Robust Structured Learning

Some of the most interesting and important problems in natural language processing (NLP) and machine learning (ML) rely on complicated inference procedures. Most practical systems are made up of many sophisticated sub-systems; these sub-systems are typically combined in an ad-hoc fashion. I am interested in developing machine learning methods for problems that feature (i) approximate inference, (ii) multiple sources of information (sub-systems) and (iii) distant supervision.

None of these challenges -- the use of approximate inference, the integration of multiple components and the use of distant supervision -- is new to ML. Yet, I believe that the availability of big interconnected data sets, ample computing resources as well as the maturity of fields such as NLP and Computer Vision bring about an unprecedented opportunity to build and deploy large ML-based systems. ML is currently missing an integrated, principled, probabilistic framework that accounts for all the complex attributes and their interactions. I am interested in developing such a framework.

Learning in Probabilistic Graphical Models

Probabilistic graphical models (PGMs) allow for the modeling of complex interaction between a large number of entities and their attributes in the presence of uncertainty. For that reason PGMs have proven suitable for many NLP tasks. Unfortunately, current methods for training (i.e., estimating parameters from data) PGMs are unsatisfactory when the model operates under adverse conditions e.g., predictions are made from the trained PGM using only inexact algorithms (since it would often intractable to identify the minimum-risk decision, or even do the exact inference needed to compute the risk of a given decision), the data on which the model is trained contains labels (observations) only for a small subset of the variables (attributes) and the structure of the model does not match exactly the structure of the data.

I am interested in developing novel methods for parameter estimation in PGMs that accounts for the adverse conditions in which the model will be used.

Entity Linking

Entity linking is the problem of matching mentions in text to their corresponding database entries. For instance, a news article may mention ``Pelosi," which should be resolved to the database entry for the US politician Nanci Pelosi. Typically, a system is provided with a database of entities and the relations between entities (e.g., Nancy Pelosi works for the US House of Representatives).

Entity linking can benefit from joint prediction: recognizing that an article mentions the US House of Representatives makes it much more likely that ``Pelosi" should be linked to Nancy Pelosi as opposed to say Italian footballer Claudio Pelosi. However, linking the congress mention is in itself an entity linking problem, so the two decisions reinforce each other and can benefit from joint inference. I am interested in applying structured prediction techniques to entity linking.

Opinion Analysis

Opinion analysis (or subjectivity analysis) is concerned with extracting information about attitudes, beliefs, emotions, opinions, evaluations and sentiment expressed in texts. Opinion analysis has been motivated by practical interest in applications that automatically mine opinion information from the large amount of available electronic text on the Web. Additionally, understanding whose opinions (or view points) is expressed in given piece of text is essential for many other natural language processing (NLP) applications such as information retrieval and question answering.

My research falls in the area of fine-grained opinion analysis, which is concerned with extracting opinions at the level of sentences, clauses, or individual expressions of opinions. I have been interested in finding ways to concisely summarize the opinions expressed in text in a way that makes them useful to other applications or for presentation to an user.

Coreference Resolution

Noun phrase coreference resolution or simply coreference resolution is the problem of identifying all noun phrases (NPs) in a document that refer to the same real-world entity. For example, in the sentence "Grown-ups never understand anything by themselves, and it is tiresome for children to be always and forever explaining things to them." the NPs "grown-ups", "themselves" and "them" all refer to the same entity, while "children" and "things" refer to different entities. The pronoun "it" is pleonastic -- it doesn't refer to any entity.

I am interested in approaches to coreference resolution that lead to globaly consistent solutions and rely on new sources of automatically derived semantic knowledge. I am also interested in improving methods for empirically evaluating and comparing different approaches to coreference resolution.

Information Extraction

Information Extraction is the task of automatically extracting structured information from unstructured text. For example, a system could process all news articles in a given day and extract all reports of natural disasters including the type of disaster (e.g., earthquake, flood, volcano), the place that it occured and the number of victims. Information extraction is an interesting application of NLP technology because it represents a task of practical importance that we can do relatively well (at least well enough that we can create useful systems).