Computing Implicit Entities and Events with Getaruns

Rodolfo Delmonte

Department of Language Sciences, Università Ca’ Foscari

Ca’ Bembo - Dorsoduro 1075, 30123 Venezia, Italy

Abstract. In this paper we will focus on the notion of “implicit” or lexically

unexpressed linguistic elements that are nonetheless necessary for a complete

semantic interpretation of a text. We referred to “entities” and “events” because

the recovery of the implicit material may affect all the modules of a system for

semantic processing, from the grammatically guided components to the

inferential and reasoning ones. Reference to the system GETARUNS offers one

possible implementation of the algorithms and procedures needed to cope with

the problem and allows to deal with all the spectrum of phenomena. The paper

will address at first the following three types of “implicit” entities and events:

- the grammatical ones, as suggested by a linguistic theories like LFG or

similar generative theories;

- the semantic ones suggested in the FrameNet project, i.e. CNI, DNI, INI;

- the pragmatic ones: here we will present a theory and an implementation for

the recovery of implicit entities and events of (non-) standard implicatures.

In particular we will show how the use of commonsense knowledge may

fruitfully contribute in finding relevant implied meanings. Last Implicit Entity

only touched on, though for lack of space, by the paper is the Subject of Point

of View which is computed by Semantic Informational Structure and

contributes the intended entity from whose point of view is expressed a given

subjective statement.

1 Introduction

The main difference existing between shallow syntactic methods mainly represented

by the dependence parsing framework, and the linguistically oriented Semantic Text

Processing systems can be gauged by the ability to compute Implicit Entities and

Events (hence IEEs) which is only viable in the latter but not in the former type of

systems. Shallow and dependency oriented approaches usually deal only with the

actual words lexically expressed in a text. On the contrary, deep approaches allow the

system to delve into the lexically and linguistically motivated IEEs by positing the

existence of empty categories. Empty categories are placeholders for a number of

different entities mostly recovered by means of chains of indices. They also respond

to grammatical principles of various kinds, as the universal need that any predicate

must have a SUBJect - more on these topics in the first section below.

Former systems make use of shallow lexica, which can be constituted by semantic

relations as the ones present in WordNet which only require word level semantic

relation matching. Using complex subcategorized lexica like COMLEX or NOMLEX

Delmonte R. (2009).

Computing Implicit Entities and Events with Getaruns.

In Proceedings of the 6th International Workshop on Natural Language Processing and Cognitive Science , pages 23-35

DOI: 10.5220/0002171600230035

 SciTePress

[2], FrameNet [3], and other similar computational lexica, requires a different

approach to the task of semantic processing which may account for IEEs.

Systems using these lexica must be able to distinguish Arguments from Adjuncts

and to build appropriate representations for Predicate Argument Structures: this in

turn requires clause level segmentation. These latter structures become then the input

to the semantic interpreter that can apply principles of grammatical, lexical and

semantic well-formedness to the analysis. It is just in this phase that the presence of

IEEs can be detected by the system and an adequate semantic representation can be

built. IEEs can be classified according to the following linguistically motivated

subdivision:

a. grammatically motivated IEEs

o as a subtype, IEEs identified by pronominal binding

b. semantically motivated IEEs

o as a subtype, IEEs identified by anaphoric binding

c. pragmatically motivated IEEs

o semantically inferred IEEs

d. discourse motivated IEEs

o Centering Main Topic IEEs

We will now give examples of the five types and comment on their status in a

theoretical and computational framework. The presentation will use LFG as linguistic

theory and FrameNet as semantic lexical theory; semantic representations are inspired

by Situational Semantics.

In the Penn Treebank II there are approximately 65000 empty categories, the

majority of which is constituted by traces – in a Chomskian sense (out grammatical

IEEs) – of moved material. Empty elements constituted by unexpressed Subjects of

untensed clauses are some 27800. Then there is some 580 elliptical empty elements.

If we consider that the PTB contains 93539 sentences we see that there is almost one

such empty element per sentence. Since 38133 are SBAR clauses, they contain each

one empty coindexed element. Hence, computing implicit elements is an important

component of any semantically viable text analysis component.

1.1 Grammatically Motivated IEEs

According to LFG theory as proposed by [1] grammatical IEEs may be classified into

three types:

- Lexical Control

- Syntactic Control

- Structural Control

In order to activate control mechanisms, two procedures need to be implemented

in the parser: an indexing function that distinguishes phrase structures from one

another; a coindexing function and a corresponding chain climbing function to

recover the semantic identity or the head of the controller.

