<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-2009">
  <Title>Recovering Coherent Intepretations Using Semantic Integration of Partial Parses</Title>
  <Section position="4" start_page="0" end_page="4" type="metho">
    <SectionTitle>
3 The Grammar
</SectionTitle>
    <Paragraph position="0"> The grammar rules used for analysis are represented using a cognitively motivated grammar formalism called Embodied Construction Grammar (Bergen and Chang, 2002). The basic linguistic unit in any construction grammar(Goldberg, 1995) is the construction. A construction is a form-meaning pair. Each pair is a structured mapping from a lexical/syntactic pattern to its corresponding semantic and pragmatic properties.</Paragraph>
    <Paragraph position="1"> Construction grammar rejects the assumption that syntax and semantic are separate processing modules(Fillmore et al., 1988). Morphemes, idioms and standard syntactic constructions like subjectauxiliary-inversion are all represented by the same kind of object - the construction. Construction Grammar defines grammaticality as a combination of syntactic and semantic well-formedness. Using Embodied Construction Grammar as the linguistic substrate therefore requires that syntactic and semantic analysis happen simultaneously.</Paragraph>
    <Paragraph position="2"> To describe constructions precisely, Embodied Construction Grammar (ECG) combines a grammar formalism and knowledge representation language within a unification-based framework. This makes it possible for both constructions and framebased, schematic knowledge (Fillmore, 1982) to be expressed succinctly in the same formalism. Linking the grammar into frame-based meaning is what makes semantic integration possible (see section 5).</Paragraph>
    <Paragraph position="3"> ECGhas twobasic units: the schema and the construction. null Constructions, as discussed, are form-meaning pairs, while schemas are used to represent meaning( like frames or image schemas  ).</Paragraph>
    <Paragraph position="4"> Schemas and constructions have roles which can be assigned an atomic value (with )or coindexed (with !).</Paragraph>
    <Paragraph position="5"> Schemas and constructions are arranged into inheritance hierarchies with the subcase of keyword.</Paragraph>
    <Paragraph position="6"> Theself keyword lets a schema or construction be self-referential  .</Paragraph>
    <Paragraph position="7"> To make this more concrete, figure 2 shows the Throw lexical construction and its associated schemas. Every construction in ECG has a form block and a meaning block. These blocks are themselves special roles that are accessed using the f and m subscripts. In the case of the Throw construction, its form pole constrains its orthography feature to the string throw. Its meaning pole is type constrained (using the colon) to be of type (Throw-Action) schema.</Paragraph>
    <Paragraph position="8"> The Throw-Action schema has roles for the thrower and the throwee. These roles correspond to the semantic arguments of a throw predicate. Roles can type constrain their fillers, and in the case of Throw-Action,thethrower must be of type Animate while the throwee is constrained to be of type  Throw-Action schema evokes the Cause-Motion-Frame and their roles are coindexed. The SPG schema is a structured representation of a path with a source (starting point), path (way traveled) and goal (end point).</Paragraph>
    <Section position="1" start_page="2" end_page="4" type="sub_section">
      <SectionTitle>
Physical-Object
</SectionTitle>
      <Paragraph position="0"> .</Paragraph>
      <Paragraph position="1"> Unique to the ECG formalism is the evokes oper- null Clearly these selectional restrictions do not apply in all cases. One can certainly throw a tantrum, for example. Treatment of metaphorical usage is beyond the scope of both this paper and the system being described.</Paragraph>
      <Paragraph position="2"> ator. The evokes operator makes the evoked schema locally available under the given alias. The Throw-Action schema evokes its frame, the Cause-Motion-Frame schema.</Paragraph>
      <Paragraph position="3"> The Cause-Motion-Frame schema is the ECG representation of FrameNet's Cause-Motion frame(Baker et al., 1998; The FrameNet Project, 2004). Because Throw is a lexical unit associated with this frame, the corresponding Throw-Action schema evokes the Cause-Motion-Frame schema so that their roles can be coindexed. In this case, the thrower is bound to the agent while the throwee is bound to the theme.</Paragraph>
      <Paragraph position="4"> The only commitment an evoking schema makes when it evokes some other schema is that the two schemas are related. In this way, The evokes operator provides a mechanism for underspecifying the relation between the evoking schema and theevoked schema. Constraints can then be used to make this relation precise.</Paragraph>
      <Paragraph position="5"> Semantic frames are a good example of where this ability to underspecify is useful. The lexical item throw, for example, only profiles some of the roles in the Cause-Motion frame. Using the evokes operator and constraints, the Throw-Action schema can pick out which of these roles are relevant to it. Evokes thus provides an elegant way to incorporate frame-based information.</Paragraph>
