<?xml version="1.0" standalone="yes"?>
<Paper uid="C88-2087">
  <Title>A Novel Analysis of Temporal Frame-Adverbials</Title>
  <Section position="1" start_page="0" end_page="429" type="metho">
    <SectionTitle>
Internet: MME@IDA.LIU.SE
Abstract
</SectionTitle>
    <Paragraph position="0"> In this paper interpretation principles for simple and complex frame-adverbial expressions are presented. Central to these principles is a distinction between phases and periods together with the temporal hierarchy, where multiple scales of time and relations can be expressed. A system, CLOCKWISE, has been implemented which interprets Swedish temporal expressions according to the principles outlined in the paper.</Paragraph>
    <Paragraph position="1"> Introduction Temporal information is expressed and conveyed in a number of ways in natural language including tense, aspect and lexical items that carry temporal information, eg. temporal adverbs. Most researchers in this field approach temporal entities in language from the perspective of tense and aspect. But there is very little in the literature on other expressions that hold temporal information, such as temporal adverbs, certain prepositional phrases and noun phrases. In most papers the meaning of a temporal adverbial such as 'next year' is merely explained as the predicate 'next year' which specifies a point or interval of time from a reference time. In objective time-modelling systems such as/Kahn &amp; Gorry 1977/and/Bruce 1973/ temporal expressions were never analyzed in their linguistic form; instead they had to be typed in as stereotyped lists. The internal structure of temporal expressions must be investigated in order to construct grammars that can capture general features and be of use in computational applications. In the paper I will focus on temporal frame-adverbial phrases, that is, expressions that refer to a temporal period in which events are located (Cf /Bennet&amp; Partee 1978/,/Hinrichs 1986/). I will not discuss the complex question of how tense, aspect and temporal adverbials interact (for an outline of the problems see/Ejerhed 1987D.</Paragraph>
    <Paragraph position="2"> TemPoral frame.adverbial phrases A large group of temporal expressions can be classified as frame-adverbial phrases. /Smith 1981/ categorizes temporal frame-adverbial phrases in the following way1:  Deictic Clock-Calendar 2 Dependent - last week, yesterday - at midnight = now, this moment + next week, tomorrow + at midnight - previously, before = the same time + later, afterwards 1 My use of + and - is the same as Smith's forward and backward arrows.</Paragraph>
    <Paragraph position="3"> 2 In an earlier paper (Smith 1980) Smith calls this class flexible  anchoring adverbials.</Paragraph>
    <Paragraph position="4"> This classification is based on the concept of anchoring, ie. a deictic is generally anchored to the time of speech (ST), a dependent is anchored to another given time in the context, and a clock calendar adverbial can anchor to either ST or to some other context-dependent time. Smith attributes a relational value to these expressions, -, = and +. These symbols stand in turn for the relational value anteriority (past), simultaneity (present) and posteriority (future).</Paragraph>
    <Paragraph position="5"> Frame-adverbials can syntactically occur as adverbs, noun phrases and prepositional phrases. They can be complex expressions such as  (1) On Monday next week at 6 pro.</Paragraph>
    <Paragraph position="6">  where temporal information is specified on several levels and we have a combination of deictic and clock calendar expreso  sions, Complex expressions have a relatively loose syntax in that the time denoted in (1) can be expressed as (2) and (3) (2) Next week on Monday at 6 pm.</Paragraph>
    <Paragraph position="7"> (3) At 6 pm next week on Monday.</Paragraph>
