News articles are one of the most often read online documents, and the amount of news stories generated daily is quite large. News search engines are then often used to retrieve relevant news articles. The understanding of the returned results can be however impaired when the returned articles are about events being parts of complex or long stories. In this paper we discuss the dificulties resulting from the missing context and information complexity that users searching in temporal news collections may face. To support users in their search and learning we propose the concept of timeline as a retrieval and ranking unit of news search. We believe that automatically generating and ranking timelines could become an efective mechanism to facilitate search and browsing in large news collections.
chronic text collections such as collections of web pages or Wikipedia articles. In this context, we believe that the News are one of the most commonly read types of online unique temporal characteristics of news archives necesdocuments nowadays. They matter much to all users sitate novel methods for efective information retrieval. who want to understand key events in the world or in Professional users typically know what they wish to find their localities. In recent years, large amounts of news when searching in or interacting with news collections articles have been published, many of which are about and they also have suficient skills and knowledge of the complex and long events or stories. Thus specialized collections (or their covered time periods) to efectively news search engines are available to users helping them retrieve information. On the other hand, general users to find relevant information. may lack necessary knowledge and expertise to be able
IR technologies have been adapted to many diverse to perform successful searches in large news collections. domains including web search, legal document search, Searchers in large news collections face to some extent code retrieval, social media search, etc. [1]. We think that similar issues as users who search in collections connovel and efective developments in the field of accessing taining documents from unknown and complex domains temporal news collections such as news archives are also (e.g., medicine or law). The lack of knowledge causes difneeded due to large amounts of news generated these ifculties in understanding and interpreting search results, days and their complex characteristics. News articles updating further queries and continuing search sessions. are a highly temporal type of documents in which time This is especially a problem when the searched events ocsignals (whether in the form of timestamps or embedded curred in more distant past. Imagine a user who wishes to temporal expressions) are of key importance, and news learn about the progress of Iraq War - the event that took tend to be often understood chronologically or in causal place over 10 years ago. Even if the user managed to find order. While text indexing or query suggestion meth- all relevant news articles, she or he will have problem ods have been already studied in the context of temporal to fully understand them, arrange them into meaningsearch such as search in web archives or in news archives ful groups as well as continue the searching process by [2, 3, 4], still there is need for research on efective re- refining subsequent queries. First, this is because the trieval and ranking approaches. Currently, the usual user most likely does not know the context of the times retrieval approach seems to be basically one based on when the war happened and the relations between its applying the same access methods as for traditional syn- individual sub-events. Second, the news articles returned by a typical search engine would be probably ranked just DESIRES 2021 – 2nd International Conference on Design of by their relevance despite that the intuitive chronological Experimental Search & Information REtrieval Systems, September ordering is useful and intuitive for the user. The lack of 15–18, 2021, Padua, Italy chronological or causal order would prevent users from " adam.jatowt@uibk.ac.at (A. Jatowt) understanding the interrelations between events and the ~ h00tt0p0s-:0//0d0s1--i7n2f3o5rm-0a6t6i5k.(uAib.kJa.atco.watt/)(A. Jatowt) overall progress of the underlying story. Next, the user © 2021 Copyright for this paper by its authors. Use permitted under Creative would also not be able to find important articles as the CPWrEooUrckReshdoinpgs IhStpN:/c1e6u1r3-w-0s.o7r3g CCoEmmUoRns LWiceonsrekAstthribouptionP4r.0oIncteerenadtiionnagl s(CC(CBYE4U.0)R.-WS.org) retrieval method probably does not utilize any notion of event importance in an explicit way. Hence, less impor- conceptualizing or refining queries and let the user better tant news articles could be ranked highly with some of understand what she or he would like to actually retrieve them perhaps being returned at the top of the ranked or further explore. list, efectively preventing the user from understanding the key events of the war. Note that manually arranging articles into meaningful sequences would likely not be 3. Ranked Timelines as Output realistically feasible. This is because of the presumed lack of knowledge that an average user has about events in arbitrary time periods, and due to potentially large numbers of articles that can be returned, especially for longer events or stories.
Incorporating the notion of event importance into the ranking mechanism could let users receive news articles ordered not just by their relevance but also by the quantiifed importance levels of the described events. However, still this would not fully help users obtain large picture of all the major events related to their query, as the returned events would not be ordered chronologically. Furthermore, the set of ranked results could contain articles describing events that are parts of diferent stories or that are related to diverse aspects of the same story. This would likely happen for ambiguous queries such as location names or names of popular entities that took parts in diverse events (e.g., a name of a country’s president), or queries about complex and long-evolving events having diverse aspects (e.g., war in Syria).
TimeLine Summarization (TLS) research helps alleviate the problem of redundancy and complexity inherent in news article collections, thereby supporting users to better understand the news landscape. Usually a timeline is in the form of short descriptions of major events which are presented in chronological order, and it may also reflect causal relations. In traditional setting, the TLS approaches [5] work on a homogeneous type of datasets (i.e., collections of documents about the same event or about events of the same story) and produce just a single timeline as output. For heterogeneous document collections such as the set of relevant search results obtained in news search, one would need to apply the generalization of TLS.
Recently, Multi-TimeLine Summarization (MTLS) [6] has extended the typical TLS settings by allowing heterogeneous document collections as an input and by generating the output in the form of multiple timelines.
