<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-0912">
  <Title>tions of Pragmatic Principles: Implications for Theories of Linguistic and Cognitive Representation. Jan Nuyts and Eric Pederson. Language and</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
1 Linguistic underspecification
</SectionTitle>
    <Paragraph position="0"> Natural language expressions deliver merely templates for the construction of a contextually relevant interpretation, namely the propositional meaning the hearer ascribes to the respective perceived utterance. What is linguistically given delivers only part of the input necessary for interpretation and has to be enriched and specified by recourse to sentence-level and discourse-level context, as well as to world-knowledge. Propositions, i.e. truth-conditional content, arise only after that enrichment process has taken place.</Paragraph>
    <Paragraph position="1"> NLP applications need to capture the meaning of a linguistic input to adequately work with it in the respective applications. But the 'meaning' relevant in this case, viz. the intended interpretation of an utterance, is underspecified by the lexically given input in at least two ways. Linguistic meaning underspecifies propositions, i.e.</Paragraph>
    <Paragraph position="2"> 'what is said' (Carston, 2002), and 'what is said' in turn underspecifies 'what is meant' (Grice, 1989). Both kinds of underspecification have to be solved by the hearer in natural discourse settings, and by the corresonding algorithms in NLP-applications. As such applications necessarily rely on linguistically given input, they have to be supplemented by a systematic account of the inferential mechanisms needed to enrich and specify the lexical schemes. A further difficulty for NLP applications is that, presuming some form of cooperativity, any utterance can be interpreted by appropriately accommodating the context. null Apparently, computing the contextually relevant interpretation of an utterance requires more than capturing and computing lexical input. It relies crucially on further information and principles of how to derive that information and combine it with what is lexically given. Information - contextual or conceptual - that is not contained in the natural language string currently being processed is subject to pragmatic actions (Perrett, 2000: 102).</Paragraph>
    <Paragraph position="3"> This paper is organized as follows: Section 2 illustrates the assumptions on pragmatic reasoning relevant here. Section 3 deals with Russian ipf aspect as one example of linguistic under-specification and sketches the semantics and pragmatics involved in the derivation of its readings. Implications for NLP will be given in section 4, and section 5 offers a short conclusion.</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2 Pragmatic Reasoning
</SectionTitle>
    <Paragraph position="0"> The twofold underspecification of natural language expressions is resolved by pragmatic reasoning: 'implicatures', derived by inference alone, enrich what is said to what is meant and 'explicatures', derived by decoding plus inference, enrich the lexical input to what is said.</Paragraph>
    <Paragraph position="1"> Both interact in mutual parallel adjustment, in that the expectation of certain implicatures constrains the possible range of explicatures. Thus, we are dealing with two kinds of semantics lexical and propositonal - and with two kinds of pragmatics - explicature and implicature (Carston, 2002):  The pragmatic actions can be captured by Levinson's (2000) heuristics, whose application gives rise to default interpretations by evoking certain inferences. The heuristics and their corresponding inferences are based on Grice's Maxims of Conversation: Q-heuristics are based on the first quantity maxim ('make your statement as informative as possible') and license inference to the negation or invalidity of a corresponding stronger expression, M-heuristics stem from violations of the manner maxim (esp. 'avoid obscurity of expression' and 'avoid prolixity'), and license the inference from marked expressions to marked interpretations. I-heuristics are based on the second quantity maxim ('do not say more than necessary') and allow for inference to a stereotype. Contrary to the Gricean view, however, these are assumed to work partly also on the subpropositional level giving rise to 'explicatures' (Carston, 2002), thereby enriching and constraining the underspecified lexical representation. null An advantage of using heuristics giving rise to default interpretations is that they capture the fact that real minds have to make decisions under constraints of limited time and knowledge. In hand-ling tasks like interpretation, humans have to use approximate methods