    <Paragraph position="8">  The issues that need considering are the following: How is the semantic well-formedness of frame-adverbial phrases determined and, if possible, what is needed to establish their temporal reference in terms of locations on the time axis. Periods and phases To account for the semantic well-formedness of flame-adverbial phrases the 'time' denoted by the phrase can be thought of as an ordered set of specifications for conventional temporal concepts, such as YEAR, MONTH, DAY, HOUR, MINUTE, SECOND. The ordering of these temporal concepts may seem to fit well into a 'Chinese box metaphor' where each concept is ~ included in a concept on a higher level and in turn includes a concept on a lower level. A semantically well-formed temporal frame-adverbial would then be described as a phrase which forms a description of an unbroken chain of temporal concepts. This is put forward by/Hinriehs 1986/in his Scoreboard of reference points and cooccurence restriction in which the construction of a reference time from a flame-adverbial phrase is possible if there is a chain of specified temporal concepts in accordance with an ordered set of concepts. However, if we included WEEK in the above set of concepts we would get into trouble. Weeks are not included in months in the same way as months are included in years and a problem of where to fit in WEEK in the Chinese box order would occur. This is due to the fact that the western calendar system is based on three sepao rate descriptive traditions: the week, the Gregorian calendar and the clock/Levy 1980/. The Gregorian calendar holds year, month and day(of month) as its basic concepts, the week is a sequence of named days, and the clock system divides the day into hours and hours into seconds, etc. The week and Gregorian calendar share the concept of day and the clock  system split.~ the day into time segments. The reason that we can express a time by specifying different concepts of separate systems is that the systems share one or more concepts with each other, tlt is also reflected in the way we refer to the day period (in S ~vedish and English). A day can be referred to in two different ways, either by specifying the day of a certain week (DW) or by picking out a day of a certain month (DATE). In English the expression on Monday next week will describe the same day on the time axis as an expression describing the same day by specifying a year, a month and a day of month. In Swedish it is possible to specify a mfique day by the yearweek-day path as the Swedish calendar has numbered weeks, giving us a year divided into 53 weeks 3.</Paragraph>
    <Paragraph position="9"> In our model we want to distinguish between periods and phases and ~;how the relations between them.</Paragraph>
    <Paragraph position="10"> A period P is a segment of time with a certain length, such as YEAR, MONTH, DAY, WEEK, HOUR, MINUTE, SECOlqD.</Paragraph>
    <Paragraph position="11"> A l)eri~xl Pi is a subperiod of a pecind P iff 1&amp;quot; can be seen as consisting of a number of periods Pi that togeflmr exhaust I'. A phase p is a specific instance of a snbperiod having a eeltain order ia tire sequence of subperiods cxhaustillg a given period P.</Paragraph>
    <Paragraph position="12"> For example, a period such as YEAR can be seen as consisting of twelw+ subperiods (MONTH). January is the first phase of those twelve subperiods.</Paragraph>
    <Paragraph position="13"> The graph below illustrates the relation between year as a period and the monthly phases.</Paragraph>
    <Paragraph position="14"> '~ ........ YEAR .....</Paragraph>
    <Paragraph position="15"> Fig. 1 The YEAR period This graph shows the year period from one point of view. We could split the year into other phases, for' instance, into periods of seasons, where summer, autumn, winter and spring would be the corresponding phases. Note that a sequence of phases is cyclically lecurring.</Paragraph>
    <Paragraph position="16"> Apart fi'c,m differentiating periods and phases we must be able to show the ,'elations between them, The following table shows some period:; with corresponding phases:  file, use of named months or named days. The analogy does not ho:Ld Idt the way of course as the week system is historically indel~mdent from the Gregoriau calendar, but we have 'tried' to force the week concept onto the concept of year by naming the weeks with numbers. In Sweden a umubered week always slarls on a Monday. The first week that holds at least four days of a new year forms week 1. Tire first of January 1987 was a qhursday and pint of week 1, whereas the first of January 1988 is a Friday which :esults in the first three days of 1988 being part of week 53 19811 This is standardized by Swedish authorities.</Paragraph>
    <Paragraph position="17"> Fox each period there are one or two possibilities to view it as a subpefiod of a higher-level period. The difference between the two different ways of referring to a DAY period is shown by the fact that there are two subperiod-of relations, one to MONTH, and one to WEEK.</Paragraph>
