Determining and Mapping Locations of Study in Scholarly Documents: A Spatial Representation and Visualization Tool for Information Discovery James Creel.

Determining and Mapping Locations of Study in Scholarly Documents:

A Spatial Representation and Visualization Tool for Information Discovery

James Creel

Katherine H. Weimer

TCDL

May 7, 2013

Geospatial Information Retrieval Challenges

• How to utilize locations represented in text?– 20% of web queries have a geographic relation

(Ahlers)

• Traditional catalog subjects and keywords do not suffice

• Location information is increasingly in demand (Reid)

• Ex. 1049 total ETDs in 2005– 300 included locations (< 30%)– 130 contained international locations (> 10%)

What about a Visual Search?

• Searching collections with a map interface?– Visual representation of research– Enable serendipitous cross-disciplinary

collaborations and networking– Enhance access to the collection

Map Prototype

2011 – Geoparsing Work Begins

1. Overarching goal is to automate geocoding

2. Find toponyms in scholarly documents

3. Look up toponyms in a gazetteer

4. Disambiguate homonymous toponyms

5. Obtain geographic coordinates from gazetter

6. Encode coordinates in item surrogates for map-based view

7. Create map with link to original text

Desired Map Functionality

1. Base map: use Google Maps and other available interfaces

2. Cluster placemarks according to zoom level

3. List the displayed placemarks

4. Dropdown menu for countries and states in the US

5. Dropdown menu for departments grouped by college1. Selection of multiple departments in more than one college

2. If selecting the college, then select all departments within the college

6. Search by author

7. Time range slider (by year)

8. Use the Web-friendly University Brand color palette

Geocoding with KML files

• The KML file with locations includes:– Author– Title– Academic department– Advisor– Degree level– Year– Place– Keywords– Url to document

• Info box displays:– Author– Title– Academic department– Degree level– Year– Place– Url to document

Beta Version of Map: Showing Google Street Maps

Clustering Mechanism

User clicks on Point of Interest Title and Metadata Appear with Link to Text

Automated Process / Geoparser

• Geoparsing addresses two key problems:1) Name extraction

2) Name disambiguation

Document text

Extracted names

Disambiguated

names

Geospatial metadata

Geoparser: Comparable Models

• Edinburgh Geoparser– Grover, et. al. used OCR with historic records,

provided the GeoCrossWalk gazetteer

• DIGMAP Geoparser– Martins, et al. used originally for DIGMAP

digital library of historic maps

Geoparser: Setting

• DSpace 1.7 supports curation tasks– Custom Java programs

• Our instantiation:– Suggest New Metadata – Generate KML

Geoparser Workflow

Geoparser: Pre-Processing

– DSpace filter-media script extracts plain-text from PDFs.

– Suggest New Metadata curation task• Partitions the document into sections using regular

expressions• Excludes sections containing non-topical toponyms

(author-affiliation locations, conference locations, etc.)

Geoparser: Name Extraction

• ‘Named Entity Recognition’ or NER– Various open-source tools/training data

• Current version uses Apache OpenNLP or Stanford NER

• Classifies substrings of the text as names • Toponym occurrences are recorded in

context and counted

Name Disambiguation

• Requires reliable data- or knowledge-base• We employ the Geonames dataset

– Conglomeration of International gazetteers• Includes GNIS (USGS)

• Several complimentary methods– Rule-based– Heuristic– Statistical

Heuristics: Overview

• Various heuristics can help indicate the probable referent of a given toponym

• Other heuristics can help pick out false positives from the classifiers

• Heuristics are based on context-clues in the text or on general observations about human discourse

Heuristics: Context-based

• One document, one sense• Unambiguous extended names i.e. “Paris,

France”• Favor locations close to other mentioned

locations• Favor locations contained in other

mentioned locations• Favor locations of mentioned feature types

Heuristics: Generalized

• Favor higher-level administrative units (countries, states, cities)

• Favor locations of larger population

Heuristics: Application

• Heuristics - grouped into refinement iterations and then applied sequentially

• Resolve obvious cases first in order to provide better data for subsequent heuristics

Geoparser Evaluation

• Comparison of human annotations to geoparser output

• Precision/Recall of name extraction

• Accuracy of name disambiguation

Evaluator Workflow

Future Work

• Explore statistical disambiguation• Explore relevance of toponyms to the

subject matter• Expand to TDL collections• Expand to other digital collections or

collection types, even the library catalog?

Much more work to be done!

References

• Ahlers & Boll, “Location Based Web Search” in The Geospatial Web (London: Springer 2007)

• Apache OpenNLP. https://opennlp.apache.org/index.html• DigMap. http://portal.digmap.edu/• Leidner, Jochen L. “Toponym Resolution in Text” (Univ.

Edinburgh 2007)• Jenny Rose Finkel, Trond Grenager, and Christopher

Manning. 2005. “Incorporating Non-local Information into Information Extraction Systems by Gibbs Sampling” (ACL 2005) http://nlp.stanford.edu/~manning/papers/gibbscrf3.pdf

• Reid, James. “GeoXwalk – A Gazetteer Server and Service for UK Academia” (ECDL 2003)

Contact:

– James Creel

[email protected]

– Kathy Weimer

[email protected]