like heuristics that guide the search and determine when it should end, and simple decision rules that make use of the information found. To behave adaptively in the environment in general, and in interpreting utterances in special, humans must be able to make inferences that are fast, frugal and accurate (Gigerenzer et al., 1999). These inferences, or the heuristics they are based on, respectively, work well insofar as they make a tradeoff on the dimension of generality vs. specificity. Their simplicity allows them to be robust when confronted with a change in environment and to generalize to new situations. Such heuristics are rather suitable means for dealing with underspecified, sense-general linguistic expressions that equally make a tradeoff between underdetermination and preciseness. Furthermore, heuristics are of the kind NLP applications can deal with (section 4).</Paragraph>
    <Paragraph position="2"> Levinson's heuristics, at least the M- and Qheuristics, are instances of such fast and frugal heuristics that give rise to default inferences which have to be made more specific in the course of interpretation. This is achieved by the rather specific I-heuristics, that give rise to inferences by referring to assumptions provided by a concept being activated by a lexical item in a certain context and discourse setting.</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="3" type="metho">
    <SectionTitle>
3 Ipf aspect in Russian
</SectionTitle>
    <Paragraph position="0"> One instance of underspecification is aspect in Russian, especially the ipf  , which gives rise to a considerable variety of readings (1a-g). Their context-dependence and defeasibility indicates their partial pragmatic character.</Paragraph>
    <Paragraph position="1"> 1a. actual-processual reading Sar medlenno podnimalsja.</Paragraph>
    <Paragraph position="2"> balloon slowly ascend:PAST:ipf  'He was a good locksmith: he could open every door.' 1g. habitual reading Po subbotam, ona chodila v banju. on saturday, she go:PAST:ipf to sauna 'She used to go to the sauna on Saturdays.' This variety makes it difficult to capture the semantics of ipf with one invariant meaning. Postulating polysemy does not solve the problem either, as this would presuppose the existence of a clearly defined number of discrete readings. This is indeed not the case, in view of the ongoing debat about how many and which readings to postulate. Instead, ipf is best considered as an instance of 'sense-generality' (Atlas 1989), i.e. as having a definite, but highly general meaning that is specified in the course of interpretation by means of context and pragmatic inference. Sense-generality and pragmatic inference constitute the semantic and pragmatic side of utterance interpretation: &amp;quot;[...] Grice's pragmatics without sense-generality is blind, and sense-generality without Grice's pragmatics is empty&amp;quot; (Atlas, 1989: 149).</Paragraph>
    <Paragraph position="3"> Accordingly, this paper aims at explaining the derivation of ipf readings by a more pragmatically oriented account, based on a 'selectional theory' of aspect (Bickel, 1996).</Paragraph>
    <Section position="1" start_page="1" end_page="1" type="sub_section">
      <SectionTitle>
3.1 Semantics
</SectionTitle>
      <Paragraph position="0"> As regards semantics, aspect is assumed to contribute to the finite part of an utterance in selecting a certain part of the descriptive content of the verb and relating it to the time an assertion is made about, the topic time TT (Klein, 1995). The relation thus obtained constitutes the aspectual value. The decisive units for this selection are phases and boundaries (Bickel, 1996). Presuming a tripartite event structure (Moens and Steedman, 1988) consisting of a preparation phase (dynamic phase ph dyn ), a culmination point (boundary t) and a consequent state (static phase ph stat ), there are three possibilities for aspect to select. English and Turkish both have ph dyn -selecting aspectual markers, Turkish also a marker for explicit ph stat  selection; Russian pf aspect explicitly selects t. The unmarked members of the aspectual oppositions may assert anything else - Russian ipf aspect may assert anything but the explicit selection of a boundary.</Paragraph>
