Date: Mon, 05 Dec 2005 02:41:11 +0530 From: Anil Kumar Singh Subject: The Language of Time
EDITORS: Mani, Inderjeet; Pustejovsky, James; Gaizauskas, Robert TITLE: The Language of Time SUBTITLE: A Reader PUBLISHER: Oxford University Press YEAR: 2005
Anil Kumar Singh, Language Technology Research Centre (LTRC), IIIT, Gachibowli, Hyderabad-500019 (A.P.), India.
Natural language (NL) can convey all kinds of information about the world, abstract or concrete, real or hypothetical. Temporal information is one of the most important aspects of whatever we want to express using NL. This book presents a collection of some of the well known papers on temporal information in NL. The focus seems to be on the computational point of view. However, papers on linguistic and philosophical issues have been included to provide enough background. Since scholars and researchers from diverse fields (philosophy, linguistics, psychology, natural language processing, artificial intelligence) have contributed to the study of time in NL, the editors have made an attempt to represent all of these. The book can be useful for anyone interested in time and/or language.
The book is divided into four parts and there is a very informative introduction to each of these parts. These introductions make the papers accessible to even those without the required background as they summarize and connect the papers. Some readers who are looking just for a quick overview of the study of time in NL might want to read only the introductions.
In all, the book contains 29 important papers. These papers cover theories as well as computational approaches for studying (or computing) time in NL. People working in the relevant fields are already familiar with many of these.
The first part has 8 papers on tense, aspect and event structure. The second part is about temporal reasoning and it consists of 6 papers. The third part has 8 papers on temporal structure of discourse. The fourth part puts together 7 papers on temporal annotation.
Chapter 1 The first part starts with Vendler's well known paper ('Verbs and Times') on verb classification with respect to time (or tense). Vendler claims (a claim more or less accepted by many authors of the following papers) that almost all the verbs can be classified into a few classes, at least in their most dominant sense. The classes that he comes up with after a very interesting and enjoyable discussion are those of states, activities (or processes), accomplishments and achievements. This four way classification is the starting point of a lot of work done on time in language. Perhaps this is why Vendler's is the first paper in the book, although it is not the earliest among those included in the book.
Chapter 2 The second paper is on 'the syntax of event structure' by Pustejovsky. He argues that grammatical phenomena make reference to the internal structure of events, and that a subeventual analysis for predicates is able to systematically capture these effects. This paper represents a major stream in semantics. The purpose is to show that verbs decompose into distinct event types with internal structure. The primary components of an event structure are the event type (of the lexical item), the rules of event decomposition and the rules to map to lexical structure. Pustejovsky considers three basic event types: states, processes and transitions. He also relates event structure with lexical conceptual structure (LCS). Problems like adverbial modification are discussed in terms of event structure. On the whole it is a long but very readable paper which makes a convincing case for looking at verbs as events with an internal structure. As the editors point out in the introduction, this approach can avoid proliferation of primitives, which is a major problem with schemes like that of Dowty's (Dowty, 1979). A limitation is that event structure is not directly related with temporal structure.
Chapter 3 Pustejovsky's paper is followed by Emmon Bach's 'Algebra of Events'. The starting point for this paper is the close parallels between the mass count distinction in nominal systems and the aspectual classification of verbal expressions. Bach expresses one of the parallels as a proportion: events:processes::things:stuff, drawing on Link's paper on count-mass-plural domain (Link, 1983). His algebra deals with the distinctions based mainly on Carlson's classification (Carlson, 1981), which subdivides achievements in two categories (happenings and culminations, e.g. recognize and die). It also, in a way, takes into account the structure of events and processes. Link's scheme treats sets of individuals ('John and Mary') as super- individuals formed by a join operation. The super-individual is made of the stuff that the individuals are made of. Bach's contention is that telic events are also formed of process 'stuff'.
