TL;DR
This paper introduces a method for analyzing and processing "compact text," a specialized form of language common in technical fields like medical records, where sentences are often grammatically incomplete. The authors demonstrate that by understanding the inherent patterns and predictable deletions in such texts, a natural language processing system can effectively extract information as if it were processing full, standard English.
What Problem Does This Paper Address?
Many specialized fields, such as medicine, rely on "compact text" for efficient communication. This text is characterized by shortened, often grammatically incomplete sentences (fragments) that are nevertheless understandable to domain experts. Traditional natural language processing (NLP) systems, designed for full, standard English, struggle with these grammatical deviations, making it difficult to automatically extract information from such valuable data sources. The challenge is to develop NLP techniques that can reliably parse and understand these specialized, condensed forms of language.
Methods
The authors employ the Linguistic String Project (LSP) system, a natural language analysis tool, to process compact medical texts. Their approach involves a three-stage process:
- Parsing: The system uses a medically tuned grammar, augmented with productions for compact sentence types and special sublanguage constructions (e.g., dosage expressions). This step identifies grammatical relations like subject-verb-object and host-modifier, using a selectional mechanism to resolve ambiguities based on domain-specific cooccurrence properties.
- Syntactic Regularization: This stage transforms the parsed sentences into a more standardized set of basic syntactic relations without altering the information content. Reduced word forms, like abbreviations, are also expanded to their full forms.
- Information Formatting: The final step maps the words from the regularized sentences into a tabular representation. This mapping is based on cooccurrence patterns and lexical information, assigning words to specific format columns that represent semantic categories.
Datasets
The primary dataset used in this study consists of eight medical discharge summaries, each approximately 1-2 pages in length. These summaries were dictated by physicians and then transcribed into computer-readable form. The authors mention that the LSP system's medical parsing dictionary was augmented for new words encountered in these texts.
Key Insights and Ideas
- "Repetitive ungrammaticality is grammatical": In specialized compact texts, certain ungrammatical constructions are recurrent and predictable. These deviations can be characterized and incorporated into parsing grammars, effectively making them "grammatical" within that specific discourse.
- "Deletions are reconstructable": The missing elements in compact text are not random. They can be systematically reconstructed based on syntactic rules and the regular patterning of domain-specific vocabulary. This allows for the recovery of full sentence meaning.
- "Texts are convergent": Compact texts within a particular domain exhibit remarkable similarity. Once a set of generalized semantic patterns is identified in an initial document, these patterns tend to recur in subsequent documents from the same domain, simplifying the analysis process over time.
- Four properties enable processing: The authors identify four key properties of compact text – grammaticality despite syntactic deviation, regular vocabulary patterning, recoverable deletions, and convergence – that collectively make it possible to process them using syntactic procedures designed for full free text.
Key Results
The analysis of the medical discharge summaries revealed that 49% of the sentences were syntactically incomplete fragments. These fragments were categorized into six types based on regularly deleted elements. The study also identified 41 cooccurrence patterns of the subject-verb-object (SVO) type, expressible in terms of 18 participating word classes. These patterns could be further grouped into six more general types. The paper demonstrates that the LSP system, using its specialized grammar and processing steps, can effectively analyze this compact text, mapping it into a structured information format.
Limitations and Future Work
The authors acknowledge that while it's not a case of "seen one, seen them all," compact texts within a given area are remarkably similar. They found that once a set of generalized semantic patterns occurred in the first document processed, no new types were recognized in the remaining seven documents. On the lexical level, they note that "seepage" (new vocabulary) tends to taper off, but changing between medical subfields can introduce a significant percentage of new vocabulary. The paper's focus is on demonstrating the feasibility of processing compact text; future work would likely involve refining the system for broader application and exploring the nuances of different sublanguages.
Summary updated