From Language to Time:A Temporal Expression AnchorerBenjamin Han, Donna Gates and Lori LevinLanguage Technologies InstituteCarnegie Mellon UniversityThis work is supported by the Defense Advanced Research Projects Agency (DARPA) under the project RADAR.Benjamin Han, Donna Gates and Lori Levin, TIME 2006, Budapest, Hungary Introduction•Many Natural Language applications need to understand the semantics of temporal expressions.•We developed a constraint-based meaning representation TCNL (Time Calculus for Natural Language).•We developed system TEA (Temporal Expression Anchorer) for normalizing temporal expressions.•Experiments on emails showed promising results.2Benjamin Han, Donna Gates and Lori Levin, TIME 2006, Budapest, Hungary Application I: EmailsDate: Thu, 11 Sep 1997 00:14:36 -0500I have put an outline out in the n10f1 OpReview directory...(omitted)We have very little time for this. Please call me Thursdaynight to get clarification. I will need graphs and prose infiles by Saturday Noon.– Maryps. Mark and John , I waited until AFTER midnight tosend this .Figure 1: A sample email (edited)used3. In contrast, explicit expressions on averageonly account for around 9.5% in the three emaildatasets. This is not surprising given that peopletend to use under-specified expressions in emails foreconomic reasons. Another thing to note is that thereare roughly the same number of relative expressionsand non-relative expressions. Since non-relative ex-pressions (including deictic expressions) can be an-chored without tracking the temporal focus over adiscourse and therefore can be dealt with in a fairlystraightforward way, we may assign 50% as a some-what generous baseline performance of any anchor-ing system4.Another difference between emails and newswiretexts is that the former is a medium for communi-cation: an email can be used as a reply, or can beattached within another email, or even be used toaddress to multiple recipients. All of this compli-cates a great deal of our task. Other notable dif-ferences are that in emails hour ambiguity tend toappear more often (“I’ll be home at 2.”), and peo-ple tend to be more creative when they composeshort messages such as using tables (e.g., an entirecolumn of numbers to denote the number of min-utes alloted for each presenter), bullet lists, abbrevi-ations, and different month/day formats (“1/9” canmean January 9 or September 1), etc. Emails alsocontain more “human errors” such as misspellings(“Thusday” to mean Thursday) and confusion aboutdates (e.g., using “tomorrow” when sending emails3Using the North American News Corpus.4This is a bit generous since solving simple calendric arith-metics such as anchoring last summer still requires a non-trivialmodeling of human calendars; see Sec. 3.around midnight), etc. Overall it is very difficult torecover from this type of errors.3 Representing Times in NaturalLanguageThis section provides a concise overview of TCNL;readers are referred to (Han and Kohlhase, 2003;Han et al., 2006) for more detail.TCNL has two major components: a constraint-based model for human calendars and a represen-tational language built on top of the model. Dif-ferent from the other representations such as Zeit-Gram (Stede and Haas, 1998), TOP (Androut-sopoulos, 1999), and TimeML/Timex3 (Saur´ı et al.,2006), the language component of TCNL is essen-tially “calendar-agnostic” - any temporal unit can beplugged in a formula once it is defined in the cal-endar model, i.e., the calendar model serves as thelexicon for the TCNL language.Fig. 2 shows a partial model for the Gregorian cal-endar used in TEA. The entire calendar model is ba-sically a constraint graph with partial ordering. Thenodes labeled with “year” etc. represent temporalunits (or variables when viewed as a constraint sat-isfaction problem (CSP) (Ruttkay, 1998)), and eachunit can take on a set of possible values. The undi-rected edges represent constraints among the units,e.g., the constraint between month and day man-dates that February cannot have more than 29 days.A temporal expression in NL is then viewed as ifit assigns values to some of the units, e.g., “Fridaythe 13th” assigns values to only units dow (day-of-week) and day. An interval-based AC-3 algo-rithm with a chronological backtracking mechanismis used to derive at the consistent assignments to theother units, therefore allowing us to iterate to anyone of the possible Friday the 13th.The ordering among the units is designated by tworelations: measurement and periodicity (arrows inFig. 2). These relations are essential for supportingvarious operations provided by the TCNL languagesuch as determining temporal ordering of two timepoints, performing arithmetic, and changing tempo-ral granularity, etc. For example, to interpret the ex-pression “early July”, we identify that July is a valueof unit month, and month is measured by day. Wethen obtain the size of July in terms of day (31) and+f{thu, night} = (19970911T18????..19970911T23????)[f +f{sat, noon}] = min/19970913T12????[f {>= -p{midnight}}] = min/(19970911..max){thu, 11_day, sep, 1997_year, 0_hour, 14_min, 36_sec} =19970911T0014363Han et al., HLT-NAACL 2006Benjamin Han, Donna Gates and Lori Levin, TIME 2006, Budapest, Hungary Mozart!s Activity in Vienna#Mozart went to Munich to compose the opera late in 1780. The next year, he was summoned from Munich to Vienna, where the Salzburg court was in residence on the accession of a new emperor. Mozart lived in Vienna for the rest of his life, until he died in 1791."T7={1791year}T6[1 year][0,!] [0,!]T5={1780year}T1[14 years]T2[3 years] [0,2 weeks]T4= {17day, jul,1787year}temporal variableT1'monthyeardayqoymonth ⊗ daysoy ⊗ month ⊗ dayhourNsoytod{morning,...}...calendar (CSP)solution: {1770year}solution: {1784year}T3solution: {1787year,apr}[3 months]solution: {(>=93200,<=93206)week}solution: {1781year}Figure 3: Did Beethoven and Mozart meet in Vienna? Mozart’s activity is shown in shaded graph.“Mozart went to Munich to compose the opera late in 1780 (T5 : {1780year}). Thenext year (T6 : { +|1year|}), he was summoned from Munich to Vienna, where theSalzburg court was in residence on the accession of a new emperor. Mozart lived inVienna for the rest of his life, until he died in 1791 (T7 : {1791year}).”Again the Interpretation Module rewrites T6 into {T5 + |1year|}, and a TCSP is formed. At thispoint we are interested to see if it is possible that