A Study on Meaning-Making in Machine Translation for Nonhumans
Summary
This thesis looks at the meaning that is sought in bioacoustics data. I argue that data practices are affecting ‘data narratives’ and that meaning is a place-based co-creating activity. I suggest ‘deep listening’: a practice that grounds data in place. I focus on projects that try to create machine translation for nonhumans, because of the impactful consequences their research has and has had, as well as the very interesting interpretations of concepts such as language that are the basis of their works.
My research question reads as follows:
How and why should we reinterpret meaning-making in data practices for machine translation for nonhumans?
I start by diving into the history of bioacoustics to explain from what vantage point these people are looking at the world. I look at the ethical and methodological concerns in current data practices in bioacoustics and ‘machine translation for nonhumans’, by doing a short case study on three entities in the bioacoustic machine translation world. I then focus on the FAIR and CARE debate to understand how data ownership and data sovereignty change the narrative and power imbalances in data. I also dive into discussions on the meaning-making of human and nonhuman worlds, and connect this to the meaning that is looked for in machine translation for nonhumans. I continue with a reading of Bakker’s deep listening, which can return and keep meaning in data. To finish, I discuss what it means to do deep listening in research that uses datasets for machine learning.
Shoot me a message if you would like to read the whole text.
References
Andreas, Jacob, Gašper Beguš, Michael M. Bronstein, Roee Diamant, Denley Delaney et al. 2022 Toward Understanding the Communication in Sperm Whales. iScience 25(6): 1-18.
Schroer, Sara A. 2021 Jakob von Uexküll: The Concept of Umwelt and its Potentials for an Anthropology Beyond the Human. Ethnos 86(1): 132-152.
This thesis was written for the BSc Cultural Anthropology and Development Sociology under the supervision of Rodrigo Ochigame.
(Non)Human Meaning-Making
In anthropology, ‘meaning-making’ is the way in which people and other subjects try to understand the world around us. To understand how others make reason, anthropologists examine societal structures at large to see how it affects the individual.
Natural language processing has a large semantic fundament. In semantics, meaning has a directional relation with a word or a sound. Machine learning models are used to map the relations between words or sounds. In datasets from bioacoustics origin, meaning is sought in the relation between sound and (re)action. Semantic models can map natural languages in multi-dimensional form (Andreas et al. 2022: 11). However, these models can only detect patterns. To make meaningful patterns, Project CETI and Earth Species Project use ethological fieldwork data and annotated datasets. Meaning should be put into the model, for an outcome of meaningful correlations. But, how do we know meaning?
In ethology, Von Uexküll has been a very influential, yet often misinterpreted, scholar. His theory on Umwelt is a great example of structural meaning-making. Von Uexküll argued, “all animals […] had to be understood as subjects and the worlds they lived in as constituted and made meaningful through their specific ways of perceiving and acting upon their worlds, their Umwelt” (Schroer 2021: 132). His deterministic view on life, can be seen in the metaphors he uses. There is the bubble, which he uses to explain that an Umwelt of one species cannot be understood from another species point of view, as they do not share the same senses through which they make meaning (Schroer 2021: 133-134). Then there is the polyphonic musical composition, which he uses to explain that the world is all these bubble-worlds together (Schroer 2021: 134). We can observe all these bubbles from the outside, but never will humans be able to understand other species, and vice-versa.
Schroer argues that his Umwelt theory can be understood differently, through a view that focuses on co-creation. She uses the term Worlding to focus on the musical analogy. With her research on falconry she shows how both human and nonhuman species create a their Umwelts together: “this musical motif nevertheless helps us to understand his conception of Umwelten as interconnected in a broader web or resonance of relations in which one organism and its corresponding Umwelt cannot be understood without reference to another” (Schroer 2021: 140).
‘Shared meaning making’ is taking this a step further, and foregoes the assumption that one’s worldview is predetermined (Schroer 2021: 141). Instead, Schroer argues, Umwelten are co-created; the conceptual bases on which we understand the world can be shaped by co-habitation (2021: 145). She uses the concept ‘worlding’ to make us understand that meaning is an active participatory process, which is shaped through but not limited by physical bodies (Schroer 2021: 145).
I agree with Schroer’s view on worlding. Worlding gives for an interpretation of meaning-making that allows meaning to be interpreted by fieldwork and participant observation. It encourages empathy in research, because an understanding between the researcher and the subject is cultivated and acknowledged. How can this view on meaning-making inform us on meaning in data?
In machine translation for nonhumans, the goal is meaningful communication between species. The key lies in meaning. On top of seeking meaning in semantic relations, I suggest to approach meaning as Schroer does. To understand the meaning that data carries, the context needs to be present. Because context gives meaning, meaning is key to interspecies communication. Machine learning models do not change meaning, because they do not use meaning as a signifier. The relations that they portray are semantic. This does not mean however, that these models are unbiased or objective. The data follows an existing narrative. On the other hand, this does mean that machine learning models cannot invent meaning. Interspecies understanding through these models can thus only rely on meaning that humans and nonhumans give to data. As we create meaning through data, it is useful to look at other possibilities to reach understanding. When we reimagine meaning as a co-created activity, it might in fact lead to different ways we could create interspecies communication.
(van Grieken 2024, 11-12)