Form-Independent Meaning Representation for Eventualities

Natural language text such as that found on the web or in newspapers can nowadays be efficiently parsed, including building meaning representations or logical forms, with somewhat usable accuracy, at speeds of hundreds or even thousands of sentences a second. And Google is parsing everything we type at it. Nevertheless, we still don’t have real question-answering (QA), so that we can ask a question such as ‘Is the president in Washington today?’, have it mapped to an equivalent query, and get a precise answer. Such an answer could in principle be based on a semantic net or knowledge graph of eventualities, continually built and updated by semantic parsers reading the newspapers. Instead, we are still presented with a bunch of snippets from pages whose words and linkages may or may not answer our question when we
ourselves do the reading. 
The central problem in using parsers to answer questions from unrestricted text like this is that the answer to our question is very likely to be there somewhere, but that it is almost certainly in a form which is not the same as that suggested by the form of our question. For example, the question ‘Is the president in Washington?’ is in fact answered by the statement in today’s paper that ‘The president has arrived at the White House’. However, understanding this requires inferences that ‘having arrived’ at a place at a time entails ‘being at’ that place at that time, that being at the White House entails being in Washington, and so on. We ourselves draw all of these inferences effortlessly when we read the latter sentence. However, the standard logical form for our question is something like present (in washington president), while that of the text is present (perfect (arrived whitehouse president)). 
Of course, the commonsense knowledge that links the two statements can be hand-engineered for specialized domains, in this case by the use of resources such as named-entity linkers, ontologies, and gazeteers, and inference rules linking arriving with being there. However, there is simply too much of it to hand-engineer the open domain. The chapter begins by briefly reviewing some early attempts to build such representations by hand. It then compares the two main alternative contemporary approaches to the discovery of hidden meaning-representations for relation-denoting content words. Section 21.3 then examines the extension of one of these approaches to the discovery of latent episodic relations such as temporal sequence and causality between such terms, and examines some extensions and limitations of the approach. A brief concluding section considers some broader implications for the theory of meaning, andits implications for practical tasks like question answering.