      <Paragraph position="1"> As truth-evaluation and assertion have to be kept apart, the selected and asserted parts have to be related an interval of time, the validation interval, where they are asserted to hold and truthconditionally evaluated. Paduceva (1996) refers to this interval with the notion of tocka otsceta (henceforth TO), an aspectual reference interval, which can be retrospective or synchronous (bounded or unbouded) with respect to the asserted part. TO is decisive for constraining the range of possible interpretations of ipf as it classifies the readings into three groups. In most cases TO is lexically given, predominantly by adverbials, and therefore constitutes an essential contextual feature for the derivation of the relevant reading. Table 1 shows the classification of ipf readings according to TO  What has been said so far provides the semantic skeleton that has to be accompanied and fleshed out by pragmatic reasoning. Note, that pragmatics does not apply 'after' semantics, but rather interleaves with it.</Paragraph>
    </Section>
    <Section position="2" start_page="1" end_page="3" type="sub_section">
      <SectionTitle>
3.2 Pragmatics
</SectionTitle>
      <Paragraph position="0"> As already mentioned before, the principles assumed to guide the pragmatic inferencing are the heuristics proposed by Levinson (2000), which are assumed to apply also at the subpropositional level giving rise to 'explicatures' (Carston, 2002) that enrich and constrain underspecified lexical representations to yield full-fledged propositions.</Paragraph>
      <Paragraph position="1"> Q-inferences are involved in deriving the meaning of unmarked forms by giving rise to scalar implicatures from the scale &lt;pf, ipf&gt;, meaning that if the speaker uses the weaker element of the scale (ipf) the hearer is entiteled to infer the invalidity of the stronger expression (pf): Using the pf aspect explicitly marks the selection of a boundary, using the ipf does not exclude that selection and thus gives rise to the  That it is indeed reasonable to distinguish these three groups is supported by a look at Turkish which has morphological means to express the respective relation (cf. Sonnenhauser, 2003).</Paragraph>
      <Paragraph position="2"> three possible values stated above in table 1. This pragmatically derived meaning of ipf is captured by the metavariable 'IPF', that comprises the possibilities for ipf listed in table 1. By contextual, discourse pragmatic or general pragmatic principles (which fix TO), this metavariable gets disambiguated as to what kind of relation it expresses. As one TO allows for several readings (cf. 2a-c), the relation obtained by the disambiguation is sense-general and needs further enrichment to yield the contextually relevant reading. This process is an instance of the drawing of I-inferences referring to a certain concept. Encyclopaedic knowledge makes certain assumptions more likely than others, and if these assumptions lead to a relevant interpretation, all other lines of interpretation are dismissed. Often encountered concepts are more likely to get accessed, as they require less effort in getting activated. null  As the potential reading (2a) is not relevant with (2b) - translators are supposed to be able to speak a certain language -, other interpretations are looked for, here the actual-processual.</Paragraph>
      <Paragraph position="3"> 2c. Perevodcik govoril i po-turecki.</Paragraph>
      <Paragraph position="4"> translator speak:PAST:ipf also Turkish 'The translator could also speak Turkish.' The insertion of i ('also') in (2c), however, creates the appropriate context for the potential reading to be relevant.</Paragraph>
      <Paragraph position="5"> The reasoning stops as soon as a relevant interpretation is achieved.</Paragraph>
      <Paragraph position="6"> Note that lexical content of the verb and aspect strictly have to be kept apart; ipf requires a certain input and if a verb is only providing a t in its semantic representation, the required phase has to be induced - again, by means of context or pragmatic reasoning. This may result in an iterative reading, or in 'zooming in' in the (inferred) preparation phase ph dyn , (cf. 3a-b; cf. also 9a-b below): 3a. Ivan vyigral gonku.</Paragraph>
      <Paragraph position="7"> Ivan win:PAST:pf race:ACC 'Ivan won the race.' The pf is applied to a verb vyigrat'('win') that provides a boundary in its semantic representation and therefore no interpretational rearrangements are necessary.</Paragraph>
      <Paragraph position="8"> b. Ivan vyigryval gonku (cetyre raza).</Paragraph>
      <Paragraph position="9"> Ivan win: PAST:ipf race:ACC four times.</Paragraph>
      <Paragraph position="10"> 'Ivan won the race four times / was winning the race.' The application of the ipf in (3b) requires a ph dyn ; the verb does not provide one, thus it has to be induced by iteration, or by inferring and 'zooming in' on it, e.g. in a context fixing a certain moment withinph dyn , e.g. a when-clause.</Paragraph>