    <Paragraph position="18"> The subperiod-of relations can be displayed in a temporal hierarchy of lhe following kind.</Paragraph>
    <Paragraph position="19">  The lines connecting temporal periods to each other show subperiod-.of relations. The hierarchy shows the two diflerent paths fi'om YEAR to DAY.</Paragraph>
    <Paragraph position="20"> :The periods and phases described above share one property and that is that they all have fixed lengths and as a consequence clear boundaries. We can easily locate the point on the time axis where January becomes February, but we cannot with accuracy specify the location on the time axis where spring is succeeded by summer. So we must distinguish be- tween bounded and fuzzy periods. There are other phasal divi-sions one can make. For example, we could split the week into weekdays and non-weekdays (week-ends?), and tim year :into feasts, such as Christmas and Easter, and non-feasts, etc. It is just a matter of fact that we divide time into different temporal scales, and some contain periods of fixed length and some do not. This is why it is problematic to try and map a phase of no fixed length, such as summer, onto another temporal scale where we are dealing with fixed boundaries sueb as the one depicted in the temporal hierarchy above.</Paragraph>
    <Paragraph position="21"> Durative expressions are mostly expressed in terans of quanti-fied periods: one week, two months, three years, etc. It is more seldom that phasal expressions are used in a durative sense. Quantified phasal expressions such as two mornings, six Thursdays, three winters, etc are primarily used to express habituality and iteration. It is however possible to infer duration fl'om these as we know the length of a phase or can infer it from its corresponding period.</Paragraph>
    <Paragraph position="22"> Phasal and deictic expressions In accordance with the previous section we can now distinguish between phasal and deictic expressions.</Paragraph>
    <Paragraph position="23"> In our terminology a phasal expression is a temporal expression that primarily describes a phase by using a corresponding word for the phase in question. A phasal expression may be, a single word (1988, January, midnight) or be more complex, such as a prepositional phrase (in 1988, in January, at midnighO.  The difference between Smith's clock calendar adverbials /Smith 1981/ and phasal expressions is that phasal expressions are just &amp;quot;providers of phases&amp;quot; and they are not primarily anchored to any other time as clock calendar adverbials are. Phasal expressions may describe absolute times if they refer to a unique phase on the time axis. So if the phasal expression 1987 denotes a unique phase of the calendar then it is possible to map it into an absolute time. Following/Hinrichs 1986/we call these phasal expressions complete or independent dates. Also included in this class we find complex phasal expressions which, if mapped on to the temporal hierarchy, would form a chain of temporal information. By chain of temporal information we mean a chain of temporal phases according to the temporal hierarchy.</Paragraph>
    <Paragraph position="24"> Deictic expressions function basically in the following way: Starting from a reference time (eg. ST) a deictic will establish a new reference time with a different phasal value for the relevant period by moving a number of steps in either a forward (+) or a backward (-) direction depending on the deictic, and keeping phasal values for higher-level periods consistent with this change.</Paragraph>
    <Paragraph position="25"> Certain deictics do not change the reference time. For example. i dr (this year), den h?ir veckan (this week), i dag (today), etc, do not move the reference time in a forward or backward direction. The new RT will be a partial copy of the old RT; today will describe a time equal to ST and specified to the DAY period level.</Paragraph>