Chapter 4 The fourth paper is Reichenbach's classic work called 'The Tenses of Verbs' (Reichenbach, 1947). The primary idea is that the times of events can be located with respect to a deictic centre, which makes them similar to pronouns (the anaphoric view of tenses). In Reichenbach's scheme, there can be references to three time points: the speech time (S), the event time (E) and the reference time (R). These three time points can be related by 'precedes' or 'simultaneous' relations. Depending on the relations between S, E and R, we can define all the possible tenses. For example, E=Rpast, whereas S=R=E would be simple present. Only seven out of the thirteen possible relations are realized in English. This is perhaps the most influential paper in the book. This approach to analyze tenses has come to be known as the Reichenbachian approach. Quite apart from its influence however, the paper is remarkable for the way it is written. It is not divided into sections, there is no conclusion, nor are there any references. One wonders how the reviewers would react if a paper like this was submitted to a natural language processing (NLP) conference in this age of LaTex stylesheets. Anyway, I found this paper to be as enjoyable as it is influential.
Chapter 5 The next paper is 'Tense Logic and the Logic of Earlier and Later' (Prior, 1968). Prior presents a tense logic which is an extension of E. J. Lemmon's minimal tense logic Kt. He also relates it to the minimal calculus of earlier-later relation. He adds two operators G ('is always going to be') and H ('has always been') to the two operators provided by Kt: P (past) and F (future). The earlier-later calculus uses operators T and U such that 'Tap' means 'It is the case at instant a that p' and 'Uab' means that 'The instant a is earlier than instant b'. The tense logic is built upon in stages in the paper so as to get four logical systems of increasing power. This paper is also an important one, but it is not as easy to read as the some of the other papers in the book, partly due to the 'somewhat opaque prefix notation'.
Chapter 6 The paper by Moens and Steedman counters the idea that linguistic categories are related to a linear model of time. Their ontology is based on a mental representation of events structured on more than purely temporal principles. Their aspectual categories are a modified form of the Vendlerian classes. States are distinguished from events, which in turn can be classified in four categories depending on whether there are consequences and whether the events are atomic or extended. Aspectual type of a proposition can change due to modifiers. Such 'aspectual coercion' can be explained in terms of a transition network, which represents an aspect calculus. Another central notion in their ontology is that of 'nucleus' ('a structure comprising a culmination, an associated preparatory process, and a consequent state'). This nucleus plays a role in many of the permissible transitions in the network. However, as the editors point out, the paper doesn't offer rules for compositional semantics.
Chapter 7 Dorr and Olsen use LCS representation of Levin's classes. The aspectual classed are defined in terms of three features (telicity, dynamicity and durativity). In their system, features can only get added, not deleted. These aspectual features can be used to help in machine translation and generation.
Chapter 8 The last paper in the first part is by Passonneau, describing the PUNDIT information extraction system. She focuses only on actual (realis) events. A three step method is used to find the actual time associated with an event (if there is any), determine the temporal structure of the situation, and locating the situation with respect to the time of text production or to the times of other situations. The features used for events are situation type (state, process, or transition event), kinesis (active or stative) and boundedness (bounded, unbounded, or unspecified).
Chapter 9 The second part begins with a long but enjoyable paper (except perhaps the example about a Republican President) by McDermott. He presents a 'temporal logic for processes and plans'. One important idea in this logic is that of temporal chronicles ('a complete possible history of the universe'). The chronicles are arranged in the form of a tree in which branching occurs only towards the future. This is meant to model future possibility. McDermott defines events as sets of intervals over which a proposition is minimally (at least once) true.
Chapter 10 'A logic based calculus of events' is presented by Kowalsky and Sergot. They concentrate on applying event calculus for database updates as well as simple narratives. Instead of rejecting an update conflicting with the existing information, the update is accepted and the conflicting information is withdrawn. Some predicates introduced are 'Holds', 'HoldsAt', 'initiates', 'terminates' and 'happens'.
Chapter 11 Indeterminant temporal anchoring and granularity is taken into account by Chittaro and Combi in framework called 'Temporal Granularity and Indeterminacy Event Calculus (TGIC)'. Event calculus determines the maximal validity intervals (MVIs) over which properties hold. Their definition of MVI accommodates a more general concept of event.
Chapter 12 Here is another very influential paper, this time by James F. Allen. It presents a 'general theory of action and time'. This theory seeks to take care of actions that involve non-activity, actions not decomposable into subactions and actions which occur simultaneously and interact with others. Allen uses thirteen basic interval relations like 'before', 'after', 'meet', 'during', etc. There are three classes (property, event and process) and one metalanguage predicate (HOLDS, OCCUR, OCCURRING) associated with each of these classes.
