NaturalLanguageProcessing-Lecture12 Instructor (Dan Jurafsky) :How about now? Oh, yeah, okay. Great, all right. So I’m Dan Jurafsky. I’m taking for Chris today as he emailed you because he had to leave town. And so we’re shifting around the schedule a little bit and we’re doing semantic role labeling today and he’ll return to more parsing on Thursday – on Wednesday and Monday. And most of these slides were taken from a really good tutorial that I recommend you look at. It has a lot more slides that I’m not covering by Scott and Kristina. Kristina was actually a student here, just graduated a few years ago, and some slides from Sameer, who is a student of mine. So a lot of students of ours have worked on semantic role labeling. So let’s just start with a – you can tell a slide from Kristina and Scott and imagine this sentence, “Yesterday Kristina hit Scott with a baseball.” There’s lots of ways that we could have expressed that sentence in different orderings or with different kinds of syntactic constructions. And the intuition should be that if we’re trying to build information extraction system or a question answerer or anything that the fact that these are syntactically different, that that have different parse trees – completely different parse trees, shouldn’t blind us to the fact that they have similar meanings. So we’d like to be able to capture this idea of shallow semantics that says that we want to know who did what to whom. That Kristina was the agent of some kind of violent act and Scott was on the receiving end and the baseball was the instrument of this violent act. I might be able to do that for any kind of parse tree of this type. Let me skip those. And so intuitively we talk about an agent of some kind of event, the active person who does the acting, a patient or theme which is the person affected by some action, the instrument and maybe some kind of time phrases or things like that. And our goal for today is gonna be can we extract these kind of – I call them shallow semantics because it’s not full logical form of something very complicated which Chris will get to later, it’s relatively shallow. Can we extract these very simply from strings of words? And you’ll see it’s very similar to the kind of parsing we’ll be talking about the last week. And in particular what we’re gonna propose is that the way to do this is to build a parse tree of the types you’ve looked at and notice that certain constituents like “with a baseball” wherever they happen to appear in the parse tree can be labeled with a particular semantic form. So we’re gonna actually build a parse tree and just map the constituents in the tree to semantic form. It’s a very simple process, in fact. So, some more examples of someparticular semantic roles, so here we have two breaking events where just to make it very clear that the subject, the syntactic subject of the sentence is different. In the first case the thing that’s broken is the object of break, it occurs after the verb, and in the second case it’s the subject of break. So we’d like to, again, extract the same semantics, figuring out that the window was the theme of the breaking event despite the – what seems like a syntactic difference. Similarly for an offering event we can have somebody offering something to somebody and we can order these things in all sorts of various ways. It could – the thing that’s offered, the guarantee, could be after the verb, it could be the second thing after the verb, it could be before the verb, and so on. So this is a passive sentence where it’s offered could say after the passive verb. So, lots of different surface realizations, same shallow semantic form. So why do we care about this? And we should always ask that for any NLP task. A lot of NLP application – a lot of NLP methodologies and tools and the math we use are very useful inside NLP for some NLP tasks. But at this point in the course, you should ask yourself, “What can I do with this that isn’t just relevant to NLP?” So the first and the most useful thing people have found semantic roles useful for was question answering. And so if I ask a question, like, you know, “What’s the first computer system that defeated Kasparov?” And the sentence in some web page that has the answer is in a different form and my goal is to pull out just the string Deep Blue as an exact answer, then a semantic parse can help me find that and find that exact answer. And more generally if I ask a question like, “When was Napoleon defeated?” I know that once I’ve found the page that has the answer I want to pull up the snippet. Let’s say I’m doing snippet generation for Google and I want to pull out just the snippet that has the answer I know that I’m searching for a case where Napoleon was the patient or theme of some defeating event and that what I want to know is – never mind the notation, I’ll come back to this – is the [inaudible] application. A lot of NLP methodologies and tools and the math we use are very useful inside NLP for some NLP tasks. But at this point in the core you should ask yourself, “What can I do with this that isn’t just relevant to NLP?” So the first and most useful thing people have found semantic roles useful for was question answering. So if I ask a question, like, you know, “What’s the first computer system that defeated Kasperov?” And the sentence in some web page that has the answer is in a different form and my goal is to pull out just the string Deep Blue as an exact answer, then a semantic parse can help me find that exact answer.And more generally if I ask a question like, “When was Napoleon defeated?” I know that once I’ve found the page that has the answer I want to pull up the snippet. Let’s say I’m doing snippet generation for Google and I want to pull out just the snippet that has the answer I know that I’m searching for a case where Napoleon was the patient or theme of some defeating event and that what I want to know is – never mind the notation, I’ll come back to this – is the temporal phrase that tells me the time of this event. Okay? So more generally any kind of who did what to whom question we can say – and this is gonna help us find exactly the answer phrase in the sentence. So the task is – the question answerer here is one more level even more specific than snippet generation