<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-1906">
  <Title>Corpus-based Induction of an LFG Syntax-Semantics Interface for Frame Semantic Processing</Title>
  <Section position="4" start_page="0" end_page="75" type="metho">
    <SectionTitle>
3 LFG for Frame Semantics
</SectionTitle>
    <Paragraph position="0"> Lexical Functional Grammar (Bresnan, 2001) assumes multiple levels of representation. Most prominent are the syntactic representations of c(onstituent)- and f(unctional)-structure. The correspondence between c- and f-structure is defined by functional annotations of rules and lexical entries.</Paragraph>
    <Paragraph position="1"> This architecture can be extended to semantics projection (Halvorsen and Kaplan, 1995).</Paragraph>
    <Paragraph position="2"> LFG f-structure representations abstract away from surface-syntactic properties, by localising arguments in mid- and long-distance constructions, and therefore allow for uniform reference to syntactic dependents in diverse syntactic configurations. This is important for the task of frame annotation, as it abstracts away from aspects of syntax that are irrelevant to frame (element) assignment.</Paragraph>
    <Paragraph position="3"> In (1), e.g., the SELLER role can be uniformly associated with the local SUBJect of sell, even though it is realized as (a.) a relative pronoun of come that controls the SUBJect of sell, (b.) an implicit second person SUBJ, (c.) a non-overt SUBJ controlled by the OBLique object of hard, and (d.) a SUBJ (we) in  VP coordination.</Paragraph>
    <Paragraph position="4"> (1) a. The woman who had come in to sell flowers overheard their conversation.</Paragraph>
    <Paragraph position="5"> b. Don't sell the factory to another company.</Paragraph>
    <Paragraph position="6"> c. It would be hard for him to sell newmont shares.  d. .. we decided to sink some of our capital, buy a car, and sell it again before leaving.</Paragraph>
    <Paragraph position="7"> LFG Semantics Projection for Frames As in a standard LFG projection architecture, we define a frame semantics projection f from the level of fstructure. We define the f -projection to introduce elementary frame structures, with attributes FRAME, FEE (frame-evoking element), and frame-specific role attributes. Figure 2 displays the f-projection for the sentence in Figure 1.4  frame. f is a function of f-structure. The verb auffordern introduces a node f (&amp;quot;) in the semantics projection of &amp;quot;, its local f-structure, and defines its attributes FRAME and FEE. The frame elements are defined as f-projections of the verb's SUBJ, OBJ and OBL OBJ functions. E.g. the SPEAKER role, referred to as ( f(&amp;quot;) SPEAKER), the SPEAKER attribute in the projection f (&amp;quot;) of &amp;quot;, is defined as identical to the f-projection of the verb's SUBJ, f (&amp;quot; SUBJ).</Paragraph>
    <Paragraph position="8"> Frames in Context The projection of frames in context can yield connected frame structures. In Figure 2, Gespr&amp;quot;ach fills the MESSAGE role of REQUEST, but it also introduces a frame of its own, CONVERSATION. Thus, the CONVERSATION frame, by coindexation, is an instantiation, in context, of the MESSAGE of REQUEST.</Paragraph>
    <Paragraph position="9"> Co-description vs. description-by-analysis In the co-description architecture we just presented f- and s-structure equations jointly determine the valid analyses of a sentence. Analyses that do not satisfy both f- and s-structure constraints are inconsistent and ruled out.</Paragraph>
    <Paragraph position="10"> An alternative to co-description is semantics construction via description-by-analysis (DBA) (Halvorsen and Kaplan, 1995). Here, semantics is built on top of fully resolved f-structures. F-structures that are consistent with semantic mapping constraints are semantically enriched - remaining analyses are left untouched.</Paragraph>
    <Paragraph position="11"> Both models are equally powerful - yet while copred(X,auffordern), null</Paragraph>
    <Paragraph position="13"> description integrates the semantics projection into the grammar and parsing process, DBA keeps it as a separate module. Thus, with DBA, semantics does not interfere with grammar design and can be developed separately. The DBA approach also facilitates the integration of external semantic knowledge sources (such as word senses or named entity types).</Paragraph>
    <Paragraph position="14"> DBA by transfer We realise the DBA approach by way of a term-rewriting transfer system that is part of the XLE grammar processing platform. The system represents f-structures as sets of predicates which take as arguments variables for f-structure nodes or atomic values. Transfer is defined as a sequence of ordered rules. If a rule applies to an input set of predicates, it defines a new output set.</Paragraph>
    <Paragraph position="15"> This output set is input to the next rule in the cascade. A rule applies if all terms on its left-hand side match some term in the input set. The terms on the right hand side (prefixed '+') are added to the input set. There are obligatory (==&gt;) and optional (?=&gt;) rules. Optional rules introduce two output sets: one results from application of the rule, the other is equal to the input set.</Paragraph>