Chapter 13 Galton's paper is a critical revision of Allen's work. He 'reinstates' instants to give a combined instant-interval scheme in which instants either fall within or limit intervals. One of the reasons for this revision of Allen's theory is to accommodate continuous change. The HOLD and OCCURS predicated are each split into three (HOLDS-ON, HOLDS-IN, HOLDS-AT).
Chapter 14 The paper by Hobbs and Pustejovsky on TimeML (originally DAML- Time), a markup language for annotating temporal information in a discourse. They also relate it to OWL-Time ontology. It takes into account the formal theories of time that have been suggested during the last many decades. This paper can be a good introduction to TimeML, which might become a standard.
Chapter 15 Though Dowty's 1979 paper is not included, there is one by him on 'the effect of aspectual class on the temporal structure of discourse'. His analysis is based on a compositional theory of aspect, i.e., aspectual classes of lexical items combine to give the aspectual classes of sentences. He also uses a pragmatic principle called Temporal Discourse Interpretation Principle (TDIP) which basically says that the reference time advances forward with each utterance, unless there is a temporal adverbial. He discusses some other pragmatic principles like the perspective of the narrator, expectations about discourse conventions (Grice's maxims) and background knowledge.
Chapter 16 Lascarides and Asher take further the work on including pragmatics in discourse (with respect to temporal relations). There work is more formal, based on the notion of 'defeasible reasoning' (based on default knowledge that can be overridden).
Chapter 17 In this chapter, Bell extends the Labovian analysis of temporal discourse structure. They focus on news stories and claim that the narrative in these stories can be segmented into 'abstract', 'attribution' and 'story', each of which can have further segmentation. The stories also contain many 'episodes', which in turn contain events.
Chapter 18 Here Webber extends the Reichenbachian notion of tense as discourse anaphor. She also related this to the work on discourse by Grosz and Sidner. The main elements of her approach are temporal focus, a tripartite ontology of events (preparation, culmination and consequence), 'specification' of entities by anaphors, and a 'focus stack' mechanism for resumption of dialog segments.
Chapter 19 Song and Cohen present an algorithm for using tense interpretation in analyzing simple narratives. They also use a modified Reichenbachian scheme for representing tenses which is more precise and unambiguous (uses one SRE triple instead of three structures for future perfect). Their algorithm is based on temporal focus, a tense hierarchy (for English) and constraints on coherent tense sequences.
Chapter 20 The automatic temporal reference resolution system described by Wiebe et al. works on the ambiguous output of a semantic parser. The domain they have considered is scheduling dialogs (''How about two?'', ''Twelve to two.''). In their model, a 'temporal unit' is associated with the current and the preceding utterance. A temporal unit has fields which are partially ordered with respect to specificity. Some resolution rules are also used.
Chapter 21 Hwang and Schubert use a novel 'fine structure' of discourse, namely 'tense trees'. Tense trees differ from simple Reichenbachian 'lists'. Given that the sentences are not 'flat' (as seems to have assumed by some, as far as studies of temporal structure in discourse are concerned), tense trees can be used to more effectively (compositionally) analyze tense and aspect.
Chapter 22 Hitzeman et al. use an HPSG implementation of discourse grammar for determining the temporal structure of the discourse. They try to take into account the mutually constraining effect of tense, aspect, temporal adverbials and rhetorical relations to reduce ambiguity. Another suggestion is to use underspecified representation of temporal or rhetorical structure, again for reducing ambiguity.
Note: The page numbers in the references given on the first page of the book for chapters 23-26 (from the ACL 2001 workshop) seem to be wrong. See the references below for the correct information.
Chapter 23 Wilson et al. present an annotation framework for automatically marking up temporal information. The annotation scheme is TIMEX2, which has been incorporated into TimeML. Their system tries to achieve cross-lingual reusability. Some updates since the publication of the paper are also mentioned.
Chapter 24 Another technique for temporal annotation is described by Katz and Arosio. In their scheme ('a radically simplified semantic formalism'), each verb is associated with a temporal interval and there are relations among these intervals. The relations are encoded with directed 'secondary edges' representing 'precedence' and 'inclusion' relations and their duals. They also discuss the relations between annotations, i.e., whether they are equivalent, consistent, or inconsistent, or whether one subsumes the other.
Chapter 25 Another temporal extraction system has been described by Filatova and Hovy. They break up news stories into clauses representing events and then assign time stamps to these events. Tense information in the clauses is used to help in timestamping.
