The annotation result of PMCID 6440663—“We discovered inactivating mutations of tumor suppressor genes, including APC, TP53, and ARID1A, in three patients.”—as shown in Fig. 2, is used as an example sentence.

The GENIA tagger outputs the base forms, part-of-speech (POS) tags, chunk tags, and named entity tags. The tagger is specifically tuned for biomedical texts such as MEDLINE abstracts. Fig. 2A is a direct output from the GENIA tagger, and 2B is a visualization of NER generated by TextAE [10], the default viewer and editor of PubAnnotation. Four different levels of tags are attached for each word in the example sentence: base forms, POS tags, chunk tags, and named-entity tags. For example, “TP53”, “TP53”, “NN”, “B-NP”, and “B-protein” indicate that the part of speech of the word “TP53” is a noun (‘NN’), that the word begins a noun phrase (‘B-NP’), and that it begins a phrase of a protein name (“B-protein”).

The last tag is a semantic-level tag to classify named entities in the text into pre-defined categories such as proteins, DNAs, RNAs, cell lines, and cell types. For named-entity tags, B/I/O notation was used, wherein the B/I/O terminology refers to the beginning of the phrase (B), internal to the phrase (I), and outside of the phrase (O).

Fig. 2 shows that APC was wrongly classified, because “APC” could refer to the adenomatous polyposis coli gene or to an antigen-presenting cell. Generally, biomedical NER faces difficulties for many reasons, prominent among which are the often-ambiguous abbreviations that are frequently used in the biomedical field.

The desired result of a dictionary-based text annotation task would be an index of the dictionary entities corresponding to the referenced target texts. PubAnnotation’s text sequencer turns a document into a sequence of characters, so that positions in the document can be specified unambiguously by character offsets. For this reason, a conversion tool between the Genia tagger output and the JSON import/export format of PubAnnotation was written in Python during BLAH6. As shown in Fig. 3, we extracted the text field from a JSON file, and tokenized it by sentence, using the Natural Language Toolkit (NLTK) package, as in lines 67–73 [11]. This tokenizer divides a text into a list of sentences by using an unsupervised algorithm to build a model for abbreviation words, collocations, and words that start sentences. In line 12–27, indexes for each word in the Genia output are calculated by adding up the white spaces and character lengths. In lines 29–60, a new list is created, containing the begin index, end index, named entity tag, and word; B-tags and I-tags are combined, and the indexes are recalculated. Finally, the list to the denotation field of the dictionary is appended and converted into JSON.

A facility of visualization and manual editing is one of the primary aspects of making the PubAnnotation ecosystem adoptable by end-users. A user can easily add a new entry or delete an entry, in a try-and-revise manner.

In Fig. 4, there are four denotations for our example sentence, T1 through T4, with its tag and span information. The first one connects span 40-62 (the text spanning from the 40th to 62nd characters) to DNA, while the fourth connects span 89-95 to Protein. The default interpretation of T4 is as follows: the text span between "span":{"begin":89, "end":95} denotes an entity T1 "id":"T1" of which the type is Protein.

Once an annotation file is prepared, TextAE can be used for manual editing of NER [12], as in Fig. 5. TextAE is a web-based graphical annotation editor, which was developed as an open-source project. APC is now tagged as a “gene,” as shown in Fig. 5, after manual editing. In addition to NER tagging, the example also presents the ease of using TextAE for manual editing of relation annotations, showing that the two entities, T1 and T2, that are introduced by the two denotations, are related to each other by the predicate “contains”, specified by the two different keys, so the relationship is directional.