      <Paragraph position="6"> Constructions with constituents are quite similar to their lexical counterparts with two exceptions. The first is the addition of a constructional block to define the constituents. The second is that the form block is now used to define the ordering of the construction's constituents.</Paragraph>
      <Paragraph position="7"> Figure 3 shows an example of a construction with constituents- the active ditransitive construction. Within the construction grammar literature, Goldberg (Goldberg, 1995) argues that the ditransitive pattern is inextricably linked with the notion of giving. This is represented in ECG by constraining the meaning pole of the ditransitive construction to be of type Giving-Frame.</Paragraph>
      <Paragraph position="8"> The Active-Ditransitive  construction has four constituents, one for each grammatical function. Its first constituent is named subject, for example, and isconstrained tobeaRefExp(referring expression)  This representation is intentionally naive in regards to the relation between active and passive. Not only is this construction easier to describe in this form, but the language encountered by the model is sufficiently simple such that the constructions do not need to be written in full generality. Though for adult language, thisis not the case. For a detailed description of how argument structure constructions can be represented compositionally see (Bryant, 2004).</Paragraph>
      <Paragraph position="9">  The form block constrains the ordering on the constituents. The meets relation means that its left argument must be directly before its right argument. In this construction, for example, the subject constituent must be directly before the verb constituent. with the semantic type Animate.</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="4" end_page="4" type="metho">
    <SectionTitle>
4 Semantic Chunking
</SectionTitle>
    <Paragraph position="0"> Chunkers (partial parsers) (Abney, 1996) use finitestate-machines (FSM) arranged into levels, to reliably recognize nonrecursive chunks of syntax. With this approach, each finite state recognizer corresponds to a simple syntactic rule. The levels control the interaction between recognizers, with higher-level recognizers depending on lower-level recognizers. null The semantic chunker that is integrated into the language analysis system uses the same processing scheme as Abney-style partial parsers, extending it to recognize the syntax and semantics associated with ECG constructions. This means that syntactic processing and semantic processing happen simultaneously. As a consequence, semantic information that looks like an NP and with a meaning pole that refers to something.</Paragraph>
    <Paragraph position="1"> is easily integrated to help minimize ambiguity.</Paragraph>
    <Paragraph position="2"> Constructions require a very different treatment than the simple syntactic patterns recognized by FSMs. In addition to the straightforward extension of the Abney algorithm to perform unification as well as using a chart, each construction is compiled into a construction recognizer.</Paragraph>
    <Paragraph position="3"> A construction recognizer searches for its constituents in the input utterance and chart. In addition to satisfying ordering constraints, candidate constituents must also satisfy the type and coindexation constraints associated with the construction. Because of this complexity, construction recognizers are more complicated than FSMs, implementing a backtracking algorithm to search for compatible constituents. For more information about the matching algorithm, see (Bryant, 2003).</Paragraph>
  </Section>
  <Section position="6" start_page="4" end_page="7" type="metho">
    <SectionTitle>
5 Integration Using Structure Merging
</SectionTitle>
    <Paragraph position="0"> Without a complete analysis of an utterance, the system must infer exactly how a set of local, partial semantic structures best fit together into a coherent, global analysis of the utterance. The approach taken here is an abductive one in that it assumes compatible structures are the same and merges them.</Paragraph>
    <Paragraph position="1"> Motivation for such an approach come from both linguistics and computational approaches to understanding. On the linguistic side, one needs to look no further than what Paul Kay calls the Parsimony Principle (Kay, 1987). The Parsimony Principle states that &amp;quot;Whenever it is possible to identify two roles in a text, the ideal reader does so&amp;quot;.</Paragraph>
    <Paragraph position="2"> On the computational side, information extraction systems like FASTUS (Hobbs et al., 1996) use an abductive structure merging approach to build up templates describing particular kinds of events like corporate transactions. Their mechanism was intended to quickly build up consistency across utterances. This approach generalizes FASTUS' approach to work on semantic frames within utterances as well as across utterances.</Paragraph>
