Transcript Slide 1
Computational Models of Discourse Analysis Carolyn Penstein Rosé Language Technologies Institute/ Human-Computer Interaction Institute Warm-Up Discussion What is social causality and what aspect of discourse analysis does it most naturally Student idea:connect to? I guess I don't understand how social What WhatI is the connection think might be between causality is different, in the way they're Girju’s approach and presenting it, from ordinary causality in interesting to do withEngagement? How would you characterize Girju’s sentences. Does it just require that both Engagement, is maybe to approach? targeted objects in each event are people? What was she They trying to good-bad do and somehow apply thismake thewhy? claim that studying What can used causality noveltoit in computational distinction tothisit.approach I'm notbeissure learn? linguistics. This is just wrong on the face of would be useful, but it could be it. I'm not really Student Comment: interesting sure how tototiesee any ifofcontracting this into really is "worse" than engagement, other than the fact that They claimed that he is a the paper confirms that affective expanding (wrt affective value, stellar PhD student although they correctly values aren't very useful in polarity). determining the sentiment of sentences. acknowledged not having any proof. Student Objection What is the function of reported events in a blog? I'm also confused about how the author thinks this analysis has predictive power for people's actions. I guess if we assume that What is the implication that event events, as reported by others, are not mentioned generalize over tens and biased and are not overly reported (likely modality? due to social influences), then we could use something like this to predict others' reactions?...I'm not sure. My somethingWhat would it do with something like “Those fishy-is-happening alarm Pakis [they] should be boiledisingoing hot oiloff. for Or am they some did.” from the Indian blog. Iwhat missing important part of the analysis? Student Comment I think Iris is right that the problem here is that it doesn't actually do much to identify facts about gender. Rather, it identifies ascribed behavior in reported events, and not even particularly generic reported events - only those events which would be reported in a very strictly structured way that only includes pronouns (and proper nouns? Student Question I'll buy that their way of identifying relations (using pronoun templates) works. It seems like a very fast algorithm with high precision (they suggest 97%, though probably with extraordinarily low recall). Also, what does it mean to "represent" 56% of the data? Student Question Any ideas on how this could be On page 68, I'm not quite sure how they got from extended to the case you use pronouns to clauses and howwhere they merged parts 2 and 4 arbitrary entities or arbitrary in with the rest named (if they did at all; this seems slightly different their approach on page nounfrom phrases rather than just69, but I might just be wildly confused). pronouns? P 69 is the next step – once you have the pair of clauses, you decide whether they are reciprocal or not. Reciprocity Reciprocal context: if a temporal order is indicated within C1 or C2 using discourse markers like “still”, “then”, etc., then it can’t be symmetric Type of eventuality [state versus event]: eventive version might sound symmetric, but it are not – “They chased each other” implies a sequencing Modality: would, should could – markers of events Temporal order of eventualities: Part 3 might indicate an ordering StudentI'mQuestion not all that enamored of their machine learning for symmetry. 84% accuracy with 78/22 split is probably around a 0.2-0.3 kappa, which is... okay? I guess? It's awkward that the features they describe were used, especially F1, since they're only using examples from the ambiguous patterns (I think). Similarly, 51% on a 3way classification task doesn't look that Reduction in error rate: great. I'm not sure what the set of classes100 - 78.33 = 21.67 Z is supposed to represent in their HMM.100 – 84.33 = 15.67 Then on the intentionality section they don't 21.67-15.67 = 6 even attempt to automate it - presumably6 / 21.67 = 27.7 because they tried, and performance was even worse than the other two classifiers.But: I 6/15.67 is around don't at all like that the different dimensions 37, so maybe the are being treated with such enormous divided by the wrong differences between them. thing? I'm also slightly confused about their symmetry results on the top of page 71; it seems that 84-37 = 47, not 78. Student Question I also don't understand how a HMM is useful for their purposes and also how the affective value works more than what I implemented for assignment 3. •Intuition: based on what you know from the sentiment lexicon, you can notice tendencies in ordering of positive and negative sentiment in eventualities, and you can use those regularities to learn sentiment associations for words where the sentiment is not known. •Note that it was not very accurate Eric: explain how this connects with what you did for assignment 3. Student question The author's final paragraph failed to convince me that social causality has much use if any at all because there is no way to normalize the affective aspects of causality. I could see potential use in suicide prevention--trying to understand blogs of people who have attempted or are at risk of attempting suicide before the blog's respective author is thrown overboard. Questions?