    <Paragraph position="16"> Figure 4 displays a transfer rule that corresponds to the co-description lexical entry of Figure 3. For matched f-structure nodes (pred, subject, object, oblique object) it defines a f-projection (by predicate 's::f ') with new s-structure nodes. For these, we define the frame information (FRAME, FEE) and the linking of semantic roles (e.g., the f-projection SemA of the SUBJ is defined as the SPEAKER role of the head's semantic projection SemX).</Paragraph>
    <Paragraph position="18"> % semantic projection for (each) FE of FEE project fe of fee(FeeID, Frame, FeID, Role) :: ti-id(X,FeeID), 's::'(X,S X), frame(S X,Frame),</Paragraph>
    <Paragraph position="20"/>
  </Section>
  <Section position="5" start_page="75" end_page="75" type="metho">
    <SectionTitle>
4 Corpus-based induction of an LFG
</SectionTitle>
    <Paragraph position="0"> frame semantics interface</Paragraph>
    <Section position="1" start_page="75" end_page="75" type="sub_section">
      <SectionTitle>
4.1 Porting SALSA annotations to LFG
</SectionTitle>
      <Paragraph position="0"> A challenge for corpus-based induction of a syntax-semantics interface for frame assignment is the transposition of the corpus annotations from a given syntactic annotation scheme to the target syntactic framework. The basis for our work are annotations of the SALSA/TIGER corpus (Erk et al., 2003), encoded in an XML annotation scheme that extends the syntactic TIGER XML annotation scheme.</Paragraph>
      <Paragraph position="1"> The TIGER treebank has been converted to a parallel LFG f-structure corpus (Forst, 2003). The SALSA/TIGER and LFG-TIGER corpora could be used to learn corresponding syntactic paths in the respective structures. Thus, we could establish the paths of frame constituting elements in the SALSA/TIGER corpus, and port the annotations to the corresponding path in the LFG-TIGER corpus.</Paragraph>
      <Paragraph position="2"> However, we could apply a more precise method, by exploiting the fact that the LFG-TIGER corpus preserves the original TIGER constituent identifiers, as f-structure features TI-ID (see Fig. 7). We use these 'anchors' to port the SALSA annotations to the parallel LFG-TIGER treebank. Thus, in a first step we extend the latter to an LFG corpus with frame semantics projection. From the extended corpus we induce general LFG frame assignment rules.</Paragraph>
      <Paragraph position="3"> This will be described in more detail in Section 4.2.</Paragraph>
      <Paragraph position="4"> Porting annotations by transfer For each sentence we extract the constituent identifiers of frame constituting elements in the SALSA XML annotations (cf. Figure 5). This information is coded into transfer rules, where we refer to the corresponding TI-ID features in the f-structure as anchors to project the frame information for a given frame annotation  instance. The first transfer rule (template) in Figure 6 defines the semantic projection of the FEE, where the correct f-structure location is referenced by the feature TI-ID. Subsequent rules - one for each role to be assigned - define the given semantic role as an argument of the FEE's semantic projection, again using the TI-IDs of the FEE and FE as anchors.</Paragraph>
      <Paragraph position="5"> We generate these frame projection rules for each sentence in the SALSA/TIGER corpus, and apply them to the corresponding f-structure in the LFG-TIGER corpus. The result is an LFG corpus with frame semantic anntations (cf. Figures 7 and 8).</Paragraph>
      <Paragraph position="6"> The basic structure of frame-inducing rules in Figure 6 was refined to account for special cases: Coordination For frame elements that correspond to coordinated constituents, as in Figure 9, we project a semantic role that records a set of semantic predicates (REL), one for each of the conjuncts.</Paragraph>
      <Paragraph position="7">  scheme allows for underspecification, to represent unresolved word sense ambiguities or optionality (Erk et al., 2003). In a given context, a predicate may evoke alternative frames (i.e. word senses), where it is impossible to decide between them.</Paragraph>
      <Paragraph position="8"> E.g. the verb verlangen (demand) may convey the meaning of REQUEST, but also COMMERCIAL TRANSACTION. Such cases are annotated with  alternative frames, which are marked as elements of an 'underspecification group'. Underspecification may also affect frame elements of a single frame.</Paragraph>
      <Paragraph position="9"> A motion (Antrag), e.g., may be both MEDIUM and SPEAKER of a REQUEST. Finally, a constituent may or may not be interpreted as a frame element of a given frame. It is then represented as a single element of an underspecification group.</Paragraph>