Chapter 26 This paper (Schilder and Habel) describes a semantic tagging system for extracting temporal information from news messages. The tagger looks for dates, prepositional phrases and situational verbs. It marks chunks of texts containing temporal information and also tries to extract this information. Temporal relations (based on Allen's theory) are identified partly by prepositions.
Chapter 27 This chapter presents in detail the specification language TimeML. While the other paper (Chapter 14) is more theory oriented, this one is more about the definition of TimeML and guidelines for annotation using this language. It is compulsory reading for anyone involved with actual annotation of temporal information, or even with annotation in general.
Chapter 28 Li et al. present a framework for mapping linguistic patterns to temporal relations. It uses 'temporal concept frames' (activity related or time related), temporal relations (relative or absolute), temporal indicators (time related words), rules for temporal references and rules for rules for resolving conflicts. They have tried their system for Chinese with good results.
Chapter 29 The last paper (Setzer et al.) is about comparing and evaluating temporal annotations using the idea of 'temporal closure', which can help in deciding whether two annotations are equivalent or not. The paper also suggests ways to make manual annotation easier by computing closure of an annotation. A precise way of evaluating annotations is presented.
This book brings together a variety of approaches, theoretical as well practical, for dealing with time in NL. The papers are among the most relevant. They have been arranged in an order which makes sense. The introductions are excellent too. Perhaps the division into four parts and an introduction for each of them is best way to bring together a lot of diversity. However, one wonders whether a small introduction to each paper (instead of a long one to each part), together with a chapter in the beginning looking at the big picture wouldn't have been better. As it is, the readers might miss the connections between the successive chapters in some cases. And one can also complain about some missing papers (like Dowty, 1979), but that is unavoidable in any collection of papers. On the whole, it was rewarding to read this book. It might become a compulsory reading for people working in the relevant disciplines, and there are quite a few of them (people as well disciplines).
Dowty D. R. (1979). Word Meaning and Montague Grammar: the Semantics of Verbs and Times in Generative Semantics and Montague's PTQ. Reidel.
REFERENCES TO THE ORIGINAL PUBLICATION OF THE CONTENTS OF THIS BOOK
1. Vendler Z. (1967). Verbs and Times. Ch. 4 of 'Linguistics in Philosophy', 97-121. Cornell University Press.
2. Pustejovsky James (1991). The Syntax of Event Structure. Cognition, 41. Elsevier.
3. Bach Emmon (1986). The Algebra of Events. 'Linguistics and Philosophy', 9, 5-16. Swets and Zeitlinger Publishers.
4. Reichenbach Hans (1947). The Tenses of Verbs. Section 51 of 'Elements of Symbolic Logic', 287-298. The Macmillan Company, New York.
5. Prior A. N. (1968). Tense Logic and the Logic of Earlier and Later. Chapter 11 of 'Papers on Time and Tense', 116-134. Oxford University Press, Oxford.
6. Moens Marc and Steedman Mark (1988). Temporal Ontology and Temporal Reference. Computational Linguistics 14(2), 15-28. Association for Computational Linguistics.
7. Dorr Bonnie J. and Olsen Mari Broman (1997). Deriving Verbal and Compositional Lexical Aspect for NLP Applications. Proceedings of the 35th Annual Meeting of the Association for Computational Linguistics, 151-158. Association for Computational Linguistics.
8. Passonneau Rebecca J. (1988). A Computational Model of the Semantics of Tense and Aspect. Computational Linguistics, 14(2), 15- 28. Association for Computational Linguistics.
9. McDermott Drew (1982). A Temporal Logic for Reasoning About Processes and Plans. Cognitive Science 6, 101-155. Cognitive Science Society.
10. Kowalski Robert and Sergot Marek (1986). A Logic-Based Calculus of Events. New Generation Computing, 4, 67-94. Ohmsha Ltd.
11. Chittaro Luca and Combi Carlo (2000). Extending the Event Calculus with Temporal Granularity and Indeterminacy. In Bettini C. and Montanari A. (eds.), 'Spatial and Temporal Granularity: Papers from the AAAI Workshop'. Technical Report WS-00-08. The AAAI Press.
12. Allen James F. (1984). Towards a General Theory of Action and Time. Artificial Intelligence, 23, 123-154. Elsevier.
13. Galton Antony (1990). A Critical Examination of Allen's Theory of Action and Time. Artificial Intelligence, 42, 159-188. Elsevier.