Outputting multiple timelines as a retrieval result in news search would be suitable to efectively organize and present search results, especially when the query is 2. Timeline for Providing Missing ambiguous, relates to complex or diverse stories, or the related news have multiple diferent aspects. The timeContext and Structuring Search lines would summarize diferent relevant stories or could Results reflect diferent aspects of the same story (e.g., timeline of the military actions, timeline of the economical aspects of We argue that presenting search results in the form of ten the Iraq war, timeline of the societal responses towards blue links ordered by the traditional notion of keyword the war, etc.). These timelines could be ranked based relevance would not sufice in many cases of news search. on their relevance degrees to the user query (e.g., comEspecially, in cases when (a) the sought news are parts of puted as the sum of relevance scores of their constituent longer ongoing stories, (b) the searched events happened events) or on their importance (e.g., computed as the sum in more distant past, (d) the issued queries are ambiguous, of importance scores of the constituent events, or simply or (c) when the user lacks necessary knowledge of event’s bound to the number of events, i.e., the timeline length). context, more efective solutions are required. We think In short, rather than ranking individual news articles, the that arranging the returned news results in the form of search engine would construct on-the-fly and rank timetimelines that summarize the key events and reflect their lines to be presented as ranked results to users. These chronological and/or causal order would be a more user- timelines would provide missing context to individual friendly way towards efective search in longitudinal events and help structure the entire news landscape that news article collections. In the above-mentioned example is relevant to user query. Finally, the timelines used as of the search intent of Iraq War the user could receive retrieval units could be then further expanded based on automatically generated timelines, instead of the usual user clicks to let the users view the detailed articles. This ranked list of news articles. Such timelines would be then style of result presentation bears resemblance to search the first step towards combating the lack of context on result clustering by their temporal aspects and ranking the user side which prevents successful search experience. in Web search as proposed by Alonso et al. [7]. After checking the timelines the user should be better Regarding the actual algorithm that could be used for informed of the event landscape (e.g. the progress of MTLS task, Yi et al. [6] have proposed a two stage Afinthe war represented by the sequence of its major events) ity Propagation approach to automatically generate the in order to perform further search. This would support set of timelines from a heterogeneous news collection. An advantage of that approach is that the number of the generated timelines does not need to be known or set beforehand, as it is dynamically obtained based on the underlying document collection. This flexibility is important in search scenarios as users can input arbitrary queries that could necessitate varying numbers of constructed timelines. search episodes).
We thank anonymous reviewers for their useful comments and suggestions. [1] W. B. Croft, D. Metzler, T. Strohman, Search engines: Information retrieval in practice, volume 520, In this position paper we propose considering timeline Addison-Wesley Reading, 2010. as an atomic retrieval1, ranking and presentation unit [2] K. Berberich, S. Bedathur, T. Neumann, G. Weikum, for news search. We think that incorporating Multiple A time machine for text search, in: Proceedings of Timeline Summarization methods as well as adding the the 30th annual international ACM SIGIR confertimeline ranking component would provide useful alter- ence on Research and development in information native to users besides the usual retrieval method that retrieval, 2007, pp. 519–526. centers on fine-grained retrieval units such as individual [3] N. K. Tran, A. Ceroni, N. Kanhabua, C. Niederée, documents. This novel approach should be especially use- Back to the past: Supporting interpretations of forful in cases of searching in longitudinal news collections. gotten stories by time-aware re-contextualization, Users could, for example, benefit from the interwoven in: Proceedings of the Eighth ACM International combination of providing timelines as search results and Conference on Web Search and Data Mining, 2015, of outputting the ranked news articles in a traditional pp. 339–348. way. For example, upon understanding the news land- [4] Y. Zhang, A. Jatowt, S. Bhowmick, K. Tanaka, Omscape based on the constructed timelines, the users could nia mutantur, nihil interit: Connecting past with switch to the usual document relevance-based search present by finding corresponding terms across time, in order to locate particular articles and satisfy detailed in: Proceedings of the 53rd Annual Meeting of the search needs. Timelines would then not only provide the Association for Computational Linguistics and the "global picture" useful for conducting search, but could 7th International Joint Conference on Natural Lanalso let users narrow down the results (e.g., by selecting guage Processing (Volume 1: Long Papers), 2015, pp. a sub-part of some timeline for subsequent search). 645–655.
There are many open problems that need to be ap- [5] D. G. Ghalandari, G. Ifrim, Examining the stateproached for using timelines as retrieval and ranking of-the-art in news timeline summarization, arXiv units including: (a) efective and eficient generation of preprint arXiv:2005.10107 (2020). timelines for arbitrary user queries, (b) timeline labelling2 [6] Y. Yu, A. Jatowt, A. Doucet, K. Sugiyama, and presentation3 for easy and quick result understand- M. Yoshikawa, Multi-timeline summarization (mtls): ing, as well as (c) the actual ranking mechanism for ar- Improving timeline summarization by generating ranging the generated timelines. multiple summaries, in: Proceedings of the Joint
Finally, worth investigation are user interaction mod- Conference of the 59th Annual Meeting of the Assoels and methods with search results composed of time- ciation for Computational Linguistics and the 11th lines. This encompasses timeline zooming in and out, International Joint Conference on Natural Language expanding timeline events to read their detailed articles, Processing, 2021. or switching between timeline search results and usual [7] O. Alonso, M. Gertz, R. Baeza-Yates, Clustering and search results. One can also imagine the possibility of exploring search results using timeline constructions, users updating/constructing their own timelines (e.g., by in: Proceedings of the 18th ACM conference on Inremoving irrelevant events/articles or adding relevant formation and knowledge management, 2009, pp. ones, perhaps the ones discovered through subsequent 97–106.
1Actually, to be correct, in the context of our proposal, the term "retrieval" should be substituted by "construction" or "generation"
2For fast overview of the ranked timelines, each timeline could have its generated label or short description
3Some of the constructed timelines may also contain shared events, the explicit indication of which could be perhaps added for more efective result presentation