Detecting Priming News Events

With the increasing text information published on the news media and social network websites, there are lots of events emerging every day. Among these various events, there are only a few that are priming and can make significant changes to the world. In this paper, we measure the priming power of the events based on some popular index that people are usually concerned with.

These real world cases indicate that there is a need to find out the priming events which drive an interested time series index. Such priming events are able to help people understand what is going on in the world. However, the discovery of such priming events poses great challenges: 1) with the tremendous number of news articles over time, how could we identify and organize the events related to a time series index; 2) several related events may emerge together at the same time, how can we distinguish their impact and discover the priming ones; 3) as time goes by, how could we track the life cycle of the priming events, as well as their impact on the evolving time series index.

Some existing work has focused on discovering a set of topics/events from the text corpus [1, 6] and tracking their life cycle [13]. But these methods make no effort to guarantee these events are influential and related to the index that people are concerned with. There is another stream of work considering the relationship between the keyword trajectory and interested time series [9, 22, 18]. However, these work can only identify a list of influential keywords for users and do not consider to organize these words into some high level meaningful temporal patterns (i.e., events).

In this paper, we study the problem of detecting priming events from a time series index and an evolving document stream. In Figure 1, we take the weekly approval index of US President Obama from Jan 20, 2009 to Feb 28, 2010 as an example to illustrate the difficulty of this problem. In Figure 1, the approval index (blue line) evolves and drops from 67% to 45% in the last 56 weeks. Particularly, in July, 2009, the index dropped from 54% to 48%. For a user who is interested in politics and wants to know what event trigger this significant change, he/she may issue a query “President Obama” to the search engine. But the result will only be a list of news articles indicating the events that President Obama participates in during these periods. In Figure 1, we tag a small part of them on the index. As we can see, in July, 2009, there are 5 events including his attempt to reset relationship with Russia, help pool nations, visit Ghana etc. Only with this information, we cannot fulfill the user’s need since we cannot differentiate the role that each event plays to change the approval index in that time period. This urges us to think about the following questions. What makes an event priming? Does it contain some elements which will attract public eyes and change their mind? Besides, if such elements do exist, could we find their existence evidence from other time period and use them to justify the importance of the local event containing them?

In this paper, we call such elements influential topics and use them as basic units to form a priming event. Specifically, we identify the influential topics at a global level by integrating information from both a text stream and an index stream. Then at a micro level, we detect such evidences and organize them into several topic clusters to represent different events going on at each period. After that, we further connect similar topic clusters in consecutive time periods to form the priming events. Finally, we rank these priming events to identify their influence to the index.

The contributions of this paper are highlighted as follows.

• •

  To the best of our knowledge, we are the first to formulate the problem of detecting priming events from a text stream and a time series index. A priming event is an object which consists of three components: (1) Two timestamps to denote the beginning and ending of the event; (2) A sequence of local influential topic groups identified from the text stream; (3) a score representing its influence on the index.

• •

  We design an algorithm that first discovers the influential topics at a global level and then drills down to local time periods to detect and organize the priming events based on the influential topics.

• •

  We evaluate the algorithm on a real world dataset and the result shows that our method can discover the priming events effectively.

The rest paper is organized as follows. Section 2 formulates the problem and gives an overview to our approach. Section 3 discusses how we detect bursty text features and measure the change of time series index. Section 4 describes the influential topic detection algorithm and Section 5 discusses how we use the influential topics to detect and organize priming events. Section 6 presents our experimental result. Section 7 reviews some related work and Section 8 concludes this paper.

Let ${\cal D}=\{d_{1},d_{2},...\}$ be a text corpus, where each document $d_{i}$ is associated with a timestamp. Let $I$ be the interested index, which consists of $|W|$ consecutive and non-overlapping time windows $W$. $\cal D$ is then partitioned into $|W|$ sets according to the timestamp of the documents. Let $F$ be a set containing all different features in $\cal D$, where a feature $f\in F$ is a word in the text corpus.

