Experiments on the tweet corpus Sample Clauses

Experiments on the tweet corpus. We then applied to the #Microposts2014 tweet dataset our observations on the Xxxx- Xxxx news corpus and realized that the results in detection degraded a lot, particularly with regard to recall (Table 3, first row). Clearly, the confidence threshold (which had been found best for news) was too high for tweets, combined with the NER filtering, so we experimented with other values and found a local optimum in 0.35 (results in second row), although the recall is still very low. We have tried to boost recall a little by combining the confidence and the condition of having been identified by XXX (Figure 5). It is still quite low, however, as often happens in real-world environments, false negatives (missed hits) are always preferable to false positives, which may make the application too noisy. So, we find that we need two different settings for the two types of text addressed by the EUMSSI project, which fits well into our implementation, since we already foresaw separate pipelines for each. Below we show the parameter combination for tweets: (confidence > 0.9 || !nertype.equals("") && confidence > 0.0.35) Figure 5: Tweets settings for NER-based filtered NEL Combination Precision in Detection Recall in Detection F-measure in Detection Accuracy in Classification of Detected Items Filtered DBSpotlight (news settings) 0.8732 0.1918 0,3145 0.90 Filtered DBSpotlight (tweet settings) 0.8449 0.275 0.4149 0.8445 Table 3: Results of re-ranking on the #Microposts2014 tweet dataset