Counting Words & Phrases - (E3) Danescu-Niculescu-Mizil et al 2012

[HyunkuKwon]

Post questions here for the the following exemplary reading:

Danescu-Niculescu-Mizil, Cristian, Lillian Lee, B. Pang, and John Kleinberg. 2012. “Echoes of Power: Language Effects and Power Differences in Social Interaction.” Proc. 21st Int. Conf. World Wide Web: 699–708.

[Raychanan]

This paper uses textual data obtained from the Wikipedia community and from the transcripts of Supreme Court debates to conduct its research. Their evidence suggests that these texts can reveal power relations or power differences between those involved in the conversation. This power may be a static disparity in status or it may be based on some kind of situational dependency. The following questions arose for me in the course of my reading.

1. Have any papers conducted further validation of this relationship between language and power differences using different textual datasets? If so, did they support or oppose the paper's conclusions? According to the authors, the identification of language coordination phenomena in large-scale on-line text was done recently by using data from Twitter. But I'm more interested in papers that focus on small-scale settings and whether their conclusions are consistent with the conclusions of this paper?

2. Do we have the ability to detect smaller power differences? As the authors describe, the power divisions between regular Wikipedia users, admins-to-be, and administrators are very clear, as are the power differences between the Chief Justice, Associate Justices, and lawyers of the Supreme Court. But can cutting-edge research identify competencies that do not have such clear-cut power differences? Programmers, for example, may have very slight differences in power depending on their age at work, even though they have both just joined the firm and hold the same title.

[RobertoBarrosoLuque]

At the very end of the paper an SVM classifier using coordination features, stylistic features and normalized frequency of each word, is trained to predict if x has higher status than y in different scenarios. Lets say an alternate classifier using the same features but in which the target is to predict weather x {judge, wiki-voter} will vote for or against y, was trained. How would we compare the results of the original classifier and this alternate one? What could we say about the underlying features and their relationship to the target variables? Are comparisons between classifiers using the same feature space valid?

[Willy624]

This is related to @RobertoBarrosoLuque 's question.

I imagine classifiers are not used only in the end but also in the original empirical analysis, cause hand labeling all the words in all those documents seems a bit unreasonable. In either case, how confident can we be about the result? Specifically, are there any norms in the content analysis community that require researchers to at least run a set of different classifiers to avoid possible p-hacking issues?

Expanding on that, the coordinate measure was introduced in Danescu-Niculescu-Mizil et al. (2011), are there any other coordinate measure in the literature that can be used to validate their result?

[romanticmonkey]

Are there theories/papers on the reverse effect of the exchange theory. That is, given two strangers (A and B) chat online, not knowing each other's faces and personal info, if A purposefully avoid to coordinate with B in language, will B conceive A as a more powerful person? Will this phenomenon differ between collaborative and competitive environment (e.g. providing collab/debate tasks)?

[egemenpamukcu]

Would it be possible to "reverse engineer" the coordination effect, in order to understand the power differences in social interactions?

For example, in contexts where the power hierarchy is not as clear as they are in the Wikipedia community or the Supreme Court, can we understand which individual or group is seen as "more powerful" by the other(s)? For example, can we estimate the "loser" of a debate, an argument, or a negotiation among equals by looking at who is the one coordinating?

[theoevans1]

In the introduction, this paper lays out three kinds of power differences: "Sometimes these power differences are embodied in formal roles... Sometimes they are based on more informal differences in the respect or authority commanded by individuals within the group. And sometimes they are more situational: x may have power over y in a given situation because y needs something that x can choose to provide or not."

The paper most explicitly addresses the first and third, power "from status" and power "from dependence." Can we think of a way to test language coordination in the second instance, power from respect or reputation? I think the results point at this in some ways, such as in the finding that coordination gradually shifts for several months after a change in status, but am interested in how this could be studied more directly.

[chiayunc]

I find this piece fascinating in that it explores the role of language and its utterance in a courtroom and manages to captures some aspects of the dynamic. I agree most with its finding in regards to the Supreme Court database. However, something that could be considered is the aspect of the Court's function. In an oral argument, the most important thing is the accuracy and effectiveness of oral deliveries. It is very common for Justices and lawyers to eventually converge on their use of language, for the sake of accuracy and effectiveness. Of course, there is a power relation, but I do not think there will be coordination on all aspects of court speech. I think if there is a more fine-grained study, it would be extremely interesting to see, with what types of words, the lawyer coordinates and with what other types the Justices coordinates.

[Bin-ary-Li]

It is interesting that they choose Wikipedia and supreme court records for their analysis. In my understanding, the literary style of the discussion on the Wikipedia or courtroom records are probably marked by argumentation, which is only one possible style of verbal exchanges. Using such a limiting dataset, how can we say that language coordination can be predictive of conversers' power status in an interdomain way?

[william-wei-zhu]

In organizations, people in the position of power are usually the ones that define norms in the organization, implicitly or explicitly. Actors who are not in the position of power feel the implicit expectation to adopt the norm (hence the imitation of linguistic style in the paper). The implicit assumption is that actors who adopt the norm (imitate speech patterns) are more likely to be awarded than actors who don't. In both the court case and the wikipedia case, are there evidence that people who imitate linguistic patterns are more likely to get promoted than people who don't?

[xxicheng]

Since the authors used exchange theory as their framework, I wonder if there are any other theories on power differences.

The authors say by focusing on function word classes rather than contents, they are able to generalize their techniques. What do you think the advantages of these two methods: focusing on function word classes and content meanings? Are there any researches proving that the technique in this article can also be applied to other datasets?

[mingtao-gao]

It's a very interesting paper on power differentials in interactions. Because the Supreme Court text data is one of the primary sources, I thought of that people with higher power may repeat the person in lower power status to clarify the meaning of his words. I think such a case often happens in courts. My question is how should we distinguish such a case discussed above from the data that reflect the actual power differentials?

[romanticmonkey]

@william-wei-zhu I think this is a very interesting question.

In the paper, the author did compare the language coordination between admins-to-be and failed-to-be when they are in the RfA discussions. Admins-to-be's seem to be "promoted" because they coordinate to their voters more. However, I'm not sure how this phenomenon will extend into the corporate world. I'm hypothesizing that voters prefer the admins-to-be because they implicitly feel more alike to the admins-to-be because of the language coordination. People prefer to group those who are similar. (I'm sure there's a theory.) Therefore, in the corporate world, maybe employees get promoted not only because of their professional ability, but also because of their level of "likeness" to their boss (part of which language coordination reveals).

[ming-cui]

I am considering if the methods the authors used can be used to study the dynamics of power & status. Power & status do change. So will those who mimic the powerful also tend to be powerful and gain status as time goes by? Is it possible to study this using computational methods?

[Rui-echo-Pan]

This paper is very interesting! While I'm also thinking to add some measurement of people's perception of the power relation, to extend and enhance the research result. The assumption may be that people with a clearer perception of the power relation may tend to coordinate more.