information and language processing systems

isla, university of amsterdam

Language Technology Project 2008: Spotting opinions in news

This project is intended for Master of Logic and AI students with specialization in language and is worth 6 ECTS (see University of Amsterdam's course catalog). The project will be running from Jan 7 to Feb 1 and will be supervised by Valentin Jijkoun (jijkoun at science uva nl). Please contact him if you intend to sign up for the project, or in case of any questions.

The kick-off meeting for the project will be on Tuesday, January 8 12:00-14:00 in REC-B 3.41 (Building "B" on Roeterseiland). All students should be present. Please tell us beforehand (by email) if you plan to take this course but cannot attend the kick-off meeting.

Project background

During the project, a team of students will work together on a set of language processing tools for detecting and attributing sentiments (opinions) in Dutch news stories. To give an example, consider the following excerpt from a news article:

De betrokkenen zelf, hun advocaten, artsen en Tweede Kamerleden reageerden verheugd op het besluit van Verdonk.
(The people involved, their lawers, doctors and the members of the Parliament, were pleased with the decision of Verdonk.)

This sentence indicates that a specific agent (members of the Parliament) is supportive (rather than critical) of another agent (Verdonk) on some issue. Extracting triples <source, sentiment, target> is the goal of this project.

Why is extracting sentiments important? Well, there are many applications in the area of media analysis:

• We can identify controversial (sub-)issues: what are the topics on which many express their opinions?,
• We can summarize points of view on a specific issue: what are different opinions and opinion-holders?
• We can summarize opinions of a specific agent: on which issues does a person, organization or party express their opinions?
• We can measure overall support/criticism on an specific issue in the media.

These and many other applications can be built on top of a sentiment extraction system.

Modules and requirements

This project is essentially a team work. Team members will have different background and will focus of specific modules within the project.

• Constructing opinionated lexicon: create a lexicon of Dutch words and phrases that indicate sentiments (e.g., verheugd), using a corpus of articles with sentiments annotated by media analysts.
  Prerequisites: fluent Dutch, linguistic curiosity.
• Sentence-level sentiment classifer capable of distinguishing between opinionated and not opinionated sentences, and between positive and negative sentiments (support and criticism).
  Prerequisites: some background in machine learning.
• Entity extractor and sentiment binder: tools that will detect possible opinion holders and link sentiments found by the sentiment classifier with appropriate holders.
  Requirements: some background in natural language processing, esp. named entity recognition and co-reference resolution.
• Evaluation and demo: methodology and tools for evaluating the modules and the project as a whole, applying it to unseen news data. We also expect a simple interactive demo as a result of the project.
  Requirements: programming (Perl, Python, PHP, Java...)

Project structure and assessment

Every week the team will have two meetings with the superviser, to track the progress, monitor possible problems, get feedback etc. Deliverables at the end of the project:

• Written report, including problem statement, relevant literature overview, description of the used methods and resources, evaluation results, and error analysis.
• Source code and lexicon developed during the project.
• Final project presentation (30 min).
• A simple online demo.

These deliverables will be taken into account in the final mark for the project. On request, an individual contribution of team members (determined by a peer review) will be used as a factor for the individual final marks.

Literature

• V. Hatzivassiloglou and K. McKeown, Predicting the Semantic Orientation of Adjectives, ACL 1997 (pdf)
• S.-M. Kim and E. Hovy, Determining the Sentiment of Opinions, COLING 2004 (pdf)

Contacts

All project-related activities and developments can be found on a dedicated wiki.

Please contact Valentin Jijkoun (jijkoun at science uva nl) if you intend to sign up for the project, or in case of any questions.