A Rule-Based System for Monitoring of Microblogging Disease Reports

People from all around the world use microblogging services on a daily basis and send messages, among others, about their current health condition. Those behaviors and tools on the Web provide us with the perfect technological-sociological background to develop a real-time disease and epidemic outbreak surveillance system. It would monitor the Twitter message data stream, decide if certain posts can be considered disease reports and cluster the appropriate ones based on the sender’s geographic whereabouts.

To motivate the underlying problem which we address, we present a couple of example messages which come from a Twitter data stream:

are messages that should be considered disease reports, whereas

are messages that definitely should not be considered disease reports.

In this paper, we present a demonstration^{Footnote 1} of our rule-based approach that can distinguish the disease reports from other microblogging messages containing the disease-related keywords.

Typical obstacles for classifiers of microblogging messages using standard machine learning approaches are:

Short Text Messages: Due to the length of the microblogging messages, the established text classification approaches, such as Naïve Bayes, turn out not to be very effective. Given the fact that the microblogging messages are rather short (140 characters is the Twitter limit), the training data set would have to be enormous for the algorithm to be able to distinguish the messages properly.

Term Frequency Equals Document Frequency: Again because of the length limit the terms are most often used only once in the messages which might make this numerical statistic not very meaningful.

Lack of the Learning Curve: The microblogging messages are not per default labelled regarding their disease relevance and the classifier does not get any feedback during its lifetime unless the messages would be marked manually – in the contrary to, e.g., spam filters, where the users help the system to learn and improve it by saying that this email is or is not spam. As a result there is no relation between experience of the classifier and its performance.

Specific Language: The colloquial Internet and microblogging diction differs from high language standards. This slang is often influenced by current offline events and thus constantly evolves requiring re-training the classifier.

The above mentioned problems clearly show that the existing text classification approaches cannot achieve high precision in the classification of the microblogging messages. Therefore we propose a new, content-based approach which performs the analysis and classification of short, specific text messages.

A classifier is required that can assign the microblogging messages to one of three classes: disease report (\(\varvec{DR}\)), no disease report (\(\varvec{NDR}\)) and possible disease report (\(\varvec{PDR}\), which means that the poster’s intentions were not clear).

We propose a rule-based classification that can classify each message by triggering multiple rules which are applied in a sequential order. A specific score is assigned to each rule and in case that it matches when analyzing a message \(m\), we add the rule score to the overall score of \(m\). At the end, we compare the calculated score to the fix thresholds and decide which class should a message belong to. The thresholds, a pair \((t_n, t_p)\), should be interpreted as follows:

Our rules are grouped in several categories. Provided that a message contains a disease-related keyword, it is processed by the following content analysis rules:

Category-1: Rules Based on Frequent Words: In the first step, we analyze the messages considering the most common words appearing in the disease report candidates.

We have set up a service which filters the Twitter stream based on the set of predefined keywords and stores the messages locally. We collected over 300 million candidates for disease reports between Jan \(16^{th}\), 2013 and Aug \(17^{th}\), 2013 and created taxonomies for each keyword by summing up the term frequencies of all words occurring in the messages containing this keyword. The stop words are removed and the remaining tokens are stemmed.

We analysed the lists of frequent words and manually assigned a classification sentiment for every \(keyword \rightarrow word\) collocation. For this we used the Twitter search engine to obtain sample messages containing both words and, based on the context, assigned values from the range \([-5, \ldots , 5]\), with -5 meaning a definitely disease-related and 5 a definitely not disease-related collocation. Having that, we find all keywords in an analysed message and look for sentiments in the corresponding taxonomies. This way we calculate the disease score of a single message.