    <Paragraph position="26"> The deictic expression n u (now) is ambiguous in respect of period level. In one context nu would map onto the year level and in another context it would refer to the hour level of RT. In our system the function that maps a deictic expression is of the form F (x, index, step) where x is either a phase or a period. Applying F results in a description of the time in terms of a set of phases which forms a chain in the temporal hierarchy with x as its most specific phase or period. Posterior deictics will have positive numbers as step, anterior will have negative numbers and simultaneous will have 0. For example, a deictic such as i morgan (tomorrow) will be mapped to a set of corresponding phases from the YEAR period down to the DAY period, leaving more specific phases unspecified. This means that while constructing a description of a time unnecessary work is avoided as we are dealing with partial descriptions and not the  temporal objects as such. Compare the following two phrases: (1) tomorrow (2) tomorrow at 3 o'clock.</Paragraph>
    <Paragraph position="27">  In both expressions tomorrow will contribute with exactly the same information, ie. a description of a certain day of a certain month in a certain year, etc.</Paragraph>
    <Paragraph position="28"> Deictics in Swedish can be categorically characterized in the following way: &amp;quot;. Lexical deicties (such as igdr (yesterday), fjol (last year) idag (today).</Paragraph>
    <Paragraph position="29"> deg Pre/postpositional phrases with a temporal NP where the pre/postposition signals the type of temporal expression. For example, am tvd veckor (in two weeks), f~r fem dr sedan (five years ago).</Paragraph>
    <Paragraph position="30"> - Noun phrases with a temporal phase or period word as head and a modifier or ordinal that signals a deictic function, such as (n?ista vecka (next week), f~rra dret (last year), etc.</Paragraph>
    <Paragraph position="31"> In many contexts prepositions are omitted in Swedish deictic expressions if the remaining NP signals a deictic function on its own.</Paragraph>
    <Paragraph position="32">  Note also that a phasal expression such as in October may, apart from specifying a phase October, also function deictically, that is, it may provide phasal values for the periods above its explicit level. However, this is regarded as a secondary use of a phasal expression.</Paragraph>
    <Paragraph position="33"> Determining the initial reference time (or index) for deictic expressions is another important consideration. Some deictic expressions are necessarily anchored to ST and others may be anchored to ST but they may also be anchored to another reference time depending on context. In Swedish expressions such as igdr (yesterday), i morgan (tomorrow), am 2 dagar (in two days) and fSrra dret (last year) are examples of expressions that must anchor to ST. Some expressions involving ndsta (next) are examples of the second group. The expressions ndista vecka (next week) and nd~sta torsdag (next Thursday) can have either ST or another context-dependent time as their initial reference time. Expressions such as ndista dag (next day.) and tvd veckor senate (two weeks later) must necessarily anchor to a reference time other than ST (see /Merkel 1987/.</Paragraph>
    <Paragraph position="34"> Next-expressions Expressions such as ndsta torsdag (next Thursday), ndista sommar (next summer) and n~ista jul (next Christmas) may be ambiguous to certain speakers. These expressions seem to be ambiguous in the sense that they either take you 1 or 2 steps from your reference time. So ntista torsdag may refer to the first or second Thursday from the initial reference time. If we changed the argument of ndsta in the above examples to words denoting periods instead of phases the corresponding reference times would not be ambiguous. Ntt'sta vecka (next week) and n?ista dag (next day) will be interpreted as having only one temporal location and not two as the first examples. How is this possible? If somebody says ntista torsdag (next Thursday) when the initial reference time is a Thursday, then there is no ambiguity; the step value will be +1 which will pick out the first phase that has the value torsdag after the initial RT. However, if ntista torsdag was uttered on a Sunday, we would have to adjust the initial RT by either starting at the Thursday preceding the initial RT orat the Thursday succeeding it, and from the 'new' initial RT move +1 to reach a temporal reference. The expression nasta jul (next Christmas) is analogous to nasta torsdag.</Paragraph>
    <Paragraph position="35"> Instead of saying that next-expressions have two possible step values, +1 or +2, we claim that there is only one value, namely +1, but that when the argument is a phasal expression, such as torsdag, the initial RT may be chosen in two ways depending on individual strategies. Some may for example choose the nearest phase as their new RT.</Paragraph>
