VerbNet Representations: Subevent Semantics for Transfer Verbs

Susan Windisch Brown Julia Bonn James Gung University of Colorado University of Colorado University of Colorado [email protected] [email protected] [email protected]

Annie Zaenen James Pustejovsky Martha Palmer Stanford University Brandeis University University of Colorado [email protected]@[email protected]

Abstract verb senses and introduce a means for automati- cally translating text to these representations. We This paper announces the release of a new ver- explore the format of these representations and the sion of the English VerbNet types of information they track by thoroughly ex- with substantially revised semantic represen- amining the representations for transfer of posses- tations designed to facilitate computer plan- sions and information. These event types are ex- ning and reasoning based on human language. We use the transfer of possession and trans- cellent examples of complex events with multi- fer of information event representations to il- ple participants and relations between them that lustrate both the general framework of the change across the time frame of the event. By representations and the types of nuances the aligning our new representations more closely new representations can capture. These repre- with the dynamic event structure encapsulated by sentations use a Generative Lexicon-inspired the Generative Lexicon, we can provide a more subevent structure to track attributes of event precise subevent structure that makes the changes participants across time, highlighting opposi- tions and temporal and causal relations among over time explicit (Pustejovsky, 1995; Pustejovsky the subevents. et al., 2016). Who has what when and who knows what when are exactly the sorts of things that we 1 Introduction want to extract from text, but this extraction is dif- ficult without explicit, computationally-tractable Many natural language processing tasks have seen representations. These event types also make up a rapid advancement in recent years using deep substantial portion of VerbNet: 37 classes of verbs learning methods; however, those tasks that re- deal with change of possession and transfer of in- quire precise tracking of event sequences and par- formation out of VerbNet’s 300+ classes, covering ticipants across a discourse still perform better 810 verbs. using explicit representations of the meanings of each sentence or utterance. To be most useful 2 Background for automatic language understanding and genera- tion, such representations need to be both automat- The VerbNet (Kipper et al., ically derivable from text and reasonably format- 2006) is a hierarchical, wide-coverage verb lexi- ted for computer analysis and planning systems. con that groups verbs into classes based on sim- For applications like robotics or interactions with ilarities in their syntactic and semantic behavior avatars, commonsense inferences needed to under- (Schuler, 2005). Each class in VerbNet includes stand human language directions or interactions a set of member verbs, the thematic roles used are often not derivable directly from the utterance. in the predicate-argument structure of these mem- Tracking intrinsic and extrinsic states of entities, bers (Bonial et al., 2011), and the class-specific such as their existence, location or functionality, selectional preferences for those roles. The class currently requires explicit statements with precise also provides a set of typical syntactic patterns and temporal sequencing. corresponding semantic representations. A verb In this paper, we describe new semantic rep- can be a member of multiple classes; for exam- resentations for the lexical resource VerbNet that ple, run is a member of 8 VerbNet classes, in- provide this sort of information for thousands of cluding the run-51.3.2 class (he ran to the store)

154 Proceedings of the First International Workshop on Designing Meaning Representations, pages 154–163 Florence, Italy, August 1st, 2019 c 2019 Association for Computational Linguistics and the function-105.2.1 class (the car isn’t run- these caused-motion frames. ning). These memberships usually correspond to Each frame also includes a semantic represen- coarse-grained senses of the verb. The resource tation that uses basic predicates to show the rela- was originally based on Levin’s (1993) analysis of tionships between the thematic role arguments and English verbs but has since been expanded to in- to track any changes over the time course of the clude dozens of additional classes and hundreds of event. Thematic roles that appear in the ”syntax” additional verbs and verb senses. should always appear somewhere in the semantic VerbNet representations previously formed the representation. Overall, this linking in each frame basis for Parameterized Action Representation of the syntactic pattern to a semantic representa- (PAR) providing a conceptual representation of tion is a unique feature of VerbNet that emphasizes different types of actions (Badler et al., 1999). the close interplay of syntax and semantics. These actions involve changes of state, changes of location, and exertion of force and can be used 2.2 Revision of the Semantic Representations to animate human avatars in a virtual 3D environ- VerbNet’s old representations included an event ment (R. Bindiganavale and Palmer, 2000). They variable E as an argument to the predicates. are particularly well suited for motion and con- Representations of states were indicated with tact verb classes, providing an abstract, language- either a bare E, as for the own-100 class: independent representation (Kipper and Palmer, has possession(E, Pivot, Theme), or During(E), 2000). The more precise temporal sequencing de- as for the contiguous location-47.8 class (Italy scribed here is even more suitable as a foundation borders France): contact(During(E), Theme, Co- for natural language instructions and human-robot Theme). Most classes having to do with change, or human-avatar interactions. such as changes in location, changes in state and changes in possession, used a path rel predicate 2.1 VerbNet in combination with Start(E), During(E), and End/Result(E) to show the transition from one VerbNet has long been used in NLP for seman- location or state to another (1). tic role labeling and other inference-enabling tasks (Shi and Mihalcea, 2005; Giuglea and Moschitti, 2006; Loper et al., 2007; Bos, 2008). In addi- (1) The rabbit hopped across the lawn. tion, automatic disambiguation of a verb’s Verb- Theme V Trajectory Net class has been used as a stand-in for verb sense motion(during(E), Theme) disambiguation (Abend et al., 2008; Brown et al., path rel(start(E), Theme, ?Initial location1, 2014; Croce et al., 2012; Kawahara and Palmer, CH OF LOC, prep) 2014). path rel(during(E), Theme, Trajectory, VerbNet’s semantic representations use a CH OF LOC, prep) Davidsonian first-order-logic formulation that path rel(end(E), Theme, ?Destination, incorporates the thematic roles of the class. Each CH OF LOC, prep) frame in a class is labeled with a flat syntactic pattern (e.g., NP V NP). The ”syntax” that follows Efforts to use VerbNet’s semantic representa- shows how the thematic roles for that class appear tions (Zaenen et al., 2008; Narayan-Chen et al., in that pattern (e.g., Agent V Patient), much 2017), however, indicated a need for greater con- like the argument role constructions of Goldberg sistency and expressiveness. We have addressed (2006). A previous revision of the VerbNet se- consistency on several fronts. First, all neces- mantic representations made the correspondence sary participants are accounted for in the repre- of these patterns to constructions more explicit sentations, whether they are instantiated in the by using a common predicate (i.e., path rel) for syntax, incorporated in the verb itself (e.g., to all caused-motion construction frames(Hwang, drill), or simply logically necessary (e.g., all en- 2014). At the request of some users, we are sub- tities that change location begin in an initial loca- stituting more specific predicates for the general tion, whether it is commonly mentioned or not). path rel predicate, such as has location, has state 1A question mark in front of a thematic role indicates a and has possession, although the subevent pat- role that appears in the syntax in some frames for the class terns continue to show the commonality across but not in this particular frame.