Lexical Control IEEs are constituted by empty categories that ensue from the

presence of a phrase structure which has been computed as argument of the governing

predicate and needs the presence of a SUBJect. An empty category is created which is

then coindexed with the lexical control argument. Relevant examples are (1), (2) and

(3) below:

(1) Mary asked John to buy a book

(2) John considers Mary an important ally

(3) Tom is a republican.

In (1) the predicate BUY is associated to an empty SUBJect which is coindexed

with the controller JOHN in force of the existence of a lexical rule that selects a

controller between a hierarchy of Grammatical Functions associated to lexically

expressed arguments of the governing predicate, in this case ASK. The hierarchy is

the Default Rule of Lexical Control and establishes the following order:

(2.1) OBJ2 < OBJ < SUBJ

and simply says that the lexical controller is an OBJ2 if present, an OBJ if present,

otherwise a SUBJ. In (2) the predicative NP “an important ally” is controlled lexically

by the OBJect Mary in force of the same Default Rule, where the argument receives

semantic role also from the lexical controller. Same situation with (3) where however

the SUBJect is the controller of the predicative NP “a republican”.

Syntactically Controlled IEEs refer to what are also called long distance

dependencies. These constructions concern two types of clauses: relative clauses and

interrogative clauses. These are too well known by computational linguists to require

a presentation. We just include two examples to complete the description:

(4) Tom wanted the book that Mary bought.

(5) Which book did Mary buy?

In both examples the predicate BUY needs the existence of an empty category

which is then filled by or coindexed with the syntactic controller, the BOOK. This

may be achieved by different procedures according to each linguistic theory, but the

final result is always the same: a chain between two structures one of which has a

control index added by the grammar.

Structurally Controlled IEEs are those SUBJects that come into existence whenever

either a predicative ADJunct in a certain structural configuration. We include here

below the relevant examples:

(6) John went to the see the movie drunk.

(6.1) John accompanied Mary to the movie naked.

(6.2) Drunk as usual John went to see the movie.

(6.3) Naked as usual John took Mary to the movie.

(7) The company has sold its assets to collect funds.

(7.1) These assets have been sold to collect funds.

**(7.2) These assets sell well to collect funds.

(8) After reading the letter Mary rushed to the school.

(8.1) Mary met John after finishing school.

(8.2) Reading books is important.

(8.2.1) Reading books is important for John.

(9) Ski John loves!

(9.1) At the corner was standing a young girl.

These examples do not exhaust all possible cases of structurally relevant IEEs. We

omitted intentionally cases of so-called “parasitic gaps” which we consider highly

rare in real texts to be taken into consideration. We also omitted the case called

OBJect intransitivization that will be discussed in the following section.

Example (6) is a case of an adjectival ADJunct which has the SUBJect as

controller. This may be due to the unsuitability of MOVIE as controller of DRUNK.

If we look at example (6.1), possible controllers are both Mary and John for NAKED.

So the OBJect Mary is taken. However, according to the position of the ADJunct

control may pass to the SUBJect John. In other words, structural control does not

answer only to grammatical criteria, but also to semantic criteria and finally to

positional ones.

Example (7) is an interesting case where we see that the controller may also be

omitted and in that case it needs to be restored from previous discourse. We deal with

such cases below. Notice here the important fact represented by (7.2) where we

included an ungrammatical case – a sentence that will not be found in real texts.

Whenever the AGENT is not lexically expressed nor can be posited by grammatical

principles no control may ensue. Crucially then, in order for a control structure like

the RESULTATIVE infinitive to be expressed in a sentence, some controller needs

also to be there. Then we have cases represented by examples under (8) which are all

gerundives. As it seems, the controller is always the SUBJect disregarding its

position. The copulative construction in (8.2) introduces another type of control, the

one called ARBITRARY control. As can be noticed, (8.2.1) is no longer a case of

Arbitrary control because of the presence of a BENEFICIARY “for John” who

becomes the controller. Final cases are those constituted by so-called inverted focus

structure – example (9) – and locative inversion in (9.1). These cases do not require

the insertion of an empty category but a shallow parser is usually unable to cope with

them appropriately [see 4]. Deep processing will impose an appropriate argument

structure by means of selectional restrictions, but also compute as SUBJect the

inverted NP in the locative construction.

