English. This paper presents the analysis of a mapping between two resources, IMAGACT and T-PAS, made through a rule-based algorithm which converts argument structures in thematic roles. Results are good in terms of Recall, while Precision values are low: an analysis of the causes is proposed.
The automatic mapping of information between two resources is not a trivial task, but indeed joining information over specific data can benefit the involved resources. This paper describes the analysis of a mapping between two linguistic resources: IMAGACT and T-PAS. The motivation behind this mapping starts with the observation that both resources deal with Italian verbs disambiguation, are corpus-based and contain pieces of information that can be integrated with each other.
IMAGACT is a linguistic ontology of actions, that are grouped in concepts and related to different verb Types. For example, the action “John takes the cup from the shelf” belongs to the concept “take an object” and refers to Type 3 of the verb to take. Each Type is also associated to one or more thematic structures (e.g. [AGENT-verbTHEME-SOURCE]) and to videos via a set of captions.
T-PAS is a repository of argument typed structures for Italian verbs. Each verb is listed with its structures, which correspond to different senses of the verb. For each structure, the specification of the expected semantic type in every argument position (e.g. for the subject) is provided.
In this paper, we describe the results of a first attempt of mapping information between these resources. Specifically, for each of the 248 verbs analysed in both resources, we aim at matching the IMAGACT Types with the corresponding typed argument structures in T-PAS. We operate this mapping by applying a set of rules which convert the information from the argument structure into a thematic-role combination, and find all the Types that match this combination.
The linking between argument and thematic
structures of a predicate is a debated complex task
in linguistic theories
The paper is structured as follows: in Section 2 we present the resources; in Section 3 we describe the mapping procedure; in Section 4 we present and discuss the results of the mapping, tested on a gold standard; in Section 5 we provide direction for future work; in Section 6 we report our conclusions. 2
In this section we describe IMAGACT and T-PAS.
Table 1 shows the total and shared quantitative
data of the two resources.
2.1 IMAGACT
IMAGACT1
Action concepts are derived by a deep
analysis of the most frequent action verbs in Italian and
English spoken corpora; this ensures the ontology
to cover the most relevant actions for our
everyday activities. Given that no one-to-one
correspondence can be established between an action
verb and an action concept
The resource is growing continuously: by now, it consists of a total of 1010 action concepts, each one with a visual representation (i.e. a scene), and 21 covered languages (9 fully-mapped, 13 underway), with an average of 730 action verbs per language. 2.2
T-PAS
T-PAS3, Typed Predicate Argument Structures
3http://tpas.fbk.eu/ typed structure (henceforth t-pas), the specification of the expected semantic type (ST) for each argument slot is provided. T-PAS accounts for the following argument positions: subject, object, indirect object, complement, adverbial and clausal. A description of the sense, in the form of an implicature, is also linked to the t-pas.
Example 1 reports the t-pas#2 of the verb abbattere: the STs [[Human]] and [[Event]] are specified for the subject position (as alternatives) and [[Building]] for the object position.
(1) [[Human Event]-subj] abbattere [[Building]-obj] implicature:[[Human Event]] distrugge, butta giu` [[Building]] example: “Il muratore abbatte la parete.” (Eng.“The bricklayer knocks the wall.”)
The STs aim at generalizing over the set of
lexical items observed in a certain position for a
particular sense of the verb. For instance, in Example
1, the ST [[Building]] generalizes over the
lexical item parete (Eng. wall). STs are drawn from
a list of about 230 types4 and are also organized
in a hierarchy, in which the elements are linked
by a “IS-A” relation
...
[[Artifact]] [[Machine]] [[Vehicle]] [[Plane]] [[Road Vehicle]] ..
Each t-pas corresponds to a distinct sense of the
verb and is identified and defined by analysing
instances of the verb in a corpus, following the
lexicographic procedure called Corpus Pattern
Analysis
4For details on the list creation see
6According to the CPA procedure, after analysing a random sample of 250 concordances of the verb in the corpus, each t-pas is defined by recognizing its relevant structure and identifying the STs for each argument slots.
T-PAS currently contains 1000 verbs. The
reference corpus is a reduced version of ItWAC
We aim at finding the best semantic match between a verb Type in IMAGACT and the t-pass of the same verb in T-PAS, the two referring to the same action concept. Notice that it is possible that a Type in IMAGACT is mapped to more than one t-pas due, for instance, to different possible verb alternations that can occur inside the same Type. Figure 1 shows an example of this mapping, in which there is a match between Type 1 and tpas#1 of the verb macinare.
The mapping is done as follows. By observing a sample of verbs in the resources, we first defined a set of simple rules to convert the t-pas in a thematic structure. Considering the ST in the argument positions of the t-pas (e.g. [Human]-subj, [Food]-obj]), the rules aim at creating a thematic structure for the t-pas of the kind AG-v-TH (dotted arrow in Figure 1). Then, we used an algorithm which applies these rules to all the t-pass of a verb, and map the derived thematic structure (derivedts) to the thematic structures (ts) of the Types in IMAGACT (horizontal arrow in Figure 1). The system thus compares all the ts in IMAGACT with all the derived-ts in T-PAS for the same verb, and retrieves the matches.7 In Figure 1, the t-pas#1 for the verb macinare have been transformed in the structure AG-v-TH and then mapped to the ts of the Type.
The mapping between IMAGACT and T-PAS is made for the 248 verbs common to the two resources.
