it to streaming data such as live chat or social‑media feeds, and (3) incorporating deep‑learning‑based contextual embeddings (e.g., BERT) to replace or augment the handcrafted DS/GS scores, potentially yielding further accuracy gains.