2 Semantically Motivated IEEs

In this section we discuss cases of lexical-semantic IEEs which are also discussed in

the FrameNet project, which is theory of lexical representation and is based on its

underlying linguistic theory, i.e. Constructional Grammar. Differently from our

approach which is mainly computational, this project is descriptive and wrongly

conflates cases of lexical semantics IEEs with cases of grammatical IEEs in the same

typology. Starting from CNI, this class of implicit entities concerns structurally

omitted constituents as can be gathered from the definition given in their Manual

(ibid. p. 54)

Under the term CNI we find three types of IEEs conflated under the same

definition, some of which have already been discussed above. Computationally

speaking, these types require totally different tools and procedures to be activated.

They may be redefined as follows:

2.1 Pronominal Binding Cases

(Big-)PRO cases as found in independent or Adjunct infinitives, participials and

gerunds, i.e. in clauses with an untensed verb.

These cases have already been presented above. The PRO SUBJect inherits lexical

properties associated to the subcategorization frame and may thus be pronominally

bound to a structural controller, if any exist. Otherwise, the PRO is computed as

generic or arbitrary: as a result, PRO cannot possibly be computed as external

pronouns that can corefer in the discourse.

Finally, for these CNI to be computed, their presence is posited by the

Interpretation Component of the system, which recovers Predicate Argument

Structures or PASs by applying grammatical completeness and other principles to the

output of the parser – in our case to c-structure. In our case, the output of the

Interpretation Component are f-structures, i.e. semantically complete PASs. Big-

PROs will then be bound by the Pronominal Binding component of the system, which

only works at sentence level, using structural information and principles of the

grammar.

2.2 Anaphoric Binding Cases

Little-PRO cases for those languages – Romance but not only - that allow a SUBJect

to be left lexically unexpressed in clauses with a tensed verb. The pronoun is added

by the Interpretation Component as above and may be bound at sentence level. In

addition, and differently from big-PRO it may become an external pronoun, which is

then bound at discourse level. In this case, discourse level processing components like

Topic Hierarchy and Centering - that will be presented below - will contribute to find

the appropriate antecedent. This case includes IMPERATIVE mood sentences which

require the SUBJect to be left unexpressed. It also includes COORDINATE structures

omitted understood SUBJect as in the example, “John went out and pro met Mary”

which we comment below, where the omitted SUBJect must be copied from the

previous clause by the grammar embodied by the parsing module, rather than by

interpretation or by lexically related principles. There is always the need to have a

SUBJect expressed with all verb predicates.

2.3 Ellipsis: On the Edge between Syntax and Semantics

We take all cases of ellipsis to be collapsed in a single computational action: copy of

the elliptical material in the place where it is missing. This can only be done once the

complete utterance that precedes the elliptical one is fully parsed. Also it is important

to remind the fact that this process can be spotted only in case there is no ambiguity,

as shown in the examples below:

(10) John went out and [John] met Mary

- SUBJECT null in coordination

(11) John often kisses Mary, and Bill does/will [kiss Mary], too.

- VP ellipsis

(12) John carefully counted the money, and Bill did/will [carefully count the money], too.

- VP ellipsis and Adverb ellipsis

(13) Harry lives in Boston and Mike [lives] in New York

- Gapping

(14) Susie wants to buy a car and my brother [wants [to buy]] a bike.

- LD Gapping

(15) Venice is the city where I live and [where] I work

- Forward Conjunction Reduction

(16) Some have served mussels to Sue while others have [served] swordfish

- PseudoGapping

(17) He takes and never gives back.

- Indefinite Null Instantiation (INI)

(18) What? Who?

- Sluices

As can be easily noticed, the only case in which ambiguity may constitute a

problem is (16), PseudoGapping, which is a case of Auxiliary ellipsis, and English

auxiliaries are ambiguous between lexical and non lexical usage. Before continuing

examining other IEEs, we need to briefly describe the system GETARUNS – which

may be freely downloaded from the www.sigsem.org Main page following the link of

the SharedTask held in Venice and then cliking on GETARUNS. The working of the

system is described in [4].

3 The System GETARUNS

Here we can only point to the fact that the system is organized as the usual pipeline of

modules, divided up into a lower and an upper level, where the lower level computes

sentence level interpretation and records the output in DAGs (Direct Acyclic Graphs).

At this level the system also computes pronominal binding and quantifier raising.

