The paper describes a rule-based information extraction (IE) system developed for Polish medical texts. We present two applications designed to select data from medical documentation in Polish: mammography reports and hospital records of diabetic patients. First, we have designed a special ontology that subsequently had its concepts translated into two separate models, represented as typed feature structure (TFS) hierarchies, complying with the format required by the IE platform we adopted. Then, we used dedicated IE grammars to process documents and fill in templates provided by the models. In particular, in the grammars, we addressed such linguistic issues as: ambiguous keywords, negation, coordination or anaphoric expressions. Resolving some of these problems has been deferred to a post-processing phase where the extracted information is further grouped and structured into more complex templates. To this end, we defined special heuristic algorithms on the basis of sample data. The evaluation of the implemented procedures shows their usability for clinical data extraction tasks. For most of the evaluated templates, precision and recall well above 80% were obtained.

Distributional Semantics postulates the representation of word meaning in the form of numeric vectors which represent words which occur in context in large text data. This paper addresses the problem of constructing such models for the Polish language. The paper compares the effectiveness of models based on lemmas and forms created with Continuous Bag of Words (CBOW) and skipgram approaches based on different Polish corpora. For the purposes of this comparison, the results of two typical tasks solved with the help of distributional semantics, i.e. synonymy and analogy recognition, are compared. The results show that it is not possible to identify one universal approach to vector creation applicable to various tasks. The most important feature is the quality and size of the data, but different strategy choices can also lead to significantly different results.

The paper addresses the detection of figurative usage of words in English text. The chosen method was to use neural nets fed by pre-trained word embeddings. The obtained results show that simple solutions, based on word embeddings only, are comparable to complex solutions, using additional information as a result of taggers or a psycholinguistic database. This approach can be easily applied to other languages, even less-studied, for which we only have raw texts available.

BackgroundHospital documents contain free text describing the most important facts relating to patients and their illnesses. These documents are written in specific language containing medical terminology related to hospital treatment. Their automatic processing can help in verifying the consistency of hospital documentation and obtaining statistical data. To perform this task we need information on the phrases we are looking for. At the moment, clinical Polish resources are sparse. The existing terminologies, such as Polish Medical Subject Headings (MeSH), do not provide sufficient coverage for clinical tasks. It would be helpful therefore if it were possible to automatically prepare, on the basis of a data sample, an initial set of terms which, after manual verification, could be used for the purpose of information extraction.ResultsUsing a combination of linguistic and statistical methods for processing over 1200 children hospital discharge records, we obtained a list of single and multiword terms used in hospital discharge documents written in Polish. The phrases are ordered according to their presumed importance in domain texts measured by the frequency of use of a phrase and the variety of its contexts. The evaluation showed that the automatically identified phrases cover about 84% of terms in domain texts. At the top of the ranked list, only 4% out of 400 terms were incorrect while out of the final 200, 20% of expressions were either not domain related or syntactically incorrect. We also observed that 70% of the obtained terms are not included in the Polish MeSH.ConclusionsAutomatic terminology extraction can give results which are of a quality high enough to be taken as a starting point for building domain related terminological dictionaries or ontologies. This approach can be useful for preparing terminological resources for very specific subdomains for which no relevant terminologies already exist. The evaluation performed showed that none of the tested ranking procedures were able to filter out all improperly constructed noun phrases from the top of the list. Careful choice of noun phrases is crucial to the usefulness of the created terminological resource in applications such as lexicon construction or acquisition of semantic relations from texts.