7Notice that the mapping is considering just this information of the resources and does not consider e.g. captions in IMAGACT or examples in T-PAS.
Datasets The rules for the conversion of a tpas in a derived-ts have been manually created by observing a sample of 15 verbs shared by the two resources (devset). We evaluated the mapping against a gold standard manually created by pairing the Types of other 14 verbs with the corresponding t-pass. We extracted the 29 verbs from the 248 shared by the two resources. The selection was made preserving the variability of the verbs in the two resources, in terms of their number of Types or t-pas. For instance, prendere (to take) is associated with 17 t-pass in T-PAS and 18 Types in IMAGACT; on the contrary bussare (to knock) has only 2 t-pass and 1 Type.
Conversion rules Table 3 synthesizes the rules we adopted. The rules consider both the ST in the argument slot and the argument slot itself, and are meant to associate a ST in an argument slot to a thematic role. For example, line 7 of Table 3 has to be interpreted as follows: if for the subject position of the t-pas the ST [[Animate]] (or a IS-A [[Animate]], according to the hierarchy of ST) is expected, then the AGENT role is selected (line 8). The rules also consider if the verb is in reflexive form (line 13). Moreover, if the t-pas registers the ST [[Abstract Entity]] (or a ST that IS-A [[Abstract Entity]]) as unique ST for any argument position (i.e. it is the only ST expected for the position), the t-pas was excluded from the mapping, as IMAGACT only accounts for physical actions which do not involve abstract entities. 4
In order to calculate Precision (P) and Recall (R) of the algorithm, we considered that DESTINATION (DE), SOURCE (SO) and LOCATION (LO) roles can not always be discriminated (for example, room is a DE in “John puts a table in the room”, a SO in “John takes the table from the room”, a LO in “John walks in the room”). The same happens for AGENT (AG) and ACTOR (AC): a human can be an agent (“John sweeps the room”) or an actor (“John bumps his head”). These limits can not be exceeded by an improvement of the rule definitions, because they are strictly dependent on the verb semantics. When calculating P and R, we grouped these derived structures together.
Precision (P) Recall (R) F-measure (F1)
0.283 0.792 0.418
We observe good values for R, while the P is very low (Table 4). A deeper analysis shows that in 34.61% of the cases, we have a full match with the gold standard and in 38.46% the results from the mapping include the ones expected by the gold standard. This means that in many cases the system is able to retrieve the correct matches.
Figure 2 shows the distribution of the main thematic structures in the Types of the whole IMAGACT ontology (in orange), in the devset (in red), compared with the derived-ts from T-PAS (in green). We verified a posteriori that the distribution of tss in the devset is strictly comparable with the one in the whole ontology, meaning that the devset is also well-balanced in terms of the thematic structures coverage (see orange and red bars in Figure 2).
By using the transformational rules we were able to recreate all the structures that are used in IMAGACT; however, there are some discrepancies in the production of AG-v-TH, TH-v (too high) and AG-v-TH-[DEjLOjSO] (too low) (see Figure 2).
The critical issue is represented by the AG-vTH structure: this is the most frequent one among the IMAGACT Types and in our test set (112 over 166 Types). For example, the following sentences belong to 4 different Types of the verb stringere, but have the same ts AG-v-TH: “Marco stringe la mano a Luca”; “Marco stringe le gambe”; “Marco stringe i pugni”; “Marco stringe la vite”. This happens also for the t-pas of stringere: 3 over the 5 derived-ts are AG-v-TH, so the system produces 12 combinations over 3 attested in the gold standard. The high frequency of this structure strongly influences the final P and R results. Moreover, the ts AG-v-TH is not distinctive of Types intra-verbs: by taking all the verbs with more than one Type, and for which AG-v-TH is a possible ts, we measured that in only 38,22% of them this ts is present in only one Type; in the other verbs (61.78%) the AG-v-TH structure appears in more than one Type. 5
Given the result in terms of Precision we presented in the previous section, we are considering to adopt other strategies that can be useful for the mapping of IMAGACT and T-PAS.
For instance, it would be possible to exploit the
examples from the corpus associated with each
tpas in T-PAS. In this sense, we hypothesize the
processing of these examples through BabelFy
On the other way round, IMAGACT captions
could also be mapped into the corresponding
tpass, by using the output of the algorithm
developed in
This method added to our rule-based strategy can be particularly useful to solve the ambiguity related to the thematic pattern AG-v-TH, for which the use of lexical information would reduce the number of possible matches. 6
In this paper we presented a first attempt of mapping IMAGACT and T-PAS by using a rule-based algorithm for the automatic conversion of T-PAS semantic types into thematic structures. We took advantage of the strong discriminative power of semantic types in their argument position to reduce the possible set of allowed thematic structures. This approach has an intrinsic limit: thematic roles are determined by verb semantics and their difference is not always reflected in the related semantic type. We also found out that the ts AG-v-TH represents the most critical issue, being the most frequent structure, and appearing in more than one Type of the same verb.
The results report a good recall and a low precision, confirming that our algorithm is not able to produce an actual mapping between the two resources, but it provides a reliable set of mapping candidates: we believe that it can be fruitfully exploited for a first step of a mapping process, in order to filter a lot of unwanted matching possibilities. We are confident that by exploiting additional linguistic information from the two resources (e.g. captions and occurrences in IMAGACT, lexical information and examples in T-PAS), the precision of this mapping will improve sensibly.