InVeRo: Making Semantic Role Labeling Accessible with Intelligible Verbs and Roles Simone Conia1, Fabrizio Brignone2, Davide Zanfardino2, and Roberto Navigli1 1Sapienza NLP Group, Department of Computer Science, Sapienza University of Rome 2Babelscape, Italy [email protected], [email protected] Abstract (Kipper-Schuler, 2005), or FrameNet (Baker et al., 1998). The linguistic intricacies of such resources Semantic Role Labeling (SRL) is deeply de- pendent on complex linguistic resources and may, however, dishearten and turn away new prac- sophisticated neural models, which makes the titioners. Regardless of which linguistic resource is task difficult to approach for non-experts. To used in the task, to further complicate the situation address this issue we present a new platform SRL has been usually divided into four subtasks named Intelligible Verbs and Roles (InVeRo). – predicate identification, predicate sense disam- This platform provides access to a new verb biguation, argument identification and argument resource, VerbAtlas, and a state-of-the-art pre- classification – but, to the best of our knowledge, trained implementation of a neural, span-based recent state-of-the-art systems do not address all architecture for SRL. Both the resource and the system provide human-readable verb sense these four subtasks simultaneously without relying and semantic role information, with an easy to on external systems (Swayamdipta et al., 2017; He use Web interface and RESTful APIs available et al., 2018; Strubell et al., 2018; He et al., 2019). at http://nlp.uniroma1.it/invero. Therefore, obtaining predicate sense and semantic role annotations necessitates the tedious orchestra- 1 Introduction tion of multiple automatic systems, which in its Since its introduction (Gildea and Jurafsky, 2002), turn further complicates the use of SRL in prac- Semantic Role Labeling (SRL) has been recog- tice and in semantics-first approaches to NLP more nized as a key task to enable Natural Language Un- generally. derstanding in that it aims at explicitly answering In this paper, we present InVeRo (Intelligibile the "Who did What to Whom, When and Where?" Verbs and Roles), an online platform designed to question by identifying and labeling the predicate- tackle the aforementioned issues and make Seman- argument structure of a sentence, namely, the actors tic Role Labeling accessible to a broad audience. that take part in the scenario outlined by a predicate. InVeRo brings together resources and tools to per- In fact, SRL has already proven to be useful in a form human-readable SRL, and it accomplishes wide range of downstream tasks, including Ques- this by using the intelligible verb senses and se- tion Answering (Shen and Lapata, 2007; He et al., mantic roles of a recently proposed resource named 2015), Information Extraction (Christensen et al., VerbAtlas (Di Fabio et al., 2019) and exploiting 2011), Situation Recognition (Yatskar et al., 2016), them to annotate sentences with high performance. Machine Translation (Marcheggiani et al., 2018), In more detail, the InVeRo platform includes: and Opinion Role Labeling (Zhang et al., 2019). Unfortunately, the integration of SRL knowl- • a Resource API to obtain linguistic informa- edge into downstream applications has often been tion about the verb senses and semantic roles hampered and slowed down by the intrinsic com- in VerbAtlas. plexity of the task itself (Navigli, 2018). Indeed, • a Model API to effortlessly annotate sentences SRL is strongly intertwined with elaborate linguis- using a state-of-the-art end-to-end pretrained tic theories, as identifying and labeling predicate- model for span-based SRL. argument relations requires well-defined predicate sense and semantic role inventories such as the • a Web interface where users can easily query popular PropBank (Palmer et al., 2005), VerbNet linguistic information and automatically an- 77 Proceedings of the 2020 EMNLP (Systems Demonstrations), pages 77–84 November 16-20, 2020. c 2020 Association for Computational Linguistics notate sentences on-the-go without having to VerbAtlas, whenever two predicate senses can bear write a single line of code. the same semantic role, the semantics of this role is coherent across the two predicate senses by defi- Notably, InVeRo also takes advantage of PropBank nition, resulting in readable labels for non-expert to get the best of both worlds, and provides an- users. VerbAtlas also clusters predicate senses into notations according to both resources, enabling so-called frames (COOK, DRINK, HIT, etc.) inspired comparability and fostering integration. by FrameNet (Baker et al., 1998), with the idea 2 The InVeRo Platform that senses sharing similar semantic behavior lie in the same frame. For non-expert users, this organi- The InVeRo platform aims at making SRL more zation has the added advantage of explicitly link- approachable to a wider audience, not only in or- ing