[On a tangent to the paper] In the recently concluded Annual HFES Meeting, Matthew Gombolay commented on “context”, that we human’s cannot explain the context behind the actions we take in a uniform manner. Politeness is one such contextual problems according to me. Why do we thank the bus driver who opens the door for us when we get off? Is it not his or her job to ferry us? Do we do this because of an inherent politeness we have as humans? Or do we do this because we saw other people do this and wanted to be a part of the group? A simple thing as thanking another individual can be contextualized in many ways. I feel that this very problem that we have not reached a consensus on contextualizing our actions is why we have not been able to teach a machine agent “context”.
Back to the topic of politeness that the paper deals with. I would like to reflect on the content in two halves, initially talking about the notion of politeness and referring to the paper where necessary and then moving to critique the procedure adopted in the research study.
I would point to two central questions that I wish to address.
What is politeness?
Is it as Brown and Levinson describe the emotional investment that we make to save our ‘face’? Is it an acknowledgment of power disparity with the person we converse with? Is it a complex mix of many things? I feel for us to converse about politeness we need to agree on the notion first. For the sake of this article, I will give the definition to which I will try and stick to.
I will define politeness as the behavioral characteristics an individual portrays, often by linguistic choice, that is civil relative to the people observing this individual.
By this definition, I give a relative measure for the word because it changes with the observers. A group might think to abstain from using cuss words as polite while another might consider the use of words like “Please”, “Thank you”, etc. in one’s conversations as a measure of politeness.
Do humans assume or adopt politeness to reach a higher power level?
From my understanding of the implications of the paper, humans tend to be polite as a means to an end. They are polite essentially to cause a disparity. This raises my question of “Are we subconsciously aware that we will be in a higher power state that the opposite party when we are polite to them?”.
My answer to the above question is that humans are polite to cause a disparity in power, but I am open to having my opinion changed.
Changing directions to critique the paper. I have the following critique:
- The paper deals with requests only. Is this enough to comment on “politeness” as a whole? Is it already sampling in a biased manner? I feel that the work did not defend its choice of requests in a sufficient manner. At least for me. I would have liked to have a mention of why they think requests generalize to other contexts more concretely.
- The test for inter-annotator agreement is interesting. The pairwise correlation test is definitely corroboratory to some of the claims that the authors have made. I failed to understand one aspect of this. When the classifier was compared to human performance by collecting more data. Was the inter-annotator agreement used as a measure of human-performance? If so is it not wrong? Humans are inherently different. A machine is always conformant to a given behavior. Is it not trivial to say based on disagreement that machine performs close to or better than humans?
- I am curious about the linguistic background questionnaire that the authors used and would definitely try to learn more about the same.
- The binary perception section mentioned the ends of the spectrum having more hits than the middle region. This reminded me of signal strength in signal detection theory. This is a Human Information Processing take where they comment that humans detect strong or weak signals very easily. But when it comes to signals that are hard to tell apart from each other, humans are bad at judging whether there is a signal or there isn’t. Only by designing the signal to have redundant dimensions can designers ensure that the right judgment of signal or no signal is made.
- I question the choice of not using any of the second domain for training. Would it not have made the more model-agnostic by using some of the data from the second domain as well?
- The paper talks about analyzing the requests made in these domains. I did not see a mention of the analysis of the responses that these requests got. I feel that an analysis of whether these requests were fulfilled would give valuable insight from the two parties involved rather than the retrospective nature of having a mechanical turker annotate it. An analysis of this sort would present a good baseline for comparison.
REFERENCES:
C. Danescu-niculescu-mizil, M. Sudhof, D. Jurafsky, J. Leskovec, and C. Potts, “A computational approach to politeness with application to social factors,” Proc. 51st Annu. Meet. Assoc. Comput. Linguist., pp. 250–259, 2013.
Penelope Brown and Stephen C. Levinson. 1978. Universals in language use: Politeness phenomena. In Esther N. Goody, editor, Questions and Politeness: Strategies in Social Interaction, pages 56–311, Cambridge. Cambridge University Press.
