Temporal and Event Analysis of Natural Language Texts Siim Orasmaa - - PowerPoint PPT Presentation

Temporal and Event Analysis

• f Natural Language Texts

Siim Orasmaa

Data

• Estonian Reference Corpus of the University
• f Tartu

– Variety of text genres (news, popular science,

legal texts, parliamentary transcripts)

– Automatically processed:

• Sentence and clause boundaries detected
• Morphological analysis provided
• Robust temporal expressions annotation

– Based on TimeML annotation language

An example of annotations

http://www.keeleveeb.ee

• I. Comparing documents by temporal

similarity

• Given a newspaper article, find temporally similar

newspaper articles - articles that refer to

• verlapping/similar time periods;
• Task:

– Preprocess/index document collection – Implement a temporal similarity measure e.g Temporal Analysis of Document Collections: Framework and Applications, Alonso et al., 2010. – Add a text similarity measure e.g Exploiting Temporal References in Text Retrieval, Arikan, 2009.

• I. Comparing documents by temporal

similarity

• Evaluation:

– Using roughly temporally parallel corpus

(newspaper articles from Eesti Päevaleht 1999 and Postimees 1999)

– Prepare some test data

• How well can you detect documents

discussing same events?

• How much the results depend on

newspaper article's category (News, Opinions, Sports, Economy etc)?

• II. Clustering temporal expression

contexts

• More fine-grained approach: an event mention should

be located somewhere near the temporal expression (e.g a verb, noun or some phrase).

• Task:

– Use an unsupervised algorithm to cluster

temporal expression contexts, e.g like in Word Sense Induction.

e.g Unsupervised corpus-based methods for WSD, Pedersen, 2006.C

– Can you detect some broad event classes? – Test the algortihm on different text genres.

• II. Clustering temporal expression

contexts

• Discussion:

– Can you propose a meaningful labeling for

found clusters?

– Can you draw parallels between found clusters

and proposed event classifications (e.g the

• ne in TimeML)?

– Does the clustering help to organize temporal

expressions for information retrieval?