Introduction.tex

\section{Introduction}
\label{sec:Introduction}
During this era of big-data and information flow, time is an important dimension to understand the sequence of historic events and their relations to the present or future events. Journalists, Anthropologists work with history all the time to understand the connections between different events. Events may convey correct or incorrect information with respect to temporal dimension. Moreover, the temporal dimension helps to distinguish current from outdated facts. The event \textit{``Germany holds the title of defending world champions in football"} may be true during the interval 2014-2018 but may not be true after 2018. The same can occur with presidents, spouses, CEOs and others.\\
A search query in the web may produce some relevant results but the temporal information of the query may be implicit or explicit. It has been reported that a significant number of searched queries in the web are filled with temporal intent. The work we are presenting here will help in understanding this intent and will provide useful temporal information based on the searched event-query.\\
Temporal information extraction \cite{uzzaman2012tempeval} focuses on recognizing and normalizing temporal expressions embedded in text into precise time points or intervals. The normalized time intervals are usually represented in a standard format like TIMEX3 \cite{uzzaman2012tempeval} of the TimeML markup language. Commonly, temporal expressions are categorized into four types: explicit, relative, implicit, and free-text. Most of the approaches \cite{alonso2011temporal,laparra2015document,strotgen2013multilingual} operate only on individual terms or phrases, and associating temporal information to larger textual units (like sentences or paragraphs).\\
Some approaches \cite{barzilay2002inferring} leverage the publication dates of the source news articles. This strategy however assumes that the documents are highly coherent and focus on a single time period indicated by their publication dates. \\
Currently, there is no way one can find out the probable years of an event's occurrence (more, with longer description) just by typing it in any search engine. So, it'll be a very useful tool to present to the user if he or she is searching an event query on the web.\\
The input is a news event query and output is a ranked list of associated years related to the query.\\
Challenges involved are sparsity of temporal information on the web, association of different events with a single time frame, and vice versa.\\
In temporal information retrieval, most approaches make use of the meta-data like creation time or publication dates associated with documents to identify their temporal scope. State of art research is purely based on statistics and graph based methods.\\
Our method works in semantic space, so works even if relevant data is not present. Neural embeddings haven't been used before in temporal embeddings. Semantic space association is a very new topic in this area of research. We can even find relevant contextual information apart from temporal information.\\
In section 2, we describe the related work done so far in this area. In section 3 we describe some basic concepts, our own approach to solve this problem. In section 4, we explain the basic tunings we have done in our experiments and analyze some per-query results among our own work. In section 5, we present the results we have achieved compared to the state of the art baselines. In section 6, we justify our results and lastly, in section 7, we present some future work that can be done using our method.
	\section{Introduction}
	\label{sec:Introduction}
	During this era of big-data and information flow, time is an important dimension to understand the sequence of historic events and their relations to the present or future events. Journalists, Anthropologists work with history all the time to understand the connections between different events. Events may convey correct or incorrect information with respect to temporal dimension. Moreover, the temporal dimension helps to distinguish current from outdated facts. The event \textit{``Germany holds the title of defending world champions in football"} may be true during the interval 2014-2018 but may not be true after 2018. The same can occur with presidents, spouses, CEOs and others.\\
	A search query in the web may produce some relevant results but the temporal information of the query may be implicit or explicit. It has been reported that a significant number of searched queries in the web are filled with temporal intent. The work we are presenting here will help in understanding this intent and will provide useful temporal information based on the searched event-query.\\
	Temporal information extraction \cite{uzzaman2012tempeval} focuses on recognizing and normalizing temporal expressions embedded in text into precise time points or intervals. The normalized time intervals are usually represented in a standard format like TIMEX3 \cite{uzzaman2012tempeval} of the TimeML markup language. Commonly, temporal expressions are categorized into four types: explicit, relative, implicit, and free-text. Most of the approaches \cite{alonso2011temporal,laparra2015document,strotgen2013multilingual} operate only on individual terms or phrases, and associating temporal information to larger textual units (like sentences or paragraphs).\\
	Some approaches \cite{barzilay2002inferring} leverage the publication dates of the source news articles. This strategy however assumes that the documents are highly coherent and focus on a single time period indicated by their publication dates. \\
	Currently, there is no way one can find out the probable years of an event's occurrence (more, with longer description) just by typing it in any search engine. So, it'll be a very useful tool to present to the user if he or she is searching an event query on the web.\\
	The input is a news event query and output is a ranked list of associated years related to the query.\\
	Challenges involved are sparsity of temporal information on the web, association of different events with a single time frame, and vice versa.\\
	In temporal information retrieval, most approaches make use of the meta-data like creation time or publication dates associated with documents to identify their temporal scope. State of art research is purely based on statistics and graph based methods.\\
	Our method works in semantic space, so works even if relevant data is not present. Neural embeddings haven't been used before in temporal embeddings. Semantic space association is a very new topic in this area of research. We can even find relevant contextual information apart from temporal information.\\
	In section 2, we describe the related work done so far in this area. In section 3 we describe some basic concepts, our own approach to solve this problem. In section 4, we explain the basic tunings we have done in our experiments and analyze some per-query results among our own work. In section 5, we present the results we have achieved compared to the state of the art baselines. In section 6, we justify our results and lastly, in section 7, we present some future work that can be done using our method.