3.1 The Upper Module

GETARUNS, has a highly sophisticated linguistically based semantic module which

is used to build up the Discourse Model. Semantic processing is strongly modularized

and distributed amongst a number of different submodules which take care of Spatio-

Temporal Reasoning, Discourse Level Anaphora Resolution, and other subsidiary

processes like Topic Hierarchy which cooperate to find the most probable antecedent

of coreferring and cospecifying referential expressions when creating semantic

individuals. These are then asserted in the Discourse Model (hence the DM), which is

then the sole knowledge representation used to solve nominal coreference. The

system uses two resolution submodules which work in a sequence: they constitute

independent modules and allow no backtracking. The first one is fired whenever a

free sentence external pronoun is spotted; the second one takes the results of the first

submodule and checks for nominal anaphora. They have access to all data structures

contemporarily and pass the resolved pair, anaphor-antecedent to the following

modules. Semantic Mapping is performed in two steps: at first a Logical Form is

produced which is a structural mapping from DAGs onto unscoped well-formed

formulas. These are then turned into situational semantics informational units, infons

which may become facts or sits. Each unit has a relation, a list of arguments which in

our case receive their semantic roles from lower processing – a polarity, a temporal

and a spatial location index. Inferences can be drawn on the facts repository as will be

discussed below.

3.1.1 Discourse Model Cases

Omitted Agent of passive sentences already discussed above, which we treat as we do

with cases of OBJect intransitivization, i.e. by adding a dummy existential quantifier.

The Agent of passive sentences will then be identified by the semantic processing

module which will look for a similar governing predicate in the context, or previous

stretch of discourse. When the predicate is found the argument will be identified and

the current existential bound to it in the Discourse Model.

3.1.2 Semantic Coreference Cases

There is no need to specify a dummy (big)-PRO, little pro or existential quantifier in

these cases because the missing element is an Adjunct and not an Argument as was

the case with the examples discussed above. So the only way to recover the identity of

the lexical entities coreferred by the optional adjuncts “evoked” by these structures is

to search in the context, or in the previous stretch of discourse – the Discourse Model

– for a similar semantic relation. When an identical predicate is found, the arguments

are recovered and their semantic identifiers used to complete the extended PAS for

the current predicate.

3.1.3 Other Discourse Model Cases

We postulate a semantic treatment of empty deleted OBJect for those transitive verbs

that allow it that is lexical. The solution of the problem lies in the lexical nature of the

phenomenon of OBJect intransitivation, which must be marked for Intransitivization,

i.e. these are transitive verbs that may become Intransitives. Seen that transitive verbs

constitute the great majority of all verbs in any language, and the ones allowing

intransitivization is a small subset, they shall have to be marked so. The empty

OBJect can then be added to the extended PAS by the semantic component. Similar

cases are constituted by the deletion of OBJ2 in ditransitive verbs, as also shown by

example (17) above. So, we prefer to consider the OBJ2 as an existential that needs to

be recovered, and leave the OBJ unexpressed.

4 Implicit Entities and Implicatures

Conversational implicatures and implications in general, are based on an assumption

by the addressee that the speaker is obeying the conversational maxims [7], in

particular the cooperative principle. The well-known example from Levinson [6,

107],

Text 1.

A: Can you tell me the time?

B: Well, the milkman has come.

requires that both interlocutors share the same spatiotemporal location, besides the

same conventions and habits. Not everywhere you can find milkmen go around

delivering milk.

Now consider the following example always from Levinson [6, p.104],

Text 2.

A: I’ve just run out of petrol.

B: Oh; there’s a garage just around the corner.

Here we see that spatiotemporal locations are even more important: if speaker A

needs fuel then the addressee indicates a spatial location, the garage, which in

addition has to be open – hence a temporal location. More on this example below.

So, we would like to regard the mechanism that recovers standard implicatures

and conversational implications in general, as a reasoning process that uses the

knowledge contained in the semantic relations actually expressed in the utterance to

recover hidden or implied relations or events as we call them. This reasoning process

can be partially regarded as a subproduct of an inferential process that takes

spatiotemporal locations as the main component and is triggered by the need to search

for coreferent or cospecifiers to a current definite or indefinite NP head. This could be

interpreted as bridging referential expression entertaining some semantic relation with

previously mentioned entities. In Text (2) the initial inference would be triggered by

the metonymy relation intervening between “petrol” and CAR. At the same time CAR

would be the trigger of the GARAGE reference, always metonymic. If we consider

now Text (1), we see that the request of the current time is itself bound to a

spatiotemporal location. Using the MILKMAN rather than a WATCH to answer the

question, is relatable to spatiotemporal triggers. In fact, in order to infer the right

approximate time, we need to situate the COMING event of the milkman in time,

given a certain spatial location. Thus, it is just the “pragmatic restriction” associated

to SPACE and TIME that is implied in the answer, that may trigger the inference.