    <Section position="1" start_page="4" end_page="7" type="sub_section">
      <SectionTitle>
5.1 Structure Merging Examples
</SectionTitle>
      <Paragraph position="0"> This section illustrates how an extragrammatical utterance and an ungrammatical utterance from the CHILDES corpus (MacWhinney, 1991) can successfully be interpreted using semantic integration through structure merging.</Paragraph>
      <Paragraph position="1"> Naomi, a child in the study, wants some flowers.</Paragraph>
      <Paragraph position="2"> Herfather then responds withIwillgive you one one flower with a restart in the description of the final NP chunk. The result of this utterance's semantic chunking analysis is shown in Figure 4.</Paragraph>
      <Paragraph position="3"> The analysis generates two semantic chunks without any links between them. The first chunk  merged because Flower is a subtype of Entity and they had the same value for the distribution feature. (the GiveAction and coindexed GivingFrame) corresponds to the I will give you one phrase in which the Giving frame's roles are filledbythespeaker, the addressee and an Entity schema corresponding to the word one.Theone flower chunk corresponds to the Flower schema.</Paragraph>
      <Paragraph position="4"> Figure 5 shows the integrated analysis. Because the Entity schema and the Flower schema had compatible types  and features, the system assumed that they were the same structure. As a consequence, semantic integration generates a complete analysis. A second more complex example illustrates some of the computational and conceptual issues associated with structure merging: Sometime later, Naomi's father is reading a book to Naomi when he utters the following ungrammatical phrase: The lamb is looking a donkey.</Paragraph>
      <Paragraph position="5"> Figure 6 shows the chunk analysis of this utter- null The lamb is looking the donkey before semantic integration. Notice that theDonkey schema isnot connected to the rest of the analysis.</Paragraph>
      <Paragraph position="6"> ance. This example generates two chunks because the subcategorization frame of the lexical item look is not satisfied by the chunk adonkey.</Paragraph>
      <Paragraph position="7"> Running semantic integration on this set of chunks, however, results in two different possible integrations. As shown in Figures 7 and 8, the donkey can be merged with the ground or the phenomenon the Scrutiny frame  This ambiguity corresponds to whether the missing word in the utterance was intended to be at or for.</Paragraph>
      <Paragraph position="8"> Notice that either integration would be an acceptable interpretation. In other words, these two integrations are equivalent in terms of their semantic acceptability. Certainly, however, leaving the a donkey structure unmerged isworsebecause moreofthe core elements of the frame would be unfilled. The semantic density heuristic (covered in the next section) formalizes this intuition.</Paragraph>
      <Paragraph position="9"> While conceptually it is satisfying for both interpretations to be acceptable, computationally speaking, it is also worrisome. Taking a single analysis and turning it into two because of a single ambiguity signals the possibility of structure merging being NP-hard. For the short utterances associated with child language, this is not problematic. For adult  The FrameNet project defines the Scrutiny frame as: This frame concerns a Cognizer (a person or other intelligent being) paying close attention to something, the Ground, in order to discover and note its salient characteristics. The Cognizer may be interested in a particular characteristic or entity, the Phenomenon, that belongs to the Ground or is contained in the Ground (or to ensure that such a property of entity is not  The lamb is looking the donkey after semantic integration. Notice that the Donkey schema is now merged with the ground role. This corresponds to the The lamb is looking at the donkey.</Paragraph>
      <Paragraph position="10">  The lamb is looking the donkey after semantic integration. Notice that the Donkey schema is now merged with the phenomenon role. This corresponds to the The lamb is looking for the donkey. language, however, this could be a serious issue. As such, an open area of research is how to design approximate merging strategies.</Paragraph>
    </Section>
  </Section>
  <Section position="7" start_page="7" end_page="7" type="metho">
    <SectionTitle>
6 Semantic Density
</SectionTitle>
    <Paragraph position="0"> The key insight behind our approach to making the integrated analyses is the realization that every utterance is trying to communicate a scene (more formally speaking, this scene is a semantic frame).</Paragraph>
    <Paragraph position="1"> Nowassuming that a better analysis is one that more fully describes the scene, one way to compare analyses is by how completely specified the frame is.</Paragraph>
    <Paragraph position="2"> Those analyses that fill out more of the frame roles should be preferred to those that fill out fewer roles.</Paragraph>
