We use the Xerox Incremental Parser (XIP, (Aït et al., 2002)) to perform robust and deep syntactic analysis. Deep syntactic analysis consists here in the construction of a set of syntactic relations from an input text. These relations, labeled with deep syn­tactic functions, link lexical units of the input text and/or more complex syntactic domains that are constructed during the processing (mainly chunks, see (Abney, 1991)).

Moreover, together with surface syntactic relations, the parser calculates more sophisticated relations using derivational morphologic properties, deep syntactic properties, and some limited lexical se­mantic coding (Levin's verb class alternations, see (Levin, 1993)), and some elements of the Framenetclassification, (Ruppenhofer et al., 2006)). These deep syntactic relations correspond roughly to the agent-experiencer roles that is subsumed by the SUBJ-N relation and to the patient-theme role sub­sumed by the OBJ-N relation, see (Brun and Ha-gège, 2003). Not only verbs bear these relations but also deverbal nouns with their corresponding argu­ments.

Here is an example of an output (chunks and deep syntactic relations):

Lebanon still wanted to see the implementation of a UN resolution TOP{SC{NP{Lebanon} FV{still wanted}} IV{to see} NP{the implementation} PP{of NP{a UN resolution}} .}

MOD_PRE(wanted,still)

MOD_PRE(resolution,UN)

MOD_POST(implementation,resolution)

COUNTRY(Lebanon)

ORGANISATION(UN)

EXPERIENCER_PRE(wanted,Lebanon)

EXPERIENCER(see,Lebanon)

CONTENT(see,implementation)

EMBED_INFINIT(see,wanted)

OBJ-N(implement,resolution)

Our parser includes a module for "standard" named entity recognition, but needs to be adapted to handle named entity metonymy. Following the guidelines of the SemEval task #8, we performed a corpus study on the trial data in order to detect lexi­cal and syntactic regularities triggering a metonymy, for both location names and organization names. For example, we examined the subject relation be­tween organizations or locations and verbs and we then classify the verbs accordingly: we draw hy­pothesis like "if a location name is the subject of a verb referring to an economic action, like import, provide, refund, repay, etc., then it is a place-for-people". We adapted our parser by adding dedicated lexicons that encode the information collected from the corpus and develop rules modifying the interpre­tation of the entity, for example:

We focus our study on relations like subject, object, experiencer, content, modifiers (nominal and prepo­sitional) and attributes. We also capitalize on the already-encoded lexical information attached to verbs by the parser, like communication verbs like say, deny, comment, or categories of the FrameNet Experiencer subject frame, i.e. verbs like feel, sense, see. This information was very useful since experi-encers denote persons, therefore all organizations or locations having an experiencer role can be consid­ered as organization-for-members or place-for-people. Here is an example of output, when apply­ing the modified parser on the following sentence: "It was the largest Fiat everyone had ever seen".

ORG-FOR-PRODUCT(Fiat)

MOD_PRE(seen,ever)

SUBJ-N_PRE(was,It)

EXPERIENCER_PRE(seen,everyone)

SUBJATTR(It,Fiat) QUALIF(Fiat,largest)

Here, the relation QUALIF(Fiat, largest) triggers the metonymical interpretation of "Fiat" as org-for-product.

This first development step is the starting point of our methodology, which is completed by a non-supervised distributional approach described in the next section.

The distributional approach proposes to establish a distance between words depending on there syntac­tic distribution.

The distributional hypothesis is that words that ap­pear in similar contexts are semantically similar (Harris, 1951): the more two words have the same distribution, i.e. are found in the same syntactic con­texts, the more they are semantically close. We propose to apply this principle for metonymy resolution. Traditionally, the distributional approach groups words like USA, Britain, France, Germany because there are in the same syntactical contexts:

(1) Someone live in Germany.

(2) Someone works in Germany.

(3) Germany declares something.

(4) Germany signs something.

The metonymy resolution task implies to distin­guish the literal cases, (1) & (2), from the meto-nymic ones, (3) & (4). Our method establishes these distinctions using the syntactic context distribution. We group contexts occurring with the same words: the syntactic contexts live in and work in are occur­ring with Germany, France, country, city, place, when syntactic contexts subject-of-declare and sub-ject-of-sign are occurring with Germany, France, someone, government, president. For each Named Entity annotation, the hybrid method consists in using symbolic annotation if there is (§1.2), else using distributional annotation (§1.3) as presented below.