More on this topic below.

4.1 The Restaurant Text

To exemplify some of the issues presented above we present a text by Sanford and

Garrod [12;13;14] called the Restaurant text. In this text, entities may be “scenario-

dependent” [15] or main characters that are independent thereof. While the authors

use the text for psychological experimental reasons, we will focus on its

computability. So first of all the sentences making up the text, here below,

Text 3

0. At the restaurant.

1. John went into a restaurant.

2. There was a table in the corner.

3. The waiter took the order.

The text has also been used in the challenge of the Shared Task associated to STEP2008, and

its full analysis is available at the link with the same name under SIGSEM main page.

4. The atmosphere was warm and friendly.

5. He began to read his book.

Here below we will only comment on implicatures and implicit arguments. The

text is also defined as a “psychological statement” text, i.e. it includes sentence (4)

that represents a psychological statement, that is it expresses the feelings and is

viewed from the point of view of one of the characters in the story. The relevance of

the sentence is its role in the assignment of the antecedent to the pronominal

expressions contained in the following sentence. Without such a sentence the

anaphora resolution module would have no way of computing “John” as the

legitimate antecedent of “He/his”. However, in order to capture such information, a

system has to compute Point of View and Discourse Domain on the basis of

Informational Structure and Focus Topic by means of a Topic Hierarchy algorithm

based on [8; 9], which has been lately evaluated in [4].

4.2 Commonsense Reasoning and IEEs

We will concentrate our attention to sentence (3) at first, which is an example of INI.

To account for the fact that whenever a waiter takes an order there is always someone

that makes the order, we compute TAKE_ORDER as a compound verb with an

optional GOAL argument that is the person ORDERing something. The system then

looks for the current Main Topic of discourse or the Focus as computed by the Topic

Hierarchy Algorithm, and associates the semantic identifier to the IEE. This latter

procedure is triggered by the “existential” dummy quantifier associated to the implicit

optional argument. However, another important process has been activated

automatically by the presence of a singular definite NP, “the WAITER”, which is

searched at first in the Discourse Model of entities and properties asserted for the

previous stretch of text. Failure in equality matching activates the bridging

mechanism for inferences which succeeds in identifying the WAITER as a Social

Role in a Restaurant, the current Main Location.

Consider now sentence (2) which introduces a TABLE as main Topic. This type

of sentences is called “presentational” and has the pragmatic role of “presenting” an

entity on the scene of the narration in an abrupt manner, or as Centering would

definite it with a SHIFT move. However, the TABLE does not constitute a suitable

entity to be presented on the scene and the underlying import is triggering the

inference that “someone is SITting at a TABLE”. This inference is guided by the

spatiotemporal component of the system. GETARUNS is equipped with a

spatiotemporal inferential module that asserts Main SpatioTemporal Locations to

anchor events and facts expressed by situational infons. This happens whenever an

explicit lexical location is present in the text. In our case, the location expressed is the

Restaurant. This can either be part of the title or just be derived from the first sentence

of the text, where it has the role of LOCATion argument of the governing verb GO

and the preposition INTO. The second sentence contains an expressed location: the

CORNER. Now, the inferential system will try to establish whether the new location

is either a deictic version of the Main Location; or it is semantically included in the

Main Location, or else it is a new unconnected location that substitutes the previous

one. The “corner” is in a meronymic semantic relation with “restaurant” and thus it is

understood as being a part_of it. This inference is the trigger of the IMPLICATURE

that the TABLE is a metonymy for the SITting event. Consequently, when the system

tries to corefer, cospecify or assert new semantic individuals, it will find an Indefinite

expression “a table” which will not just constitute literally that the text presents a new

entity TABLE, but that the IE is involved with a related event. The Entity implied is

again understood as the Main Topic of current Topic Hierarchy, i.e. JOHN.