    <Paragraph position="3">  frame on the top has a semantic density score of .75 and the Commercial-Event on the bottom has a score of 1. Thus the second frame would be considered better because more of the frame elements are filled.</Paragraph>
    <Paragraph position="4"> This is the motivation for the ranking heuristic that we call semantic density.</Paragraph>
    <Paragraph position="5"> Semantic density compares constructional analyses based upon their semantic content. Analyses that have a higher ratio of filled slots to total slots in their semantic analysis are considered better analyses according to semantic density. Figure 9 shows a simple example of the semantic density metric in use.</Paragraph>
    <Paragraph position="6"> Let's reconsider the example from the last section regarding The lamb is looking a donkey.In that case, there were two possible integrations, one where the donkey was the ground and the other where the donkey was the phenomenon. Applying the semantic density metric to those two competing analyses shows that they have equivalent semantic density. This is consistent with the intuition that either analysis is an equally acceptable interpretation of the input utterance.</Paragraph>
    <Paragraph position="7"> The higher-level point here is that there are many ways to semantically analyze a given utterance, and some ways are better than others. While the two semantically integrated interpretations of the lamb sentence were equally good, both were better than the unintegrated utterance. Given a preference for one interpretation over the other, it makes sense to consider semantic interpretation to be a graded phenomenon much like grammaticality.</Paragraph>
    <Paragraph position="8"> Keller (Sorace and Keller, to appear) defines the graded nature of grammaticality in terms of constraint reranking within an optimality theory context. Since structure merging and semantic density also define a gradient, they could also be stated within an optimality framework, with the most dense analyses being considered optimal.</Paragraph>
  </Section>
  <Section position="8" start_page="7" end_page="8" type="metho">
    <SectionTitle>
7 Related Work
</SectionTitle>
    <Paragraph position="0"> Semantic analysis has a long tradition within NLP.</Paragraph>
    <Paragraph position="1"> While a broad overview describing logic, frames and semantic networks is beyond the scope of this paper  , this work builds on the frame-based tradition as well as the tradition of robust analysis to make progress towards robust semantic analysis. One related traditional approach is Lexical Functional Grammar ((Dalyrimple, 2001)). LFG introduced notions of completeness and coherence for its feature structures. These principles (intuitively speaking) require that certain features of an analysis (including all the semantic arguments of a predicate) be filled before an analysis can be considered well formed. Such a constraint is akin to a binary version of the semantic density metric.</Paragraph>
    <Paragraph position="2"> Structure merging also builds on historical work. Even before FASTUS, the employment of the parsimony principle within understanding systems goes back to the completely unrelated FAUSTUS system (Norvig, 1987). Norvig's work, however, used graph based algorithms within a semantic network to perform abductive inference.</Paragraph>
  </Section>
  <Section position="9" start_page="8" end_page="8" type="metho">
    <SectionTitle>
8 Future Work
</SectionTitle>
    <Paragraph position="0"> While implementation is complete, only initial testing of the analyzer against the CHILDES data has been started. To get more complete results, a test grammar covering the parents' utterances must be completed (or learned). Once this has been finished, the number of semantic relations correctly extracted with and without structure merging can be measured.</Paragraph>
    <Paragraph position="1"> If verified, the ideas in this paper would have to be appropriately extended to adult language. The structure merging phase, for example, would have to redefined to use an approximate search algorithm find a good integration. Investigation of the approach described by Beale (Beale et al., 1996) seems promising.</Paragraph>
    <Paragraph position="2"> Extremely intriguing is the extension of the semantic density metric. Currently, it is merely a first approximation of semantic preference. One obvious direction is weight different frame roles in accordance with lexical preferences. According to FrameNet data for the lexical item look, the phenomenon frame role is 50% more likely to be expressed than the ground role (The FrameNet Project, 2004). By including such preferences, a more complete notion of semantic preference can  See (Jurafsky and Martin, 2000) for such an overview or (Allen, 1995) for an introduction with logic.</Paragraph>
    <Paragraph position="3"> be defined. Narayanan and Jurafsky (Narayanan and Jurafsky, 1998) take the first steps in this direction, integrating syntactic and semantic preferences within a probabilistic framework for a small scale preference task.</Paragraph>
  </Section>
class="xml-element"></Paper>