    <Paragraph position="36"> There is a class of expressions that behaves tl~e same way as next-expressions, for example, expressions involving modifiers such as fSrra (~ last), nastnasta (the next but one).</Paragraph>
    <Paragraph position="37"> Interpretation principle By looking at the properties of the different parts of a complex temporal expression interpretation principles for frame-adverbial expressions can now be set up. Making use of the distinction between the deietic and phasal expressions together with the temporal hierarchy we can outline a Frame-adverbial interpretation principle which will explain when a frame-adverbial expression is interpretable: It is possible to interpret a temporal frmne-adverbial phrase au- null tanomcusly iff 1. a) \] he phrase consist8 of olin or several phasal expressions and b) 't'he eombimttion of phasal values wilt constitut~ a chain of 1,3mpot-al infonnafi(m according to the icmpmal hierarchy. and c) ')'he top-most phase of the chain specifies a unique phase. or 2. a) 'Yhe phrase con~;ists of a combhlation of one or several deieties  and an arbitrtay numbes of phasal expressions and b) hMex is lmown for each deictic ~mt~ e) 'the deietics and the phasal expressions each specify consistent phase-of relations. The phasal values provided will constitnte a ch~,in of temlm~al infommtion according to rim tcmlmmI hierarchy.</Paragraph>
    <Paragraph position="38"> IRt( ( d) A ddctic provides tonporal information for the highest level of the chain.</Paragraph>
    <Paragraph position="39"> The abow: principle will permit intc~pretation of expressions such as O) (2) 1987 i maj den 24 maj 1987 i mo/gon kl kvart Over' fyra ngtsh~ reeks pd torsdag kl .~vart 6vet fyra for 20 dr sedan idag rid den htir tiden i morgon 1987 in May May 24th, 1987 tomorrow at a quartet' past four ~ext week on l&amp;quot;hursday at a quarter past four 20 years ago today at this time tomorrow fSr 39 dr&amp;quot; och fyra t*~m~der sedaza 39 years and four ~umths ago om t,M veckor oeh fyra dagar in two weeks and four days The following expressions will be ruleA out by the principle:  (3) kl kvart i~ver 5 1987 at a q~tarter past Jive 1987 pd kvdllen i jnnua~i in the evening in Januacy (4) fOrra dret pd eftermiddagen last year in the afternoon  nasta recks k112 next week at nvelve o' clock The fact that an expression i.,; ruled out by the principle does not mean that into'pretation is blocked completely. The expressions in (3) and (4) are only incomplete in the sense that some perk,ds lack a phasal value and it may turn out that contextual knowledge will provide this. Inteq~retation is stack if the expression yields inconsistent phasal values. For example, the expression pd mdndagen den 23 november i dr (on Momqay Novembe~ 23rd this year) provides inconsistent phasal values for the I)AY period, ie. November 23rd i,~ a Wednesday in 1988.</Paragraph>
    <Paragraph position="40"> However, there are some borderline cases that the above principle would permit such a,q (5) for .39 dr sedan ig&amp; n~'~a reeks n&amp;~ta dr for 2 ar sedan fOr 2 nu~uuler sedan (6) for 39 &amp; sedan i ~u~rgon om s,.x dr fOr 2 mdnader se&amp;m 39 ),ears ago yesterday next week next year two years ago two months ago 39 year's ago tomorrow in six years 2 months ago \[n (5) two deictics with the same direction are combined, either + +, or .... . And in (6) there is a combination of deictics with different relational values, that is an anterior is joined with a posterior deictic. In principle it seems possible to combine de-ictics in this way, although some Swedish speakers do con-. sider the examples in (5) and (6) as being odd.</Paragraph>
    <Paragraph position="41"> The examples in (5) and (6) are all complex frame-adverllial expression holding more than one deictic phrase, but the im-portant thing is that the different deictics together specify the same tinre. In other words, if we interpret a phrase such as (7)fiOr tvd dr sedan igdr (two years ago yesterday) where the two involved deictics both have ST as obligatory index the first deietic will pick out a description of a reference time two years before ST and the second will modify that description by adding phases for DAY and MONTH based on the partial description of a phase one day before ST. This view makes it possible to claim that certain deictics