The procedure invoked by the system to produce such an implicature makes a call

to WordNet that checks all possible inclusion relations: hyponymy, hyperonymy,

meronymy, etc. Then we look for current main location specified for spatial “place”

locations and recover the Predicate the “restaurant”. Then the semantic index – Idy -

of the Main Topic is searched and passed down to the predicate that will compute the

implicature. The following procedures produce the other semantic index associated to

TABLE and assert its semantic properties; it asserts an inclusion for the location

Table into the Main location.

Then there is the final call which has the task to search for unexpressed relations

intervening in the current spatiotemporal location. To solve this problem in a

principled matter we needed commonsense knowledge organized in a

computationally tractable way. This is what CONCEPTNET 2.1 [11] actually

constitutes. ConceptNet - available at www.conceptnet.org - is the largest freely

available, machine-useable commonsense resource. Organized as a network of semi-

structured natural language fragments, ConceptNet consists of over 250,000 elements

of commonsense knowledge. At present there are 19 semantic relations used in

ConceptNet, representing categories of, inter alia, temporal, spatial, causal, and

functional knowledge. The representation chosen is semi-structured natural language

using lemmata rather than inflected words. The way in which concepts are related

reminds “scripts”, where events may be decomposed in Preconditions, Subevents and

so on, and has been inspired by Cyc [10].

ConceptNet can be accessed in different ways, we wanted a strongly constrained

one. We chose a list of functions that encode pieces of knowledge and use those

functions together with the information available at a certain point of the computation

to derive Implicit Information. In other words, we assume that what is being actually

said hides additional information which however is only implicitely hinted at. What

we need is a predicate constrained by a conceptual function and other predicates. So

first of all the list of functions,

allsceneryevents([‘SubEventOf’,‘FirstSubeventOf’,

‘DesiresEvent’,’Do’,CapableOf’,’FunctionOf’,’UsedFor’,’EventRequiresObject’,Location

Of’]

Then the call that searches ConceptNet for implicit information,

create_infer_rel(NoFr, MainLoc, AgentId, CurrLocatId, Temp, Loc):-

allsceneryevents(CondEvents),

member(Type, CondEvents),

MatchScenery=..[Type, [go, Prep, MainLoc], [Event, Preps, CurrLocat ])

…

Infon=..[fact, EvId, Event, [actor:AgentId, locat: CurrLocatId], 1, tes(Tr3), Loc],

assert(Infon),

If the call has success, we end up by recovering a predicate SIT in the slot Event,

and use this predicate to assert an additional property associated to the Topic of

discourse. So eventually, the system checks for implicatures because it is triggered by

the unsuitability of the current entity – the TABLE – as topic of discourse.

With a similar strategy can be resolved the non-standard implicature involved in

Text 2, that we repeat here below,

Text 2.

A: I’ve just run out of petrol.

B: Oh; there’s a garage just around the corner.

There are a number of missing conceptual links that need to be inferred, as

follows:

Inf1: the CAR has run out of petrol

Inf2: the CAR NEEDS petrol

Inf3: garages SELL PETROL for cars

In addition, in order to use ConceptNet we need to translate “petrol” and “garage”

into “gas/gasoline” and “gas station” respectively. This passage is not just a mere

translation but requires contextual information to tell apart the two meanings

associated to the word “garage” – that is the place where to keep your car, and the

place where to get gas. Now we can query the ontology as we did previously and will

recover the following facts. The whole process starts from the first utterance and uses

RUN OUT OF GAS,

(Do "car" "run out of gas")

Then we can use GAS STATION and CAR to build another query and get,

(Do "car" "get fuel at gas station")

where FUEL and GASoline are in IsA relation. We may still get additional

information on the reason why this has to be done,

(Do "person" "don't want to run out of gas")

(SubeventOf "drive car" "you run out of gas")

(Do "car" "need gas petrol in order to function")

(Do "gas station" "sell fuel for automobile")

These may all constitute additional commonsense knowledge that may be used to

further explain and clarify the implicature.

A brief comment on Schank’s approach [17] – but see also Mueller [19] - and the

restaurant text. Schank introduces scripts and a theory of conceptual dependencies

which are based on primitive actions which are very close to FrameNet’s Frames. In

Schank’s perspective, plans are the means for satisfying goals and they are composed

of scripts. To understand a story one needs scripts and a plan. However, differently

from what we do here, scripts are organized with metadata that contain for instance

preconditions, instrumental relations etc. and we certainly make no use of

preorganized conceptual structures. We do not even make use of Frames, in the literal

sense, in order to produce lexical inferences. All we do is built by the actual

Predicate-Argument Structures instantiated in a given text as it is analyzed and

represented in situational semantics in a Discourse Model by GETARUNS.