Given the interested index $I$ and the text corpus $\cal D$, our target is to detect the priming events that trigger the movement of the index. As discussed in Section 1, the first step is to discover the influential topics. A possible approach [6, 10] is to first retrieve all the topics from the documents using the traditional approach in topic detection and tracking (TDT). Then we detect the influential topics by comparing the strength of the topic with the change of the index over time. We argue that this approach is inappropriate because the topic detection is purely based on the occurrence of the words and ignores the behaviors of the index. Consider the feature worker. Typical TDT approach would consider its co-occurrences with other features when deciding to form a topic. By enumerating all the possibilities, it will form a topic with features such as union because worker union frequently appears in news documents. However, if we take the presidential approval index into consideration, the most important event about worker related to the index would be that President Bush increased the Federal Minimum Wage rate for workers ever since 1997. This event helped him to stop the continuous drop trend of the approval index and make it stay above 30%. Therefore, it is more favorable to group worker with wage rather than with union. This example urges us to consider how to leverage the information from the index to help us organize the features into influential topics. These influential topics take in not only the feature occurrence information but also the changing behavior of the index. We formally define such topics as follows.

Based on the definition of influential topics, the next step is to represent priming events using these topics. One simple and direct way is to take each occurrence of a topic $\theta_{i}$ as a priming event. However, this approach has one major problem. We observe multiple topics at a time window $w$ and these topics are actually not independent but correlated and represent the same on-going event. Our topic detection algorithm may not merge them into a single topic because they only co-occur at that certain window but separate in other windows. For example, the topic of {strike target} would appear together with the topic {force troop afghanistan} in 2001. But in 2003, when the Iraq war starts, it co-occurs with the topic {gulf war} instead. Therefore, in order to capture such merge-and-separate behavior of topics, we define the local topic cluster as follows.

Based on the definition of local topic cluster, we further define a priming event as follows.

Our algorithm is designed to detect event with high priming event score and there are three major steps: (1) Data Transformation, (2) Global Influential Topic Detection, (3) Local Topic Cluster Path Detection and 4) Event extraction by grouping similar topic cluster paths. Details are given in the following sections.

In this section, we present how we transform and normalize the features $F$ and the index $I$.

Given the feature set $F$ and the probability volatility index $PVI$, our task here is to identify a set of influential topics $\{\theta_{1},...,\theta_{k}\}$, where each topic $\theta_{k}$ is formed by a set of keywords $F_{\theta_{k}}={f_{\theta_{k},1},...,f_{\theta_{k},|\theta_{k}|}}$. The problem can be solved by finding the optimal $\theta_{k}$ such that the probability of the influential bursty features grouped together is maximum for the text stream $\mathcal{D}$ and $PVI$. Below, we formally define the probability of $\theta_{k}$:

Since $P(\mathcal{D},PVI)$ is independent of $\theta_{k}$, we only consider the numerator of Eq. (3). We use the topic $\theta_{k}$ to account for the dependency between $\mathcal{D}$ and $PVI$. Therefore, given $\theta_{k}$, $\mathcal{D}$ and $PVI$ are independent. Our objective function then becomes:

Some observations can be made on Eq. (4). The second component $P(\mathcal{D}|\theta_{k})$ represents the probability that the influential topic generates the documents. And intuitively, we expect the document overlap of the features from the same topic to be high. The third component of $P(\theta_{k})$ represents the probability of the features to be grouped together. And two features should be grouped together if they usually co-occur temporally. Therefore, these two components basically require the features of $\theta_{k}$ to be coherent at both the document level and the temporal level. So generally, if more features are grouped together, the values of the second and the third components will decrease. And the first component $P(PVI|\theta_{k})$ represents the probability that the influential topic triggers the volatility of the time series index. Obviously, if the features in the group cover more windows with high volatility probability, the value of the first component will be higher. This will make the algorithm look for the features with high potential impact on the index. Below, we show how we estimate these three components.