Category-2: Rules Based on the Types of Detected Named Entities: We use DBPedia Spotlight [8] for the extraction of semantic resources from the text. Other existing public services like AlchemyAPI, OpenCalais and Zemanta cannot be used because of heavy processing load of the message stream. We use an internal mirror of the DBpedia and the DBpedia spotlight on a cluster of hosts. After the identification of the semantic sources, we query their types, like rdf:type (dbpedia-owl:Disease), dbpedia-owl:type to check the type of the recognized semantic concepts. This approach lets us derive more meaning from the tweets, e.g., take disease-related words from outside our taxonomies or the synonyms into account. Let us consider the following tweet: In this example the message that would not be classified as disease report without the help of semantic concepts. The word happy would be the only match in the sentiments list (see Category-3) and no collocation from the taxonomy of fever would be found. DBPedia Spotlight however annotates “glandular fever” (dbpedia:Infectious_mononucleosis) and “Ross river fever” (dbpedia:Ross_River_fever) as diseases.

Category-3: Rules Based on General Mood of Messages: To improve the classification precision, we apply further rules to extract the sentiment of the message.

We use the word list by Hansen et al. [3] to look up the message’s tokens and thus calculate its general mood. Unfortunately, the disease report candidates generally tend to have a negative score calculated by this rule because of the diction that consists mainly of words with negative sentiment.

Category-4: Other Rules: The microblogging messages are often enriched with emoticons (also called smileys, conventional symbols for expressing emotions) to put emphasis on the author’s mood. For that reason we check the tweet for the presence of smileys.

Furthermore, when people tell about their illness when having a fever, they sometimes mention its height. We could assume that a number found in a message (restricted to ranges which correspond to fever temperatures, both in Fahrenheit and Celsius scale) means with high probability that it concerns a disease.

The studies by Signorini [10] and Chew [2] exemplarily explain how massively the Twitter data stream is influenced by current real life events. Among many other triggers, like for example the American Idol contest, the authors of these publications took a look at the 2009 Swine Flu (H1N1) outbreak.

Stewart and Diaz [11] present different approaches of health-related Twitter surveillance. Discussing Early Warning, as well as Outbreak Control and Analysis Systems, they introduce several biosurveillance algorithms and techniques (Khan [6]; Hutwagner et al. [5]; Basseville and Nikiforov [1]) and use them to analyse the crowd’s behavior during the 2011 enterohemorrhagic Escherichia coli (EHEC) outbreak in Germany. Their main focus was to detect aberration patterns when the observed variable (here: tweets containing the “EHEC” keyword but more generally: the number of tweets regarding diseases that do not reveal seasonal patterns) exceeds an expected threshold value. They used four different biosurveillance algorithms for early detection, each one of which proved to be at least one day faster than well-established early warning systems, like e.g. The Early Warning and Response System of the European Union^{Footnote 2}, MedISys^{Footnote 3} or ProMED-mail^{Footnote 4}.

Lampos and Cristianini [7] investigate the 2009 Swine Flu outbreak. Hu et al. [4] aim to cluster Twitter messages by topic and extract meaningful human-readable labels for each cluster. They decompose the unstructured text using NLP and then transform the syntactic feature space (parse [sub]trees) into semantic feature space using WordNet and Wikipedia. Saif et al. [9] propose to add the semantic concepts of extracted entities as additional features for sentiment analysis.

Hansen et al. [3] analysed which tweets attract the biggest attention and are most likely to be retweeted. As a part of this publication one of the authors, Finn Årup Nielsen, prepared a list of 2477 English words rated for valence with an integer between minus five (negative) and plus five (positive).

The main difference of our approach with the existing approaches are that our approach is based on the collected heuristics, manually added sentiment scores and the semantic types of the extracted entities in the messages. Its advantage is that it does not require to have a learning loop (only an update of the keyword list / taxonomies) and has no shortcomings when applied to the very short messages.

On this basis, we demonstrate a real-time system for classification of disease messages from mass of microblogging messages. Our system is a live service connected to the Twitter stream that receives messages and visualizes the disease reports on a map provided that they were enriched with the geographic whereabouts of the sender (sent from a mobile device).

Given that our system could receive a complete data stream, after some time it could extract anomalies based on the daily/weekly number of messages originating from a certain area. Such a tool could be a great complement to the well-established health surveillance systems.