Ingrid Falk

Combining formal concept analysis and translation to assign frames and semantic role sets to French verbs

In Natural Language Processing, verb classifications have been shown to be useful both theoretically (to capture syntactic and semantic generalisations about verbs) and practically (to support factorisation and the supervised learning of shallow semantic parsers). Acquiring such classifications manually is both costly and errror prone however. In this paper, we present a novel approach for automatically acquiring verb classifications. The approach uses FCA to build a concept lattice from existing linguistic resources; and stability and separation indices to extract from this lattice those concepts that most closely capture verb classes. The approach is evaluated on an established benchmark and shown to differ from previous approaches and in particular, from clustering approaches, in two main ways. First, it supports polysemy (because a verb may belong to several classes). Second, it naturally provides a syntactic and semantic characterisation of the verb classes produced (by creating concepts which systematically associate verbs with their syntactic and semantic attributes).

Federating clustering and cluster labelling capabilities with a single approach based on feature maximization: French verb classes identification with IGNGF neural clustering.

Classifications which group together verbs and a set of shared syntactic and semantic properties have proven to be useful in both linguistics and Natural Language Processing tasks. However, most existing approaches for automatically acquiring verb classes fail to associate the verb classes produced with an explicit characterisation of the syntactic and semantic properties shared by the class elements. We propose a novel approach to verb clustering which addresses this shortcoming and permits building verb classifications whose classes group together verbs, subcategorisation frames and thematic grids. Our approach involves the use of a recent neural clustering method called IGNGF (Incremental Growing Neural Gas with Feature maximization). The use of a standard distance measure for determining a winner is replaced in IGNGF by feature maximisation measure relying on the features of the data that are associated with clusters during learning. A main advantage of the method is that maximised features used by IGNGF during learning can also be exploited in a final step for accurately labelling the resulting clusters. In this paper, we exploit IGNGF for the unsupervised classification of French verbs and evaluate the obtained clusters (i.e., verb classes) in two different ways. The first way is a quantitative analysis of the clustering process relying on a usual gold standard and on complementary unbiased clustering quality indexes. The second way is a qualitative analysis of the cluster labelling process. Relying on an adapted gold standard, we evaluate the capacity of the IGNGF clusters labels (i.e., subcategorisation frames and thematic grids) to be exploited for bootstraping a VerbNet-like classification for French. Both analyses clearly highlight the advantages of the approach.

Towards a lexicon of event-selecting predicates for a French FactBank

This paper presents ongoing work for the construction of a French FactBank and a lexicon of French event-selecting predicates (ESPs), by applying the factuality detection algorithm introduced in (Saurí and Pustejovsky, 2012). This algorithm relies on a lexicon of ESPs, specifying how these predicates influence the factuality of their embedded events. For this pilot study, we focused on French factive and implicative verbs, and capitalised on a lexical resource for the English counterparts of these verbs provided by the CSLI Group (Nairn et al., 2006; Karttunen, 2012).

Automatic Identification of Aspectual Classes across Verbal Readings.

The automatic prediction of aspectual classes is very challenging for verbs whose aspectual value varies across readings, which are the rule rather than the exception. This paper sheds a new perspective on this problem by using a machine learning approach and a rich morpho-syntactic and semantic valency lexicon. In contrast to previous work, where the aspectual value of corpus clauses is determined on the basis of features retrieved from the corpus, we use features extracted from the lexicon, and aim to predict the aspectual value of verbal readings rather than verbs. Studying the performance of the classifiers on a set of manually annotated verbal readings, we found that our lexicon provided enough information to reliably predict the aspectual value of verbs across their readings. We additionally tested our predictions for unseen predicates through a task based evaluation, by using them in the automatic detection of temporal relation types in TempEval 2007 tasks for French. These experiments also confirmed the reliability of our aspectual predictions, even for unseen verbs.

Enhancing NLP Tasks by the Use of a Recent Neural Incremental Clustering Approach Based on Cluster Data Feature Maximization

The IGNGF (Incremental Growing Neural Gas with Feature maximisation) method is a recent neural clustering method in which the use of a standard distance measure for determining a winner is replaced in IGNGF by cluster feature maximization. One main advantage of this method as compared to concurrent methods is that the maximized features used during learning can also be exploited in a final step for accurately labeling the resulting clusters. In this paper, we apply this method to the unsupervised classification of French verbs. We evaluate the obtained clusters (i.e., verb classes) in three different ways. The first one relies on an usual gold standard, the second one on unsupervised cluster quality indexes and the last one on a qualitative analysis. Our experiment illustrates that, conversely to former approaches for automatically acquiring verb classes, IGNGF method permits to produce relevant verb classes and to accurately associate the said classes with an explicit characterisation of the syntactic and semantic properties shared by the classes elements.