First, we define the document similarity of two features using Jaccard Coefficient [3].

where $D_{f_{i}}$ is the document set containing feature $f_{i}$. Then the $P(\mathcal{D}|\theta_{k})$ can be estimated as below:

Second, in order to compute $P(\theta_{k})$, we estimate the temporal similarity of two features by comparing their co-occurrence over the whole set of windows $W$ as below:

where $p(f_{i})=\{p(f_{i},1),p(f_{i},2),...,p(f_{i},|W|)\}$ is the bursty probability time series of $f_{i}$ computed in Section 3.1. Then the probability of a set of features belonging to $\theta_{k}$ can be estimated by the average similarity for each pair of features:

Finally, in order to estimate the $P(PVI|\theta_{k})$, we define the influence probability for a feature $P(PVI|f)$ as:

Since the denominator of Eq.(9) holds the same for all the features, we just take the numerator for computation. And $P(PVI|\theta_{k})$ can be estimated as

Finally, the topic can be extracted using a greedy algorithm by maximizing Eq. (4) for each topic in a similar way as in [6].

However, Eq. (4) is different from the objective function defined in [6] since we extract topics with respect to an interested time series index $I$ rather than purely based on text documents. Consider the worker example again. The document similarity and the temporal similarity of worker and union are 0.31 and 0.25, while those of worker and wage are 0.35 and 0.1. If we do not consider the index $I$ by setting $P(PVI|\theta_{k})=1$, by Eq. (9), $P(worker,union|\mathcal{D},PVI)=0.31\times 0.25=0.0775$ and $P(worker,wage|\mathcal{D},PVI)=0.35\times 0.1=0.035$. As a result, the algorithm would combine worker and union. However, since $P(PVI|union)=30$ and $P(PVI|wage)=74$, by considering the feature influence to the index $I$, we have $P(worker,union|\mathcal{D},PVI)=0.0775*30=2.325$ and $P(worker,union|\mathcal{D},PVI)=0.035*74=2.59$. In this way, the algorithm will instead group $worker$ and $wage$ together to make an influential topic with respect to $I$.

As shown above, since influential topics carry the volatility index information, it brings benefits to the priming event detection. In the above example, if we detect a new event containing the common topic {worker, union}, the event may be trivial since the topic has a low influence probability in the history. But if we detect an event with {worker, wage} instead, this event has a higher probability to be a priming event. This is because the high influence probability of the topic indicates that the events with such topic attracted the public attention and changed people’s mind in the past.

After extracting each $\theta_{k}$, the bursty rate of $\theta_{k}$ in a window $w$ can be computed as below:

The bursty period of $\theta_{k}$ is determined in a similar way to detecting the bursty period of features in Section 3.1. And we use $\theta_{w}$ to denote all the bursty topics in window $w$.

A priming event $pe$ consists of three components: (1) Two timestamps to denote the beginning and the ending of the event in the window $W_{pe}$; (2) A sequence of local correlated topic clusters $C_{w},w\in W_{pe}$; (3) A score representing its influence on the index which is defined by considering the following three aspects of information:

1. 1.

   High event intensity $B_{pe}$: The event contains high bursty topics. The intensity at each window $w$ is estimated as follows:

   $$B_{pe,w}=\frac{1}{|\theta_{k}|}\sum_{\theta_{k}\in\theta_{pe,w}}P(\theta_{k},w)$$

   (12)

   And the intensity over the whole period is measured by $\|B_{pe}\|$ where $\|$ means taking 2-norm on the vector.

2. 2.

   High index volatility: The index must have high volatility during the bursty period of this event which is measured by $\|PVI_{pe}\|$.

3. 3.

   High index and event co-movement rate: The index and event intensity time series should be highly correlated which is measured by their linear correlation $Corref(B_{pe},PVI_{pe})$.

