ORAL HISTORY INVESTIGATION SESSION IN LEICESTER ON 28 JUNE

We discussed what could be, for us, a ‘good enough’ proof of concept that can provide design specifications for the National Collection from the perspective of Oral History. We realized that, despite a preliminary proof of concept is already accomplished (but it would require fine tuning with a NLP specialist), there is still a question to be answered:

Within what’s been done, how do we start to provide the structured data that enable to connect OH recordings with other datasets?

We agreed that in order to complete the proof of concept, we don’t need further Oral History datasets, but we would need to work on two different levels:

• Including collection data into the exploratory tool to explore the linkage with museum collections. Once the collection data are in the same vector space, we can see how the new clusters correspond/differ from the previous OH data. In this way, we can potentially identify for each of the sentences, some objects that might relate to them. Considering the OH dataset, there are two potential collections that can be use for this stage: the National Coal Mining Museum collection, and the collection of the former Snibston Museum (which now should be hosted by the Leicestershire County Council).
• Fine Tuning the Model: improving the pre-processing, experimenting with other language models to understand what is the best model to use. This would require for Stef to work with a NLP specialist. We discussed this opportunity with Neslihan Suzen, Lecturer in Applied Mathematics at University of Leicester who already collaborates with Stef*.*

We reflected on the fact that visualizations usually have two steps, one is exploratory and the second is presentational for the public realm. Until now, the tool we are working is for expertise/internal use only, and further work needs to be done to reach the second level. There might be scope to apply this pipeline to the textile OH dataset for the Bradford Exhibition and develop a presentational visualization.

We agreed on the following next steps:

• Work on an academic paper and internal report where we can explain how we have worked, what we have achieved, what are the limitations and potential next steps
• Complete the proof of concept, by fine tuning the process with Neslihan, and providing Stef with the collection data from the National Coal Mining Museum (and, potentially, the former Snibston Museum). Working with this data would require an amendment to the original ethic application submitted to the University of Leicester, and the agreement with the Science Museum signed (this might still take a couple of months)
• Organize a Workshop in October as a concluding moment of this proof of concept and a generative session for future development of the visualizations and next steps

We also talked about a criticality which was raised at the Museum Analytics Conference in May, on the risk to limit to textual and numerical data a sonic historical form such as Oral History. We discussed the opportunity to bring the audio dimension back into the exploratory tool, connecting the clusters to the correspondent audio extracts. Stef highlighted that this is not a problem to implement, but it takes time and it is something that can be potentially developed in a future stage. We agreed that, for the purpose of the Worskhop, we could manually mix different voices related to one of the topics identified, to provide a use case scenario to discuss.

In the last part of the day, Colin Hyde and Sally Horrocks joined us, and we discussed the role of archive owners and Oral Historians who know the data very well and that can have a fundamental role in making sense of the clusters (also considering that, with the dimensionality reduction, you lose a lot of information). For example, there are particular words in each mining communities that only oral historical knowledge would be able to understand (Sally provided the example of the word ‘snap’ to mean lunch in the East Midlands mining community). In the CE vision of collaboration, human is as important as the automated process, and we need to identify what kind of expertise and technical capabilities contribute to refine the topics. There is also important information at the paralinguistic level (for example the tone of voice would tell you if one sentence is a sarcastic comment or not). The role of pauses and silences are also important, very often they are the most interesting part of the interviews, as well as the pace of different sections.

Sally and Colin also suggested further potential areas where the exploratory tool could help:

• Sensitivity review: understanding when people are talking about someone else, and mentioning someone who might be identifiable
• Unexpected topics: this tool could be used to help identifying topics which are not necessarily closed to the focus of the OH project (example of one interview talking about food in Mines of Memory, which might be of interest for somebody working on food history who might not have thought to look at a OH mining project)
• Keyword search: being able to search into the tool using keywords (and not only across the automated generated topics)
• Identifying gaps: if you are half a way through in a OH project, you could use this tool to look at the gaps in the topics, to address what’s missing and identify people who are not yet represented.
• Geographical representation of people’s lives: this is full of potential but, because place names are not easily retrieved by automatic transcription, it would require a lot of work and developing this dimension can become an independent project on its own.