such as igdr (yesterday), i morgon (tomorrow), ete, always have ST as their index. /Smith 1978/ has proposed that interpretation of expressions such as (7) is done in two steps, that is, that the first deictic forms the RT for tile second. Smith's strategy would mean that igdr (yesterday) in (7) cannot have ST as its obligatory index; with our strategy this is indeed possible. hnplementations The principles outlined for the interpretation of temporal I~'amcadverbials have been implemented in a system, CLOCKWISE, which interprets frame-adverbial expressions into a temporal description in terms of phasal wdues. Tht: first version of CI,OCKWISE consists of a parser, based on finite ~';tatc machinery, and a 'temporal expert' that will make use of its knowledge about temporal phases and periods and infer temporal information that is missing explicitly in the expressicms. A temporal representation is constructed in a notepad dm'ing the parsing process. The notepad contains information about the periods and phases denoted by the expression and~ if the expression is deictic, also index, step value and direction. Thc notepad is structured according to the temporal hierarchy and will therefore support the interpreter according to the l,lterpretation principle. The first version can, however, only deal with one type of index, namely speech time. CI_,OCKWISI-!, 1 has been usexl as a module in a natural language and graphi-cal interface to a booking system/J6nsson 1987/.</Paragraph>
    <Paragraph position="42"> The second version of CLOCKWISE is under construction. The temporal representation is built by an LFG-type gimmnar to gether with a frame-based knowledge representation whmv temporal periods and phases am treated as semantic objects (of /Ahrenberg 1988/. The grammm&amp;quot; is written in a way to syntactically filter out file temporal expressions, ie. temporal prepositional phrases are functionally distinct front other prepositional phrases in the functional structures (fstructures). The result of a successful interpretation of a temporal frame-adverbial is a directed acyclic graph (DAG) where a set of phasal values forms a description of the time denoted by the expression. For example, a frame-adverbial such as niista vecka pd fredag kl 12.30 (next week on Friday at 12.30) has a functional structure of the following kind:  The above DAG is mapped to an intermediate DAG by the Frame system. The intermediate DAG has separated deictic phrases from phasal ones, and also provided a frame for the deictic mapping function F.</Paragraph>
    <Paragraph position="43"> rEaP icT,c \] / r \]/</Paragraph>
  </Section>
  <Section position="2" start_page="429" end_page="429" type="metho">
    <SectionTitle>
EVHASAL *&amp;quot;OUR12 2
EPHASAL SMINUTE30~
</SectionTitle>
    <Paragraph position="0"> Fig. 4 Intermediate structure By using a more general and powerful framework we are able to incorporate other types of temporal expressions and develop the above principles in relation to tense and aspect. Durative expressions have recently been incorporated in the system.</Paragraph>
    <Paragraph position="1"> The development of CLOCKWISE is part of the grammar development project in the LINLIN project at LinkSping University (/Ahrenberg 1987/).</Paragraph>
    <Paragraph position="2"> CLOCKWISE is at the moment unable to handle expressions such as (1) The first week of January CLOCKWISE will get stuck on this expression due to the fact that weeks cannot be mapped onto months, whereas the first day of January would result in a description. One solution would be to force a mapping of weeks onto months (in the same way as weeks are mapped onto years in the Swedish calendar) giving us five subperiods of MONTH each of week length, with corresponding phases where the first and the fifth week phase sometimes would be partial.</Paragraph>
  </Section>
  <Section position="3" start_page="429" end_page="429" type="metho">
    <SectionTitle>
Acknowledgements
</SectionTitle>
    <Paragraph position="0"> This research has been supported by the NationaV Swedish Board for Technical Development.</Paragraph>
    <Paragraph position="1"> I would like to thank Lars Ahrenberg, Ntis Dahlbltck, Arne Jtnsson, Mats Wirtn och Ivan Rankin for helpful comments on drafts of this paper. I am also indebted to Jim Goodwin who has provided me with the knowledge representation tool used in</Paragraph>
  </Section>
class="xml-element"></Paper>