Method: We constructed a distributional space with the 100M-word BNC. We prepared the corpus by lemmatizing and then parsing with the same robust parser than for the symbolic approach (XIP, see sec­tion 3.1). It allows us to identify triple instances. Each triple have the form w1.R.w2 where w1 and w2 are lexical units and R is a syntactic relation (Lin, 1998; Kilgarriff & al. 2004). Our approach can be distinguished from classical distributional approach by different points. First, we use triple occurrences to build a distribu­tional space (one triple implies two contexts and two lexical units), but we use the transpose of the classical space: each point xi of this space is a syn­tactical context (with the form R.w.), each dimen­sion j is a lexical units, and each value xi(j) is the frequency of corresponding triple occurrences. Second, our lexical units are words but also complex nominal groups or verbal groups. Third, contexts can be simple contexts or composed contexts. We illustrate these three points on the phrase pro­vide Albania with food aid. The XIP parser gives the following triples where for example, food aid is considered as a lexical unit:

OBJ-N('VERB:provide','NOUN: Albania'). PREP_WITH('VERB: provide ','NOUN:aid'). PREP_WITH('VERB: provide ','NP:food aid').

From these triples, we create the following lexical units and contexts (in the context 1.VERB: provide. OBJ-N, "1" mean that the verb provide is the gov­ernor of the relation OBJ-N): Words: Contexts:

VERB:provide 1.VERB: provide. OBJ-N NOUN:Albania 1.VERB: provide. PREPWITH NOUN:aid 2.NOUN: Albania.OBJ-N NP:food aid 2.NOUN: aid. PREPWITH 2.NP: food aid. PREP_WITH 1.VERB:provide.OBJ-N+2.NOUN:aid. PREPWITH 1.VERB:provide.OBJ-N+2.NP:food aid. PREPWITH 1.VERB:provide.PREP_WITH +2.NO:Albania.OBJ-N

We use a heuristic to control the high productivity of these lexical units and contexts. Each lexical unit and each context should appear more than 100 times in the corpus. From the 100M-word BNC we ob­tained 60,849 lexical units and 140,634 contexts. Then, our distributional space has 140,634 units and 60,849 dimensions.

Using the global space to compute distances be­tween each context is too consuming and would induce artificial ambiguity (Jacquet, Venant, 2005). If any named entity can be used in a metonymic reading, in a given corpus each named entity has not the same distribution of metonymic readings. The country Vietnam is more frequently used as an event than France or Germany, so, knowing that a context is employed with Vietnam allow to reduce the meto-nymic ambiguity.

For this, we construct a singular sub-space de­pending to the context and to the lexical unit (the ambiguous named entity):

For a given couple context i + lexical unit j we construct a subspace as follows:

Subcontexts = list of contexts which are occur­ring with the word i. If there are more than k con­texts, we take only the k more frequents.

Sub_dimension = list of lexical units which are occurring with at least one of the contexts from the

Sub_contexts list. If there are more than n words, we take only the n more frequents (relative fre­quency) with the Sub_contexts list (for this applica­tion, k = 100 and n = 1,000).

We reduce dimensions of this sub-space to 10 dimensions with a PCA (Principal Components Analysis).

In this new reduced space (k*10), we compute the closest context of the context j with the Euclid­ian distance.

At this point, we use the results of the symbolic approach described before as starting point. We at­tribute to each context of the Sub_contexts list, the annotation, if there is, attributed by symbolic rules. Each kind of annotation (literal, place-for-people, place-for-event, etc) is attributed a score corre­sponding to the sum of the scores obtained by each context annotated with this category. The score of a context i decreases in inverse proportion to its dis­tance from the context j: score(context i) = 1/d(context i, context j) where d(ij) is the Euclidian distance between i and j.

We illustrate this process with the sentence pro­vide Albania with food aid. The unit Albania is found in 384 different contexts (|Sub_contexts| = 384) and 54,183 lexical units are occurring with at least one of the contexts from the Sub_contexts list (|Sub_dimension| = 54,183).

After reducing dimension with PCA, we obtain the context list below ordered by closeness with the given context (1.VERB:provide.OBJ-N):

Contexts d symb. annot.

1.VERB:supply.SUBJ-N_PRE 1.47 literal 1.VERB:support.SUBJ-N_PRE 1.70 place-for-people etc.

Score for each metonymic annotation of Albania:

place-for-event

The score obtained by each annotation type al­lows annotating this occurrence of Albania as a place-for-people metonymic reading. If we can't choose only one annotation (all score = 0 or equal­ity between two annotations) we do not annotate.