Chambers and Jurafsky, ACL 2010
Template-Based Information Extraction without the Templates is a paper by Chambers and Jurafsky which be found online.
Citation
Chambers and Jurafsky. Template-Based Information Extraction without the Templates. In ACL 2010.
Summary
This paper proposed a solution finding and filling out templates in an unsupervised manner. Templates are extremely important in information extraction where they are a step above semantic role labelling. Templates, in a general sense, indicate what is happening and who is involved. While in most cases templates are either given out (to be filled) or engineered for a specific task, this paper tried to do both unsupervised. The dataset was the MUC dataset where the task is to find templates about terrorists in South America.
Method
While the actual method the paper followed can appear to be a mishmash of steps, the underlying ideas are quite clear. The first main idea is that events that occupy one template are bound to happen near each other. This means that if the verbs "kidnap" and "taken" are found close together in the corpus, that is a piece of evidence that "kidnap" and "taken" have something to say in the same template. If we find a lot of evidence for certain verbs to be together, then we group them in a template.
The other main idea is that while they need to get the verb relations from the given corpus, there are not enough examples in the corpus to find patterns about the verbs. For this, they expanded the corpus at hand by looking at the New York Times and the Reuters section of the Gigaword corpus. They only considered text that had to do with a template which they were considering.
With these two main ideas, let's consider the details. They found related event-like objects (verbs and nouns with wordnet type noun.event) and clustered them. With these clusters, they looked up more text from the Gigaword corpus for examples of these verbs. Using semantic role labelling, they found the object and subject of each verb. Using a list of subjects and objects from each verb, they clustered based on reference similarity in order to determine that the "kidnap.obj" is the same as "disappeared.subj". These define the template. When identifying a template, they saw how many roles could be extracted. When slot filling, they just looked up where the subject and object of each of the important verbs would go into the template.
Performance
There are really two tasks being done: template evaluation (how well did they identify what should be in the template (the roles)) and slot filling (how well did they fill the template). While template evaluation was a bit tough to do, since the dataset only asked about four big templates, the method performed rather nicely. In some situations the method found potentially valid slots that were not included by the task.
As for slot filling, they did slightly worse than semi-supervised methods in Patwardhan and Riloff's papers, but still brought home impressive results for a completely unsupervised method.
Related Work
Other work has been done in finding templates in an unsupervised manner. Good work has been done in Patwardhan and Riloff, EMNLP 2007 and Patwardhan and Riloff, EMNLP 2009.
