NER and the Invisible in Plain Sight
In this blogpost, we want to share our experience about the application of a digital tool during one of our projects, namely the diary written by the local elite Zhang Gang.
One of the tasks in this project is to create a list of people who appear in a particular context of the diary. For this, we used the enpchina r-package to perform Named Entity Recognition (NER) on the text. We prepared a subset of sentences from the diary and let the algorithm process each sentence and extract the relevant text segments that were most likely to represent a person’s name. Although the algorithm is trained on Modern Chinese and the diary was written between 1888 and 1942, the language employed by Zhang Gang tends to be more vernacular than classical, which still brought relevant results that required only minimal cleaning.
After going back and forth between the text and the list, we came to the conclusion that extracting only names did not necessarily result in capturing every person of interest in the text. In fact, this way of collecting data tended to skip a particular group of people who are mostly ‘nameless’ in the diary: female family members.
Unlike the case for the male counterparts, Zhang Gang deployed terms that empathize their relationship, rather than referring to them by (some part of) their name. One example is how he mentioned daughter-in-laws and son-in-laws throughout the diary. Whereas the son-in-laws are referred to by their whole name, given name or just a part of the name combined with the family term (and sometimes place) , the given names of the daughter-in-laws are rarely mentioned. The names most commonly used for them consist of a family term and rank of seniority, once in a while paired with their surname.
|婿诚斋, 姚婿, 薛里姚婿, 薛里姚婿诚斋 or 四川姚婿||长媳|
|叶墨山, 叶婿, 叶婿墨山, 墨山, 墨山婿 or 墨婿||次媳|
|李公谅 , 瑞安李壻, 瑞安李壻公谅 or 李壻||三媳, 唐氏媳|
|戴腾秋, 鲍田戴婿, 戴婿腾秋 or 戴婿||四媳, 四媳蔡氏|
Despite their frequent occurrence in the text, these names did not correspond to the type of name ‘expected’ by the NER algorithm. As the NER algorithm is highly likely to be trained on the presence of surnames and/or common patterns of given names in a sentence, mentions of female family members in the text get (partly) ignored.
My two-cents about that…
At a surface level, it seems that a difference in writing habits points out a bias in the algorithm with regards as to what counts as a name. In that case, we can address the issue by retraining the model of NER on more fuzzy types of names and creating a new list of people.
However, I personally think that this application of NER reveals something more than what the issue let on. Our extraction process is not merely influenced by what the algorithm ‘expects’ as a name. The language in the historical source still shapes, to a certain degree, the way we obtain and curate our data. To return to our problem, even after we could list all the types of names that occurred in the diary, we could not automatically remove the ‘cloak of obscurity’ that the names create for these women. These names can still not be understood outside the context of the diary and, unless we identify each of the female family members with other external sources (if available), it is hard to re-use or enrich the data about these women.
To summarize, not all individuals present in digitized historical sources are ‘equally accessible’ to us and have the potential to be studied outside their original context when we only rely on automatized digital tools. If this problem goes unchecked, it will affect who will be included in the final analysis of someone’s research or in the search index of a database. Those who already enjoyed a great presence in historical research will have a greater chance to be noticed, while those who are clearly “there” in the sources, are rendered completely invisible.