      <Paragraph position="11"> This coercion process is to be located at the transition from 1 to 2 (figure 2) and may be captured in terms of M-inferences: marked expressions warn 'marked situation'. Again, we are dealing with an interaction of implicature and explicature: 'Marked situation' is the implicature, that gives rise to the explicatures that explain the markedness  of the situation in deriving a relevant reading, i.e. steps 2 and 3 are carried out as in the non-coercion case. Note further that this reasoning process is not carried out step by step, but in parallel adjustment, which may be modelled within a game-theoretic framwork or  lly unmarked. The notion of orm notions of morphological and semantic markedness.</Paragraph>
      <Paragraph position="12"> Note, that the ipf aspect remains semantica pragmatic markedness has to be distinguished f  Q- and M-inferences differ crucially from I-inferences in that the latter arise from the interplay of semantic representation with context, while Q- and M-inferences do not need any context for derivation, but are based on sets of alternates with essentially the same form and contrastive semantic content in the case of Qinferences, and with contrasting forms but essentially the same inherent semantic content in the case of M-inferences. That is, Q- and M-inferences can be drawn on the basis of what is lexically given and on the sets of possible alternates. They give rise to utterance-type meaning as they are based on general expectations on how language is normally used (Levinson, 2000).</Paragraph>
      <Paragraph position="13"> This property is useful for NLP applications as will be shown in section 4 below. It is due to this difference, that Q-/M- and I-implicatures come into play at different levels of meaning composition, inference and interpretation.</Paragraph>
      <Paragraph position="14"> Figure 2 sketches the pragmatic processes involved in the derivation of the specific readings. What becomes obvious from figure 2 is the fact that there is one level at which ambiguity is to be postulated and one level where we are dealing with sense-generality: The three relations the IPF meta-variable may give rise to are instances of ambiguity; the disambiguated relations, however, are still sense-general and have to be further specified in the course of interpretation. Once detected, the ambiguity is easy to handle, as in most cases it gets disambiguated by lexical means specifying TO. The specification of the sense-generality involved is more difficult to deal with.</Paragraph>
    </Section>
    <Section position="3" start_page="3" end_page="3" type="sub_section">
      <SectionTitle>
3.3 Interaction
</SectionTitle>
      <Paragraph position="0"> What has been outlined so far suggests that a more pragmatically oriented analyses of closed classes of morphems, aspect in the present case, yields interesting insights and resolves puzzles that arise from the apparently elusive and variable content of these expressions (Levinson, 2000: 98).</Paragraph>
      <Paragraph position="1"> Much of the confusion within aspectology, and presumably within further grammatical categories, arises due to a neglect of pragmatics and to definitions of aspect that take pragmatic inferences as part of its semantics (Bickel, 1996). The division of labour advocated in this paper keeps the semantics of aspect rather simple, but has to be supported by general pragmatic principles.</Paragraph>
      <Paragraph position="2"> Such an analysis is cross-linguistically applicable and facilitates systematic comparison of different aspectual systems. Furthermore, such a simple semantics provides a suitable input for NLP applications. null The division of labour between semantics and pragmatics, i.e. the relation between the input provided by the lexicon and the contribution of pragmatic reasoning, varies from language to language, but the semantics stated in the selectional theory, and the inferential heuristics themselves remain stable across languages.</Paragraph>
    </Section>
  </Section>
  <Section position="6" start_page="3" end_page="8" type="metho">
    <SectionTitle>
4 Implications for NLP
</SectionTitle>
    <Paragraph position="0"> Without a principled account of pragmatic, i.e.</Paragraph>
    <Paragraph position="1"> inferential, principles, applications in natural language processing will necessarily fail. This means, that the lexically given facts currently being processed have to be combined with information from former parse-states - as the interpretation is modelled as processing incrementally from left to right (Kempson et al., 2001) -, with general heuristics of pragmatic reasoning and with access to conceptual knowledge.</Paragraph>
