Jonas Sjöbergh

Recognizing Humor Without Recognizing Meaning

We present a machine learning approach for classifying sentences as one-liner jokes or normal sentences. We use no deep analysis of the meaning to try to see if it is humorous, instead we rely on a combination of simple features to see if these are enough to detect humor. Features such as word overlap with other jokes, presence of words common in jokes, ambiguity and word overlap with common idioms turn out to be useful. When training and testing on equal amounts of jokes and sentences from the British National Corpus, a classification accuracy of 85% is achieved.

Robots Make Things Funnier

We evaluate the influence robots can have on the perceived funniness of jokes. We let people grade how funny simple word play jokes are, and vary the presentation method. The jokes are presented as text only or said by a small robot. The same joke is rated significantly higher when presented by the robot. We also let one robot tell a joke and have one more robot either laugh, boo, or do nothing. Laughing and booing is significantly funnier than no reaction, though there was no significant difference between laughing and booing.

Unsupervised evaluation of parser robustness

This article describes an automatic evaluation procedure for NLP system robustness under the strain of noisy and ill-formed input. The procedure requires no manual work or annotated resources. It is language and annotation scheme independent and produces reliable estimates on the robustness of NLP systems. The only requirement is an estimate on the NLP system accuracy. The procedure was applied to five parsers and one part-of-speech tagger on Swedish text. To establish the reliability of the procedure, a comparative evaluation involving annotated resources was carried out on the tagger and three of the parsers.

Geospatial Visual Analytics of Traffic and Weather Data for Better Winter Road Management

Sapporo is a city with two million citizens that gets 6 m of snow per year. This means that winter road management is very important for sustaining economic and social activities during the winter. We believe that an exploratory and iterative analysis and visualization approach is useful to support the decision making, to improve the winter road management strategies. We propose using a huge library of tools and services, and a framework that allows users to freely federate tools and services improvisationally (“mash-up”) to create custom visualization and analysis environments and to apply these on appropriately selected data sets. Unlike conventional macro analysis approaches, we focus on micro analysis of winter road conditions. We use probe car data, speed readings etc., automatically collected from taxis and private cars. Geospatial visualization of the average speeds of all the road segments shows how different roads are affected by heavy snowfall, by snow plowing, and by snow removal. Combining geospatial visualization with knowledge discovery algorithms is a potential approach in this area. An example would be clustering the road segments based on similarity of the impact snowfall has to group roads into groups that can be maintained using similar strategies.

Comparing manual text patterns and machine learning for classification of e-mails for automatic answering by a government agency

E-mails to government institutions as well as to large companies may contain a large proportion of queries that can be answered in a uniform way. We analysed and manually annotated 4,404 e-mails from citizens to the Swedish Social Insurance Agency, and compared two methods for detecting answerable e-mails: manually-created text patterns (rule-based) and machine learning-based methods. We found that the text pattern-based method gave much higher precision at 89 percent than the machine learning-based method that gave only 63 percent precision. The recall was slightly higher (66 percent) for the machine learning-based methods than for the text patterns (47 percent). We also found that 23 percent of the total e-mail flow was processed by the automatic e-mail answering system.

Visual Data Exploration Using Webbles

We describe a system for visual exploration of data built using pluggable software components called Webbles. The system specifies a small common interface, a set of slots the plugins are expected to have, and any Webble following this interface can be plugged in at runtime. The system contains several types of visualization components, some built from scratch, some built by writing Webble wrappers for existing software, and some built by writing small interface wrappers for existing Webbles. The visualization components allow for interactive exploration of data, and selections or grouping of data in one visualization component are propagated to other components automatically. Interaction is done through direct manipulation of the visualization results.

Visualizing Clinical Trial Data Using Pluggable Components

We describe our system for visualizing data from clinical trials. The system is intended for clinicians with little or no knowledge of statistics, data mining, etc. The system is built using pluggable components, and it is easy to add more types of visualization. Currently the system supports interactive data exploration using for instance parallel coordinates, image maps, charts, and life tables. Grouping and filtering patients or subsets of data is easy and any changes are immediately reflected in all visualization tools currently used.

What types of translations hide in Wikipedia

We extend an automatically generated bilingual Japanese-Swedish dictionary with new translations, automatically discovered from the multi-lingual online encyclopedia Wikipedia. Over 50,000 translations, most of which are not present in the original dictionary, are generated, with very high translation quality. We analyze what types of translations can be generated by this simple method. The majority of the words are proper nouns, and other types of (usually) uninteresting translations are also generated. Not counting the less interesting words, about 15,000 new translations are still found. Checking against logs of search queries from the old dictionary shows that the new translations would significantly reduce the number of searches with no matching translation.

From Multiple Linked Views to Multiple Linked Analyses: The Meme Media Digital Dashboard

We describe a system called the Digital Dashboard that uses multiple linked views of data. All views allow interaction with the visualization results and interaction is done through direct manipulation. The system has been extended to allow new complex data to be generated in analysis components at runtime, e.g. By statistical analysis or data mining of parts of the data. The resulting data can be used in other linked views or analysis components, so when e.g. A data mining parameter is changed, all linked views (or analysis components) are automatically updated as soon as the new calculations are finished, and when something changes in linked components (e.g. A different subset of the data is selected), the calculations are automatically redone (if necessary).

Automated email answering by text-pattern matching: Performance and error analysis

Automated answering of frequent email inquiries can be implemented as a text categorization task with narrow text categories, where all messages in 1 text category have the same answer. Such email categorization should be optimized for high precision and at least acceptable recall. One such high-precision email categorization method is matching of surface text patterns to incoming email messages. In order to assess the upper performance limits of text-pattern matching, we conducted extensive tests with almost 10,000 messages. Our results show that automated email answering with precision around 90% and recall 50–75% is feasible. In order to achieve this, however, the system must work with multiword expressions rather than stand-alone words. Furthermore, we argue that the system has to distinguish the context of an email inquiry from the actual need that created the inquiry—a question, request, or complaint. We have discovered and analysed 12 reasons why text-pattern matching may fail.