Abstract
We have developed tools and applied methods for automated identification of potential news from textual data for an automated news search system called Software Newsroom. The purpose of the tools is to analyze data collected from the internet and to identify information that has a high probability of containing new information. The identified information is summarized in order to help understanding the semantic contents of the data, and to assist the news editing process.
It has been demonstrated that words with a certain set of syntactic and semantic properties are effective when building topic models for English. We demonstrate that words with the same properties in Finnish are useful as well. Extracting such words requires knowledge about the special characteristics of the Finnish language, which are taken into account in our analysis.
Two different methodological approaches have been applied for the news search. One of the methods is based on topic analysis and it applies Multinomial Principal Component Analysis (MPCA) for topic model creation and data profiling. The second method is based on word association analysis and applies the log-likelihood ratio (LLR). For the topic mining, we have created English and Finnish language corpora from Wikipedia and Finnish corpora from several Finnish news archives and we have used bag-of-words presentations of these corpora as training data for the topic model. We have performed topic analysis experiments with both the training data itself and with arbitrary text parsed from internet sources. The results suggest that the effectiveness of news search strongly depends on the quality of the training data and its linguistic analysis.
In the association analysis, we use a combined methodology for detecting novel word associations in the text. For detecting novel associations we use the background corpus from which we extract common word associations. In parallel, we collect the statistics of word co-occurrences from the documents of interest and search for associations with larger likelihood
in these documents than in the background. We have demonstrated the applicability of these methods for Software Newsroom. The results indicate that the background-foreground model has significant potential in news search. The experiments also indicate great promise in employing background-foreground word associations for other applications.
A combined application of the two methods is planned as well as the application of the methods on social media using a pre-translator of social media language.
It has been demonstrated that words with a certain set of syntactic and semantic properties are effective when building topic models for English. We demonstrate that words with the same properties in Finnish are useful as well. Extracting such words requires knowledge about the special characteristics of the Finnish language, which are taken into account in our analysis.
Two different methodological approaches have been applied for the news search. One of the methods is based on topic analysis and it applies Multinomial Principal Component Analysis (MPCA) for topic model creation and data profiling. The second method is based on word association analysis and applies the log-likelihood ratio (LLR). For the topic mining, we have created English and Finnish language corpora from Wikipedia and Finnish corpora from several Finnish news archives and we have used bag-of-words presentations of these corpora as training data for the topic model. We have performed topic analysis experiments with both the training data itself and with arbitrary text parsed from internet sources. The results suggest that the effectiveness of news search strongly depends on the quality of the training data and its linguistic analysis.
In the association analysis, we use a combined methodology for detecting novel word associations in the text. For detecting novel associations we use the background corpus from which we extract common word associations. In parallel, we collect the statistics of word co-occurrences from the documents of interest and search for associations with larger likelihood
in these documents than in the background. We have demonstrated the applicability of these methods for Software Newsroom. The results indicate that the background-foreground model has significant potential in news search. The experiments also indicate great promise in employing background-foreground word associations for other applications.
A combined application of the two methods is planned as well as the application of the methods on social media using a pre-translator of social media language.
| Original language | English |
|---|---|
| Journal | Journal of Print Media Technology research |
| Volume | 2 |
| Issue number | 3 |
| Pages (from-to) | 141-156 |
| Number of pages | 15 |
| ISSN | 2223-8905 |
| Publication status | Published - 7 Nov 2013 |
| MoE publication type | A1 Journal article-refereed |
Fields of Science
- 113 Computer and information sciences
- social media
- data mining
- topic analysis
- word associations
- 6121 Languages
- linguistic analsysis
- named entities