    <Paragraph position="2"> To make the findings so fare suitable for NLP, the decisive input factors have to be stated as well as the principles of their interaction. Contrary to human interpretation of underspecified forms, where meaning is constructed predominantly by abductive reasoning, in computer driven interpretation meanings have to be selected from an a priori established list of possibilities (cf. Sonnenhauser, 2004). For ipf this means that a list of readings has to be compiled (see table 1), the factors involved their derivation have to be fixed and rules of interaction have to be stated that can be expressed in the deductive, propositional logic form A - B (cf. Vazov and Lapalme, 2000), where A is the premise, i.e. it states the conditions for the rules to apply, and B the conclusion, i.e. the derived reading.</Paragraph>
    <Section position="1" start_page="3" end_page="5" type="sub_section">
      <SectionTitle>
4.1 The input
</SectionTitle>
      <Paragraph position="0"> Input factors for algorithms are the following: Verbs indexed for the ph and t they contain  ; lexical items indexed for whether they add ph or t, and aspectual selectors indexed for what they select and for their status within the language specific semantic markedness relation. That is how Q-inferences are drawn. TO constrains the possible interpretations of the unmarked aspec- null Languages differ as to whether this lexical specification already gives rise to 'default aspect' (Bohnemeyer and Swift, 2001) as in Russian, or not, as in Turkish.</Paragraph>
      <Paragraph position="1"> tual partner in disambiguating IPF. As it is in most cases lexically given, adverbials are to be annotated as to what kind of TO they fix.</Paragraph>
      <Paragraph position="2"> To capture M-inferences, the default combinations of base and selector have to be stated, as well as rules for resolving possibly occurring mismatches. The M-inferences then can be pinned down by coercion operators (Pulman, 1997; Thomas and Pulman, 1999). As pointed out above, the metalinguistic Q- and M-inferences can be handeled purley on the basis of the lexically given input and on the basis of the possible alternatives that have to be stated as well.</Paragraph>
      <Paragraph position="3"> The presumably most difficult problem is how to specify verbs for the conceptual knowledge they provide access to, which is indespensable for I-inference to be drawn. One means is corpus analysis in order to detect regularities and coocurrences of lexical items that might hint to a conceptual connection  . As the factor 'probability' cannot be completely eliminated, a condition has to be implemented preferring the shortest line of reasoning (Thomas and Pulman, 1999).</Paragraph>
      <Paragraph position="4"> Furthermore, a mechanism must be implemented that parallels the stop mechanism in human interpretation, i.e. a mechanism that stops the search for information to be built in in the interpretation process. The reasoning stops as soon a truth-conditional interpretation is achieved, i.e. as soon as the requirement for that propositional content is achieved. It is this requirement that crucially drives the interpretation process (cf. the requirement '?Ty(t)' in the Dynamic Syntax approach as advocated by Kempson et al., 2001).</Paragraph>
      <Paragraph position="5"> Interpretation of aspectual forms processes incrementally, i.e. information once provided and processed cannot be undone.</Paragraph>
    </Section>
    <Section position="2" start_page="5" end_page="6" type="sub_section">
      <SectionTitle>
4.2 The 'default' case
</SectionTitle>
      <Paragraph position="0"> The default case is a match of basis and marker  , where the verbal basis provides the necessary input for the marker to apply. For ipf, the conditions have to be stated under which the three possibilites (figure 1) get activated. Here, TO primarily specified by temporal or manner adverbials (e.g. vse bol'se 'more and more', choroso 'well') - is decisive. Adverbials of car- null Assuming an associative organization of memory, where related things are probabilistically associated reflecting generalizations about regularities in an organism's perception of and interaction with its environment (Robinson, 1997), this statistical approach indeed seems to be on the right track.  'Marker' is the general term that subsumes also cases of default aspect, where there is no overt morphological marking, as frequently encountered in Russian.</Paragraph>
      <Paragraph position="1"> dinality and duration fix TO as retrospective and the reading as being out of group III. The rule for this line of interpretation can be stated as follows (adopted from Vazov and Lapalme, 2000): 4. IF ipf is applied to a verb providing a phase AND if there is an adverbial fixing TO as retrospective THEN the reading is out of group III.</Paragraph>
      <Paragraph position="2">  Ja guljala ot trech to pjati.</Paragraph>