      <Paragraph position="10"> We model underspecification as disjunction, which is encoded by optional transfer rules that create alternative (disjunctive) contexts. Optionality is modeled by a single optional rule. Figure 10 displays the result of underspecified frame element assignment in an f-structure chart (Maxwell and Kaplan, 1989). Context c1 displays the reading where Antrag is assigned the SPEAKER role, alternatively, in context c2, it is assigned the role MEDIUM.</Paragraph>
      <Paragraph position="11"> In a symbolic account disjunction doesn't correctly model the intended meaning of underspecification. Yet, a stochastic model for frame assignment should render the vagueness involved in underspecification by close stochastic weights. Thus, under-specified annotation instances provide alternative frames in the training data and can be used for fine-grained evaluation of frame assignment models.</Paragraph>
      <Paragraph position="12"> Multiword Expressions The treatment of multiword expressions (idioms, support constructions) requires special care. For idioms, the constituting elements are annotated as multiple frame evoking elements (cf. Figure 11 for &amp;quot;uber die Ladentheke gehen - go over the counter (being sold)). We define semantic projections for the individual components: the main frame evoking predicate (FEE) and the idiom-constituting words, which are recorded in a set-valued feature FEE-MWE. Otherwise, idioms are treated like ordinary main verbs. E.g., like sell, the expression triggers a COMMERCE SELL frame with the appropriate semantic roles, here GOODS.</Paragraph>
      <Paragraph position="13"> Asymmetric Embedding Another type of non-isomorphism between syntactic and semantic rep-Figure 12: Asymmetric embedding (example (2)) resentation occurs in cases where distinct syntactic constituents are annotated as instantiation of a single semantic role. In (2), PP and NP are annotated as the MESSAGE of a STATEMENT, since they jointly convey its content. Projecting distinct constituents to a single semantic node can, however, lead to inconsistencies, especially if both constituents independently project semantic frames.</Paragraph>
      <Paragraph position="14"> (2) Der Geschaeftsfuehrer gab [PP MO als Grund fuer die Absage] [NP OBJ Terminnoete] an.</Paragraph>
      <Paragraph position="15"> The director mentioned [time conflicts] [as a reason for cancelling the appointment] In the SALSA annotations asymmetric embedding at the semantic level is the typical pattern for such double-constituent annotations. I.e., for (2), we assume a target frame structure where the MESSAGE of STATEMENT points to the PP - which itself projects a frame REASON with semantic roles CAUSE for Terminn&amp;quot;ote, and EFFECT for Absage.</Paragraph>
      <Paragraph position="16"> Such multiple-constituent annotations arise in cases where frame annotations are partial: since corpus annotation proceeds frame-wise, in (2) the REASON frame may not have been treated yet.</Paragraph>
      <Paragraph position="17"> Moreover, annotators are in general not shown complete(d) sentence annotations.</Paragraph>
      <Paragraph position="18"> We account for these cases by a simulation of functional uncertainty equations, which accommodate for a potential embedded frame within either one of the otherwise re-entrant constituents. We apply a transfer rule set that embeds one (or the other) of the two constituent projections as an embedded role of an unknown frame, to be evoked by the respective 'dominating' node. We introduce an 'unknown' role ROLE for the embedded constituent, which is to be interpreted as a functional uncertainty path over variable semantic roles.</Paragraph>
      <Paragraph position="19"> Figure 12 displays the alternative (hypothetical) frame structures for (2), where the second one with FRAME instantiated to REASON and ROLE to CAUSE - corresponds to the actual reading.</Paragraph>
      <Paragraph position="20"> Overview of data Our current data set comprises 12436 frame annotations for 11934 sentences. Table 1 gives frequency figures for the special phecoord usp mwe asym &gt;dbl all  nomena: coordination, underspecification, multi-word expressions and double constituents (asym).5 We successfully ported 11713 frame annotations to the LFG-TIGER corpus, turning it into an LFG corpus with frame annotations.</Paragraph>
    </Section>
    <Section position="2" start_page="75" end_page="75" type="sub_section">
      <SectionTitle>
4.2 Inducing frame projection rules
</SectionTitle>
      <Paragraph position="0"> From the enriched corpus we extract lexical frame assignment rules that - instead of node identifiers use f-structure descriptions to identify constituents and map them to frame semantic roles. These rules can then be applied to the f-structure output of free LFG parsing, i.e. to novel sentences.</Paragraph>
      <Paragraph position="1"> We designed an algorithm for extracting f-structure paths between pairs of f-structure nodes that correspond to the s-structure of the frame evoking element and one of its semantic roles, respectively.</Paragraph>