We combine these three measures and define the priming event score as below:

In this section, we proposed a two-phase algorithm to detect events with high priming event score. We first look for potential important topic cluster at each local time period by extracting and grouping bursty topics. (High event intensity). We then take the topic cluster at the period with most significant index volatility as seed and probe the whole priming event path (High index volatility and High index and event Co-movement rate). Below, we discuss the two phases of the algorithm in detail.

In this section, we evaluate our algorithm on real-world dataset from political and finance domain.

We archive the news articles from the ProQuest database¹¹1http://www.proquest.com/. In ProQuest, for political news, we take “President Bush” and “President Obama” as the query keywords and extract $15,542$ and $1,643$ news articles respectively from Jan. 1, 2001 to June. 1, 2010. For the financial news, we want to study the priming events during the financial crisis period, so we take “Finance” as the query keywords and extract $9,416$ news articles from June. 1, 2007 to June. 30, 2010.

For the preprocessing of these news articles, all features are stemmed using the Porter stemmer. Features are words that appear in the news articles, with the exception of digits, web page address, email address and stop words. Features that appear more than 80% of the total news articles in a day are categorized as stop words. Features that appear less than 5% of the total news articles in a day are categorized as noisy features. Both the stop words and noisy features are removed.

For the time series index, we archive the President Approval rating index from the Gallup Poll²²2http://www.gallup.com And the poll is taken every 10 days approximately. And for President Obama, we take a weekly average rating based on the Gallup daily tracking. For the financial index, we use the popular U.S. market index, the S&P 500 index and also take a weekly sample based on its daily index. We further identify the bursty probability time series and volatility time series according to the methods introduced in Section 3.1 and Section 3.2.

After these preprocessing, we have three real datasets.

• •

  Bush. It contains 1186 bursty feature probability streams and 1 volatility time series with equal length of 281 in the 8 years of Bush’s administration.

• •

  Obama. It contains 1197 bursty feature probability streams and 1 volatility time series with equal length of 56 in the 13 months of Obama’s administration.

• •

  Tsunami. It contains 961 bursty feature probability streams and 1 volatility time series with equal length of 156 in the 3 years of financial crisis.

We implemented our framework using C# and performed the experiments on a PC with a Pentium IV 3.4GHz CPU and 3GB RAM.

The problem of Topic Detection and Tracking (TDT) [1] is a classical research problem for many years. The first stream of work used graphical probabilistic models to capture the generation process of document stream [2, 7]. [20] extended the LDA model and incorporated location information. [21] analyzed multiple coordinated text streams and detected correlated bursty topics. [15] added the social network information as a regulation into the topic detection framework. On the other hand, there are another stream of work which detected topic based on the bursty features [11]. [6] detected bursty features and clustered them into bursty events. [4] further built an event hierarchy based on the bursty features. [10] analyzed the characteristics of bursty features (power and periodicity) and detected various types of events based on it. [13] proposed an algorithm to track short phrased and organized them into different news threads. Although these work can detect topics and track event efficiently. They can not tell the users which topics/events are changing the real world’s interested time series, President Approval Rating, Stock Market Index.

In contrast, there are some work studying the relationship between the text stream and interested time series. [12] studied the relationship between various topics over the volatility of presidential approval index. [9] attempted to find out the relationship between the online query and its related sales rank. [5] proposed a model for mining the impact of news stories on the stock prices, by using a $t$-test based split and merge segmentation algorithm for time series preprocessing and SVM for impact classification. [19] studied the relationship between the news arrival and the change of volatility of stock market index. [18] further explored this relationship to rank the risk of stocks. However, these work are either relying on analysts to extract topics manually or just identifying a list of influential keywords. In our work, we attempt automatically identify these events by first organizing these words into high level influential topics and then detect priming events based on them.

In this paper, we study the problem of detecting priming events based on a time series index and an evolving document stream. We measure the effect of the priming events by the volatility rate of the time series index. We propose a two-step framework to detect the priming events by first detecting influential topics at a global level and then forming priming events using detected influential topics at a micro level. The experimental results on the political and finance domains show that our algorithm is able to detect the priming events that trigger the movement of the time series index effectively.