Best NLP papers

As the main COLING 2014 conference opens today, the guest blog series continues with some insights from one of the winning paper authors. Daojian Zeng, PhD candidate at National Laboratory of

There exists a significant amount of unstructured texts on the web, including newswires, blogs and scientific documents. How could a machine be able to read and understand the meanings of these texts?

One possible approach is to automatically convert these unstructured texts into structured data, where the key is to know the sematic relations between pairs of nominals. To this end, the relation classification is defined as a crucial step.

The most representative methods for relation classification use supervised paradigm. Such methods are further divided into two kinds including feature-based methods and kernel-based methods. These approaches are effective because they leverage a large body of linguistic knowledge. However, the extracted features or elaborately designed kernels are often derived from the output of preexisting natural language processing (NLP) systems (such as syntactic parse trees), which may lead to the propagation of the errors in the existing tools. Therefore, we wonder whether there is some method to extract features that is independent upon existing NLP tools?

The above problem is the issue that we addressed in our Coling 2014 paper. The paper, titled “Relation Classification via Convolutional Deep Neural Network”, employed a convolutional Deep Neural Network (DNN) to automatically identify the semantic relation types between pairs of nominals, which takes all of the tokens as input to generate features without complicated pre-processing.

Deep Learning (or Feature Learning) had excellent ability to capture the data characteristics and learn useful features automatically. In the task of the relation classification, to identify the relations between pairs of nominals, it is necessary to skillfully combine lexical and sentence level clues from diverse syntactic and semantic structures in a sentence. For example, in the sentence “The [fire]e1 inside WTC was caused by exploding [fuel]e2 ”, to identify that fire and fuel are in a Cause-Effect relationship, we usually leverage the marked nouns and the meanings of the entire sentence. How does the convolutional DNN learn such features?