      <Paragraph position="2"> Table 2 gives an example for the frame projection in Figure 13. Starting from the absolute f-structure path (f-path) for (the f-structure projecting to) the FEE MITTEILEN we extract relative f-paths leading to the roles MESSAGE and SPEAKER. The f-path for the MESSAGE (&amp;quot;OBJ) is local to the f-structure that projects to the FEE. For the SPEAKER we identify two paths: one local, the other non-local. The local f-path (&amp;quot;SUBJ) leads to the local SUBJ of mitteilen in Figure 13. By co-indexation with the SUBJ of versprechen we find an alternative non-local path, which we render as an inside-out functional equation ((XCOMP&amp;quot;) SUBJ).</Paragraph>
      <Paragraph position="3"> Since f-structures are directed acyclic graphs, we use graph accessibility to distinguish local from non-local f-paths. In case of alternative local and non-local paths, we choose the local one. From alternative non-local paths, we chose the one(s) with shortest inside-out subexpression.</Paragraph>
      <Paragraph position="4"> Generating frame assignment rules We extracted f-path descriptions for frame assignment from the enriched LFG-TIGER corpus. We compiled 9707 lexicalised frame assignment rules in the format of Figure 4. The average number of distinct assignment rules per FEE is 8.38. Abstracting over the FEEs, we obtain 7317 FRAME-specific rules, with an average of 41.34 distinct rules per frame.</Paragraph>
      <Paragraph position="5"> Due to the surface-oriented TIGER annotation format, the original annotations contain a high number of non-local frame element assignments that 5Role assignment to more than two constituents (&gt;dbl) constitute a rather disparate set of data we do not try to cover.</Paragraph>
    </Section>
  </Section>
  <Section position="6" start_page="75" end_page="75" type="metho">
    <SectionTitle>
FEE XCOMP PRED &amp;quot;
MSG XCOMP OBJ &amp;quot;OBJ local
SPKR SUBJ (XCOMP&amp;quot;)SUBJ nonlocal
XCOMP SUBJ &amp;quot;SUBJ local
</SectionTitle>
    <Paragraph position="0"> are localised in LFG f-structures. The f-paths extracted from the enriched LFG corpus yield 12.82% non-local (inside-out) vs. 87.18% local (outside-in) frame element assignment rules.</Paragraph>
    <Paragraph position="1"> As an alternative rule format, we split frame assignment into separate rules for projection of the FEE and the individual FEs. This allows assignment rules to apply in cases where the f-structure does not satisfy the functional constraints for some FE. This yields improved robustness, and accounts for syntactic variability when applied to new data. For this rule format, we obtain 960 FEE assignment rules, and 8261 FEE-specific FE assignment rules. Abstracting over the FEE, this reduces to 4804 rules.6</Paragraph>
    <Section position="1" start_page="75" end_page="75" type="sub_section">
      <SectionTitle>
4.3 Reapplying frame assignment rules
</SectionTitle>
      <Paragraph position="0"> We reapplied the induced frame assignment rules to the original syntactic LFG-TIGER corpus, to control the results. The results are evaluated against the frame-enriched LFG-TIGER corpus that was created by explicit node anchoring (Sec. 4.1). We applied 'full frame rules' that introduce FEE and all FEs in a single rule, as well as separated FEE and FE rules. We applied all rules for a given frame to any sentences that had received the same frame in the corpus. We obtained 93.98% recall with 25.95% precision (full frame rules), and 94.98% recall with 45.52% precision (split rules), cf. Table 3.a. The low precision is due to overgeneration of the more general abstracted rules, which are not yet controlled by statistical selection. We measured an ambiguity of 8.46/7.83 frames per annotation instance.</Paragraph>
      <Paragraph position="1"> 6In the future we will experiment with assignment rules that are not conditioned to FEEs, but to frame-specific syntactic descriptions, to assign frames to 'unknown' lexical items. full frame rules FEE and FE rules rec prec amb rec prec amb  We finally apply the frame assignment rules to original LFG parses of the German LFG grammar. The grammar produces f-structures that are compatible with the LFG-TIGER corpus, thus the syntactic constraints can match the parser's f-structure output. In contrast to the LFG-TIGER corpus, the grammar delivers f-structures for alternative syntactic analyses. We don't expect frame projections for all syntactic readings, but where they apply, they will create ambiguity in the semantics projection.</Paragraph>
      <Paragraph position="2"> We applied the rules to the parses of 6032 corpus sentences. Compared to the LFG-TIGER corpus we obtain lower recall and precision (Table 3.b) and a higher ambiguity rate per sentence. Drop in precision and higher ambiguity are due to the higher ambiguity in the syntactic input. Moreover, we now apply the complete rule set to any given sentence. The rules can thus apply to new annotation instances, and create more ambiguity. The drop in recall is mainly due to overgenerations by automatic lemmatisation and functional assignments to PPs in the TIGER-LFG corpus, which are not matched by the LFG parser output. These mismatches will be corrected by refinements of the TIGER-LFG treebank.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>