Sentence Level Temporality Detection using an Implicit Time-sensed Resource

Sentence Level Temporality Detection using an Implicit Time-sensed Resource Sabyasachi Kamila, Asif Ekbal, Pushpak Bhattacharyya Department of Computer Science and Engineering, IIT Patna Patna, India {sabysachi.pcs16,asif,pb}@iitp.ac.in Abstract Temporal sense detection of any word is an important aspect for detecting temporality at the sentence level. In this paper, at first, we build a temporal resource based on a semi-supervised learning approach where each Hindi-WordNet synset is classified into one of the five classes, namely past, present, future, neutral and atemporal. This resource is then utilized for tagging the sentences with past, present and future temporal senses. For the sentence-level tagging, we use a rule-based as well as a machine learning-based approach. We provide detailed analysis along with necessary resources. Keywords: Temporal Sense Detection, Semi-supervised Machine Learning, Sentence Level Temporality Detection 1. Introduction Over the last few years, ‘temporality’ has drawn a signifi- cant attention to the community of Natural Language Pro- cessing (NLP) and Information Retrieval (IR). Time is an intrinsic property that aids in ordering events in a sequen- tial order from the past to present to future. This ordering of events is very crucial in analyzing a document. Some of the applications where temporality plays an important role include automatic summaries (Allan et al., 2001), question- answering (Schockaert et al., 2006), clustering (Alonso et al., 2009), similarity of documents (Jatowt et al., 2013) etc. Queries can specify temporality both explicitly and implic- itly. Queries like “World Cup 2011”, “Indian Prime Minis- ter 2000” etc. denote the temporality explicitly, whereas the queries like “Chomsky’s childhood”, “Recent Bolly- wood songs” etc. correspond to implicit temporality. All these highlight the significance of time in refining and rank- ing the results retrieved from a search engine. In a survey, Joho et al. (2013) claimed that most of the time user queries need to be addressed with recent infor- mation. However, many situations demand the past or fu- ture related information. For example, the query “�लोबल वा�म�ग क� वत�मान िःथित (globala vArmiMga kI varta- mAna sthiti - The current status of global warming.)”1 re- quires present related information whereas the queries like “�डिजटल अथ��यवःथा म� भारत के �लए अवसर �ा है? (Diji- Tala arthavyavasthA meM bhArata ke lie avasara kyA hai?- What are the opportunities for India in the Digital Econ- omy?)”, “अशोक का इितहास (ashoka kA itihAsa- History of Ashoka)” need future and past related information, respec- tively. Here, tense related information does not help but the implicit temporal keywords ‘current’, ‘opportunities’ and ‘history’ help in finding the temporal information of the re- spective queries. 1.1. Motivation and Problem Definition Most of the earlier studies, for example, TempEval tasks (Verhagen et al., 2009; Verhagen et al., 2010; UzZaman 1Henceforth, all the Hindi examples are represented by Hindi glosses, ITRANS representations and using equivalent English translation. et al., 2013) in the computational linguistics, have concen- trated on identifying the temporal expressions, event ex- pressions and various relations among these. These studies tried to address the temporal aspects of information with the help of linguistic …

Content truncated to 500 words