      <Paragraph position="3"> I go-for-a-walk:PAST:ipf from three:Gen to five:Gen 'From three to five, I went for a walk.' Both interpretations can be overridden if TO is turned into a synchronous one by adverbials of the type vsegda ('always') or obycno ('usually'): 6a. habitual reading Ja obycno guljala ot trech to pjati.</Paragraph>
      <Paragraph position="4"> I usually go-for-walk:PAST:ipf from three to five 'I usually went for a walk from three to five.' This shows the incremental way of interpretation, whereby the inner parts are left intact:  The outermost TO is the one relevant for discourse advancement or non-advancement, respectively. null A synchronous TO may be bounded or unbounded (group I and II, table 1), cf. (7) and (8): 7. IF ipf is applied to a verb providing a phase AND if there is an adverbial fixing TO as synchronous bounded/unbounded THEN the reading is out of group I/II 8a. actual-processual reading V vosem' casov, ja citala knigu.</Paragraph>
      <Paragraph position="5"> at eight o'clock, I read:Past:ipf book:Acc 'At eight o'clock, I was reading a book.' 8b. inactual reading Ran'se, on rabotal v universitete.</Paragraph>
      <Paragraph position="6"> before he work:PAST:ipf at university 'He used to work at university.' (= 'He was working as a teacher.') Depending on the semantic representation of the verb, implicatures or presuppositions may arise. Ipf with the structure [ph t] leaves the reaching of the boundary as an implicature, ipf with [t ph] leaves the initial boundary as presupposition. That is how the semantic representations of verbs provide background and frame for pragmatic reasoning.</Paragraph>
    </Section>
    <Section position="3" start_page="6" end_page="8" type="sub_section">
      <SectionTitle>
4.3 The 'coercion' case
</SectionTitle>
      <Paragraph position="0"> Whenever an aspectual marker is applied on a basis not providing the relevant feature (ph or t) for it to apply, that feature is semantically or pragmatically induced in order to eliminate that mismatch. Coercion operators capture this recategorization process (Pulman, 1997): 9a. Ivan vyigral gonku.</Paragraph>
      <Paragraph position="1"> Ivan win:PAST:pf race:ACC 'Ivan won the race.' Here, pf is applied to a verb that provides a t; no coercion is necessary.</Paragraph>
      <Paragraph position="2"> 9b. Ivan vyigryval gonku (cetyre raza).</Paragraph>
      <Paragraph position="3"> Ivan win: PAST:ipf race:ACC (four times).</Paragraph>
      <Paragraph position="4"> 'Ivan won the race four times / was winning the race.' The application of ipf in (9b) requires a ph, which the verb vyigrat' ('win') does not provide. So it has to be induced by iteration or by zooming in on ph dyn . Two coercion operators may be applied: &amp;quot;iterate / stretch: point - process&amp;quot; (Pulman, 1997). In most cases, context provides the necessary cues for disambiguation (e.g. an iterative adverbials or a when-clause), if not, one has to rely on the 'probability-condition'. 9c. V vosem casov ona uze vysla.</Paragraph>
      <Paragraph position="5"> at eight o'clock she already leave:PAST:pf 'At eight o'clock, she had already left.' (= she was gone) For the consequent-state reading in (9c) to arise, the prefix vy- first has to induce the boundary required for pf to apply. Uze ('already') here triggers the application of the coercion operator &amp;quot;add-cstate: X - &lt;X, state&gt;, where X is point or process&amp;quot; (Pulman 1997). The rules for (9b) are: 10a. IF ipf is applied to a verb providing no phase, AND a lexical item indicating iteration is present THEN induce the phase by application of 'ite-rate' 10b. IF ipf is applied to a verb providing no phase AND and adverbial/clause indicating incidence is present THEN induce the phase by application of 'stretch' The application of ipf onto a verbal basis providing merely a t (prior to coercion) is both pragmatically and morphologically marked, but ipf does not lose its semantic unmarkedness. Though interpretation in terms of coercion is compositional, the specific reading this coercion gives gives rise to depends on linguistic context and world-knowledge (de Swart, 1998); cf. (11)  the exercise.' Whereas (11c) can be disambiguated by fixing TO as retrospective, (11a) and (11b) cannot be distinguished by TO alone as both require it to be synchronous. The distinction between the possible readings is left to contextual disambiguation and world-knowledge. Gaining probability values and assigning them to interpretations by a statistical approach taking into account judgements of native speakers seems to be a possible way (cf. Glovinskaja, 1982 for the habitual, potential and actual-processual reading), but the probability rankings can be overriden by the lexical content of verbal phrases.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>