The ever increasing amount of articles and the limits of human cognition require automated approaches to analyzing the sentiment expressed in user-generated content. As part of opinion mining, sentiment detection identifies and aggregates polar opinions – i.e., positive or negative statements about facts. For achieving accurate results, one needs to deal with the inherent ambiguities of human languages. webLyzard’s method to determine sentiment automatically has continually been optimized since 2003, directing particular attention to the context of opinionated terms when resolving such ambiguities.

webLyzard not only uses sentiment information to enrich visualizations such as tag clouds, geographic maps and information landscapes, but also offers high-performance data services for tagging third-party content.

Selected Publications

• Gindl, S., Weichselbraun, A. and Scharl, A. (2010). “Cross-Domain Contextualization of Sentiment Lexicons”, 19th European Conference on Artificial Intelligence (ECAI-2010). H. Coelho et al. Lisbon, Portugal: IOS Press: 771-776.
• Scharl, A. and Weichselbraun, A. (2008): “An Automated Approach to Investigating the Online Media Coverage of US Presidential Elections”, Journal of Information Technology & Politics, 5(1): 121-132.
• Scharl, A., Pollach, I. and Bauer, C. (2003). “Determining the Semantic Orientation of Web-based Corpora”, Intelligent Data Engineering and Automated Learning, 4th International Conference (IDEAL-2003), Hong Kong (Lecture Notes in Computer Science, vol. 2690). Ed. J. Liu et al. Berlin: Springer. 840-849.

• Frequency (Total, Positive and Negative). The displayed frequency values represent the number of occurrences in the last seven days. Selecting positive or negative frequency not only changes the data displayed in the chart, but also activates a global sentiment filter that affects the entire portal. Once activated, the footer of the portal shows the current filter status. Selecting total frequency deactivates the sentiment filter.
• Sentiment (Mean, Standard Deviation). While the mean value measures average sentiment towards a term within a sample, standard deviation reflects how contested a particular topic is (the term ‘oil spill’, for example, tends to have a low standard deviation since everyone agrees on its negative connotation).

Term Selection

The category drop-down menu enables users to select a category and show the five highest-ranking keywords within this category, or pick a set of predefined terms. Clicking on a keyword in the legend triggers a full-text query; clicking on the colored series marker hides the corresponding series from the chart. The term selection drop-down menu in the lower right corner allows accessing additional keywords computed for the selected period, and select an arbitrary set of these keywords to be displayed in the chart.

Display Features

The vertical axis is rescaled automatically, a feature that is particularly useful if one keyword dominates the coverage and therefore obscures the distributions of the other keywords.

Hovering above a data point displays the associated keywords. This mouse-over effect identifies topical trends and shows their impact on the individual peaks in the chart. Clicking on a data point triggers a search for this particular topic cluster in the preceding week.

Display of Search Results

Once a user has entered a search query, the system lists all sentences containing the search term and groups them by document. The column headers show the publication date, the sentiment of the document, the currently selected and total number of matches, and the average sentiment of the displayed search results. Users can sort the results by their date of publication as well as the sentiment on both the document level and the sentence level (the color of the term reflects sentiment on the sentence level).

Export Function

By clicking on their publication date, users can select and deselect the documents to be exported via the corresponding button in the upper right corner of the active document view. The average sentiment in the header is updated accordingly.

Impact on Other Views

Based on a full-text query, the system also updates the similar topics view and displays the summary of the highest-ranking document in the Active Document view, including its source and target geography (initially, these results remain hidden behind the keyword graph).

• The title field of the documents can be queried using the title: operator; searching for “carbon dioxide title:ocean” will only return documents which contain “ocean” in their title.
• The site: operator limits to search to a given domain; searching for “carbon dioxide site:cnn.com” returns documents on “carbon dioxide” mirrored from the CNN Web site.
• Documents in a certain time interval can be searched by using the from: and to: operators; searching for “carbon dioxide from:2010-01-01 to:2010-01-31” will only return documents published in January 2010.
• Source and target geography of the documents can be considered with the source: and target: operators; searching for “carbon dioxide source:US target:Kyoto” will return documents about carbon dioxide published in the US and referencing Kyoto as the main location in the text.

Clicking on the text block extends the quote, clicking on the circular marker on the left activates that particular document. Back and forward buttons enable the user to browse the list of related documents. The RSS links provide a continuously updated list of related documents (see below for a more detailed description).

The tag cloud displays more frequent terms in darker shades, using a larger font, less frequent terms use a smaller font and lower opacity settings. The color of individual terms indicates their sentiment (positive = green; neutral = black; negative = red). This allows investigating the “spin” across sources; e.g. typically balanced news media coverage compared to the very positive slant characteristic for corporate publications found on the Web sites of Fortune 1000 companies.

Interface Elements

• Captions. The keywords for each peak are calculated automatically, based on the content of surrounding documents.
• Document Markers. The markers [1-5] show the position of the documents from the Nearby Locations window.
• Document Selection. Hovering above any location in the map shows a preview of the closest matching document in the active document window, clicking selects this document (each of the small gray dots, which become visible after zooming in, represents one document).

After entering a search term, the set of results is visualized in the geographic map. Circular markers show the target geography of the found articles. The diameter of the marker represents the number of matches for a given location, its color the average sentiment of these matches. Using a color range from yellow to either green or red, trajectories link the source and target geography of an article (source geography = location of the publisher; target geography = main location referenced in the document). The little ‘+’ symbol on the right side of the window allows users choosing alternative base maps (e.g. NASA Blue Marble, Political Borders, Google Terrain, etc.), as well as deactivating the circular markers or trajectories.

• Similar Topics. Documents on similar topics relative to the active document are available in RSS format at www.ecoresearch.net/climate/rss/topic/[source]/[contentid], where [source] is the currently selected source (e.g. “climate.media” for media documents) and [contentid] is the identifier of the active document.
• Nearby Locations. Documents on nearby locations relative to the active document are available using the URL www.ecoresearch.net/climate/rss/location/[source]/[contentid].
• Tags. Documents matching a particular [tag] are available as an RSS feed using the URL www.ecoresearch.net/climate/rss/tag/[source]/[tag].