14. Hobbs Jerry R. and Pustejovsky James (2003). Annotating and Reasoning About Time and Events. Proceedings of the AAAI Spring Symposium on Logical Formalization of Commonsense Reasoning. Stanford University, CA.
15. Dowty David R. (1986). The Effects of Aspectual Class on the Temporal Structure of Discourse: Semantics or Pragmatics? Linguistics and Philosophy, 9, 37-61. Swets and Zeitlinger.
16. Lascarides Alex and Asher Nicholas, (1993). Temporal Relations, Discourse Structure, and Commonsense Entailment. Linguistics and Philosophy, 16, 437-493. Kluwer Academic Publishers.
17. Bell Allan (1999). News Stories as Narratives. In Jaworski A. and Coupland N. (eds.), 'The Discourse Reader', 236-251. Routledge.
18. Webber Bonnie Lynn (1988). Tense as Discourse Anaphor. Computational Linguistics, 14(2), 61-73. Association for Computational Linguistics.
19. Song Fei and Cohen Cohen, (1991). Tense Interpretation in the Context of Narrative. Proceedings of the Ninth National Conference on Artificial Intelligence (AAAI-91), 131-136. American Association for Artificial Intelligence.
20. Wiebe Janyce, O'Hara Tom, McKeever Kenneth, Ohrstrom- Sandgren Thorsten (1997). An Empirical Approach to Temporal Reference Resolution. Proceedings of the Second Conference on Empirical Methods in Natural Language Processing (EMNLP-97), 174- 186. Association for Computational Linguistics and SIGDAT.
21. Hwang Chung Hee and Schubert Lenhart K. (1992). Tense Trees as the Fine Structure of Discourse. Proceedings of the 30th Annual Meeting of the Association for Computational Linguistics, 232-240. Association for Computational Linguistics.
22. Hitzeman Janet and, Moens Mark and Grover Claire (1995). Algorithms for Analyzing the Temporal Structure of Discourse. Proceedings of the Annual Meeting of the European Chapter of the Association for Computational Linguistics, 253-260. Association for Computational Linguistics.
23. Wilson George, Mani Inderjeet, Sundheim Beth and Ferro Lisa (2001). A Multilingual Approach to Annotating and Extracting Temporal Information. Proceedings of the ACL 2001 Workshop Temporal and Spatial Information Processing, 81-87. Association for Computational Linguistics.
24. Katz Graham and Arosio Fabrizio (2001). The Annotation of Temporal Information in Natural Language Sentences. Proceedings of the ACL 2001 Workshop Temporal and Spatial Information Processing, 104-111. Association for Computational Linguistics.
25. Filatove Elena and Hovy Eduard, (2001). Assigning Time-Stamps to Event-Clauses. Proceedings of the ACL 2001 Workshop Temporal and Spatial Information Processing, 88-95. Association for Computational Linguistics.
26. Schilder Frank and Habel Christopher (2001). From Temporal Expressions to Temporal Information: Semantic Tagging of News Messages. Proceedings of the ACL 2001 Workshop Temporal and Spatial Information Processing, 65-72. Association for Computational Linguistics.
27. Pustejovsky James, Ingria Robert, Sauri Roser, Castano Jose, Littman Jessica, Gaizauskas Rob, Setzer Andrea, Katz Graham, Mani Inderjeet (2004). The Specification Language TimeML. This volume, 545-557. Oxford University Press.
28. Li Wenjie, Wong Kam-Fai and Yuan Chunfa (2001). A Model for Processing Temporal References in Chinese. Proceedings of the ACL 2001 Workshop Temporal and Spatial Information Processing, 33-40. Association for Computational Linguistics.
29. Setzer Andrea, Gaizauskas Robert and Hepple Mark (2003). Using Semantic Inference for Temporal Annotation Comparison. Proceedings of the Fourth International Workshop on Inference in Computational Semantics (ICOS-4), 185-196.
ABOUT THE REVIEWER:
ABOUT THE REVIEWER
Anil Kumar Singh is working towards PhD in Computational Linguistics at the Language Technology Research Centre (LTRC), IIIT, Hyderabad, India. His research interests are in dealing with temporal information in NL, statistical NLP, corpus linguistics and, last but not the least, NL engineering. But there is much more to life than research interests.