5 Conclusions

We have presented a complete treatment of implicit entities and events that

encompasses all possible semantically relevant lexically unexpressed elements. This

has been implemented in a system called GETARUNS which deals with all these

phenomena in a principled way by means a theoretically validated division of labour

between the different modules that make up the whole pipeline. We subdivided IEEs

into different categories according to both theoretical and computational criteria. In

this way grammatical IEEs are taken care before lexically semantically motivated

ones. In turn these latter come before the need to to carry out pronominal binding and

anaphora resolution. Finally, when the semantic components are completing their

mapping and search the preceding Discourse Model for coreferring/cospecifying

entities, procedures that look for implicatures are activated and inferences are fired.

This can only be done in presence of a full-fledged semantic interpretation of the

current utterance, because it is just by means of its PAS that the appropriate implicit

events may be recovered. World knowledge is represented by two repositories: a

generic semantic network like WordNet and the commonsense ontology ConceptNet.

Again, other similar repositories may be used, but the mechanisms to access them

should be the same: no implicature may be recovered without a full semantic

interpretation of the triggering utterances. Dialogues and texts are full of IEEs either

as elliptic material or as implicated events and we are currently experimenting with

Multiparty Meetings Dialogues from ICSI in order to verify what impact may they

have on the overall interpretation process.

References

1. Bresnan, J., 2000. Lexical-Functional Syntax, Blackwell.

2. ComLex:- http://nlp.cs.nyu.edu/comlex

3. Collin F. Baker, Charles J. Fillmore, and John B. Lowe. 1998. The Berkeley FrameNet

project. In Proceedings of COLING-ACL-98, Montreal, Canada.

4. Delmonte R., (2007), Computational Linguistic Text Processing, Nova Science Publishers,

New York.

5. Fellbaum, Christiane, (ed.)(1998), WordNet: An Electronic Lexical Database. MIT Press,

Cambridge MA.

6. Levinson, Stephen C. 1983. Pragmatics. Cambridge, Cambridge University.

7. Grice, H. P., 1975. Logic and Conversation. in P. Cole & J. L. Morgan, Syntax and

Semantics, Vol. 3: Speech Acts. New York : Academic Press, 41-58.

8. Grosz, B. 1981 Focusing and description in natural language dialogues. In Joshi, A., B.

Webber & I. Sag (eds) Elements of discourse understanding. Cambridge UP.

9. Grosz, B. & Sidner, C. 1986 Attention, intentions and the structure of discourse.

Computational Linguistics 12: 175-204.

10. Lenat, D.B. (1995). CYC: A large-scale investment in knowledge infrastructure.

Communications of the ACM, 38(11).

11. Liu, H., Singh, P. (2004). ConceptNet: A Practical Commonsense Reasoning Toolkit.

At:http://web.media.mit.edu/~push/ConceptNet.pdf

12. Sanford, A.J., & Garrod, S. (1998). The role of scenario mapping in text comprehension,

Discourse Processes, 26, 159 - 190.

13. Anderson, A., S. C. Garrod, A. J. Sanford, The accessibility of pronominal antecedents as a

function of episode shifts in narrative text, 1983.

14. The Quarterly Journal of Experimental Psychology Section A, Volume 35, Issue 3 May

1983, pp. 427–440.

15. Sanford, A.J. and Garrod, S.C., Thematic subjecthood and cognitive constraints on

discourse structure, 1988. Journal of Pragmatics, Volume 12, Issues 5-6, December 1988,

pp. 519-534.

16. Sanford, A.J. and Garrod, S.C., 1981. Understanding written language: Exploration in

comprehension beyond the sentence, Wiley, Chichester.

17. Schank R C & Abelson R P. 1977. Scripts, plans, goals and understanding: an inquiry into

human knowledge structures. Hilsdale, NJ: Erlbaum.

18. Sidner, C. 1983. Focusing in the comprehension of definite anaphora. In M. Brady & R.

Berwick (eds) Computational models of discourse. MIT Press, Cambridge, MA: 267-330.

19. Erik T. Mueller, 2007. Modelling Space and Time in Narratives about Restaurants, Literary

and Linguistic Computing, Vol. 22, No. 1, 67-84.