Verbs tenses was the first idea that came into my mind to specify the time of a news article, however, working on multiple sentences level makes huge amount of ambiguity, specially that the articles could be mentioning a history of some named entity (person, country, organization) but then they could end the article with a simple prediction. In addition, this task could only give simple prediction of (past, future, or current time). Using verb tenses seems to be useful on sentence level, but it also faces more ambiguity with long sentences such as: “He kept failing since 2010, but now he achieved success”.
Therefore, the idea of predicting text time by using verb tenses should not be considered as a main approach. Instead, it could be used as an auxiliary task to help other noun-based method. It is also possible to try deep learning model to predict time of the text, but it is too difficult to get an annotated text from the same domain to extract time from verb tenses only.
A model such as LSTM could rely instead on nouns or on date formats, trying to achieve the prediction by using verb tenses only may require a special corpus that contains only verbs. The success of such model is under doubt to work on paragraph or article level, it could be successful only on sentence level. But again, the prediction of sentences tenses may not be useful alone for the majority of the tasks.
The common approach of predicting time of the text is to rely on a set of regular expressions to get numbers and date formats. DARPA and other institutions introduce such an approach since 90s. TIMEX2 system is one of these systems that gives XML tags to the text based on two categories of indicators, the triggering indicators and non-triggering indicators. The first group could be described as a direct indicator of date such as (yesterday, Tuesday, 10-10-2000) while the other one (the non-triggers) such as (on, in, when) may lead to triggers but it cannot be used directly nor it is used specifically with triggers.