predicate senses that are otherwise unrelated, der to promote advances in the area of SRL itself, like make.01 and create.01 in PropBank which, in- but also to encourage the integration of semantics stead, are part of the same frame MOUNT-ASSEMBLE- into other fields of NLP. The two main barriers to PRODUCE in VerbAtlas and, therefore, also share the this objective are the complexity of i) the linguistic same semantic roles. In a bid to make SRL more resources used in SRL which are, however, indis- accessible, the InVeRo platform adopts the intelli- pensable for the definition of the task itself, and ii) gible verb senses and semantic roles of VerbAtlas. the complexity of the recently proposed techniques. Section 2.1 explains how InVeRo takes advan- 2.2 An All-in-One Solution for SRL tage of the intelligible verb senses and semantic As already mentioned in Section1, the traditional roles of VerbAtlas to gently introduce non-expert SRL pipeline consists of four main steps: predicate users to SRL, while Section 2.2 details how the identification, predicate sense disambiguation, ar- InVeRo model for SRL can make semantic role gument identification and argument classification. annotations accessible to everyone. While some of the above steps are considered easier than others, each of them features distinct peculiar- 2.1 Intelligible Verb Senses and Roles ities, which has driven recent works to focus on One of the most contentious points of discussion in improving only specific aspects of the entire SRL SRL is how to formalize predicate-argument struc- pipeline. Instead, little attention has been paid to tures, that is, the semantic roles that actors can systems that can tackle all the above-mentioned play in a scenario defined by a predicate. Prop- steps at the same time. As a result, anyone wishing Bank (Palmer et al., 2005), one of the most popular to take advantage of SRL annotations in another predicate-argument structure inventories, uses an NLP task has to choose, mix and match multiple enumerative approach where each predicate sense automatic systems in order to obtain sentences fully has a possibly different roleset, e.g., for the predi- annotated with predicate sense and semantic role cate make, the sense make.01 (as in “making a prod- labels. Understandably, this has been a major de- uct”) bears the semantic roles ARG0 (creator), ARG1 terrent for the integration of semantics into down- (creation), ARG2 (created from) and ARG3 (benefi- stream applications. ciary), whereas make.02 (as in “cause to be”) bears As part of the InVeRo platform, not only do only ARG0 (impeller) and ARG1 (impelled). This we introduce an all-in-one model that addresses exhaustive approach, however, requires an expert the complete SRL pipeline with a single forward linguist to tell which roles share similar seman- pass, but we also make this model available through tics across senses (e.g., ARG0 is an agent in both a Web interface to let everyone label sentences make.01 and make.02) and which do not (e.g., ARG1 with SRL annotations without the need to install is a product in make.01 but a result in make.02). any software. In other words, a user only has to On the other hand, VerbAtlas (Di Fabio et al., provide a raw text sentence; the InVeRo all-in-one 2019), a recently proposed predicate-argument model for SRL takes care of the rest, making the structure inventory, in contrast to the enumerative predicate sense and role labeling process accessible approach of PropBank and the thousands of frame- and effortless. specific roles of FrameNet, adopts a small set of ex- plicit and intelligible semantic roles (AGENT, PROD- Model Design. The InVeRo all-in-one system for UCT, RESULT, DESTINATION, ::: , THEME) inspired by SRL is based on the ideas put forward by He et al. VerbNet (Kipper-Schuler, 2005). As a result, in (2018) in that, unlike other works that used word- 78 level BIO tagging schemes to label arguments (He and made available as prepackaged downloads, e.g. et al., 2017; Strubell et al., 2018; Tan et al., 2018), SENNA2, or as online demos, e.g., AllenNLP’s it directly models span-level features. In particular, SRL demo3. However, recent BERT-based online we follow He et al.(2018) by letting the neural systems, such as AllenNLP’s SRL demo, do not model learn span-level representations from the perform predicate sense disambiguation (in addi- word-level representations of the span start and tion to predicate identification, argument identi- span end words, while also adding a span-length fication and argument classification), which is a specific trainable embedding. More formally, the crucial step in SRL, especially when considering span representation sij from word i to word j is that the PropBank roles ARG0, ARG1, through ARG5 obtained as follows: become meaningful only if they are associated with a PropBank predicate sense (see Section 2.1). s w w l s sij = W (ei ⊕ ej ⊕ ej−i) + b Results.