155 Second, similar event types are represented with troduced a new event variable, e¨, to distinguish a similar format; for example, all states are rep- a process from other types of subevents, such as resented with E, never with During(E). Finally, states. For example, see the motion predicate in predicates are given formal definitions that apply (2). across classes. (2) The rabbit hopped across the lawn. In order to clarify what is happening at each Theme V Trajectory stage of an event, we turned to the Generative has location(e , Theme, ?Initial Location) Lexicon (Pustejovsky, 1995) for an explicit the- 1 motion(e¨ , Theme, Trajectory) ory of subevent structure. Classic GL character- 2 ¬has location(e , Theme, ?Initial location) izes the different Aktionsarten in terms of struc- 2 has location(e , Theme, ?Destination) tured subevents, with states represented with a 3 simple e, processes as a sequence of states char- acterizing values of some attribute, e1...en, and Although people infer that an entity is no longer transitions describing the opposition inherent in at its initial location once motion has begun, com- achievements and accomplishments. In subse- puters need explicit mention of this fact to accu- quent work within GL, event structure has been rately track the location of an entity. Similarly, integrated with dynamic semantic models in order some states hold throughout an event, while others to more explicitly represent the attribute modified do not. Our new representations make these dis- in the course of the event (the location of the mov- tinctions clear, where pre-event, while-event, and ing entity, the extent of a created or destroyed en- post-event conditions are distinguished formally in tity, etc.) as a sequence of states related to time the representation. points or intervals. This Dynamic Event Model Elsewhere (Brown et al., 2018), we discuss in (Pustejovsky and Moszkowicz, 2011; Pustejovsky, more detail the Dynamic Event Model, show the 2013) explicitly labels the transitions that move an effect of the new subevent structure on the inter- event from frame to frame. pretation of the role of the Agent, and give fur- Applying the Dynamic Event Model to Verb- ther examples of the new change of location and Net semantic representations allowed us refine the change of state representations. event sequences by expanding the previous tripar- tite division of Start(E), During(E), and End(E) to 3 Change of Possession an indefinite number of subevents. These num- In this section, we closely examine the represen- bered subevents allow very precise tracking of par- tations for events involving changes in posses- ticipants across time and a nuanced representation sion. These representations illustrate the greater of causation and action sequencing within a sin- clarity and flexibility we have gained by adopting gle event. In the general case, e1 occurs before e2, the conventions described in section 2. They also which occurs before e3, and so on. We’ve intro- show some of the choices we have made to capture duced predicates that indicate temporal and causal the underlying semantics while maintaining a con- relations between the subevents, such as cause(ei, nection to the varying surface forms. We discuss ej) and co-temporal(ei, ej). both one-way transfers (give) and two-way trans- We have made other refinements suggested by fers (sell). We also address the different perspec- the GL Dynamic Event Model. For example, tives verbs can impose on a transfer event, such as we greatly expanded the use of negated predi- the difference between Mary gave John the book cates to make explicit the opposition occurring and John obtained the book from Mary, in which in events involving change: e.g., John died is the Agent of the event is the Source or the Recipi- analyzed as the opposition halive(e1,Patient), ¬ ent of the item, respectively. These variations have alive(e2,Patient)i. Compare the new representa- interesting analogs in the Transfer of Information tion for changes of location in (2) to (1) above. In classes (Fig. 1), which we discuss in Section 4. (2), we use the opposition between has location The semantic representations for changes of and ¬has location to make clear that once the possession in VerbNet assume a literal, non- Theme is motion (in e2), it is no longer in the metaphoric use of the verbs in question. Metaphor Initial location. In order to distinguish the event may select only some of the source domain’s par- type associated with a semantic predicate, we in- ticipants or entailments. For example, She stole

156 for one-way transfers in which one party was re- sponsible for initiating the entire change, but was insufficient when more than one transfer occurred. There was no way to show that one party could initiate one transfer while another party initiated another. Two-way transfer representations either attributed all causation to one party, or omitted the cause predicate entirely. The ability to omit the predicate led one to wonder why it was ever nec- essary to include it. 3.2 New Representations Three predicates form the core of the change of possession representations:

Figure 1: Primary distinctions made in the VerbNet • has possession(e, [slot-1], [slot-2]) representations for events involving transfer • transfer(e, [slot-1], [slot-2], [slot-3])

• cause(ei, ej) John’s car entails that John no longer has pos- We define has possession broadly as involving session of his car, whereas She stole John’s heart ownership or control over a thing; e.g., I have a does not entail his loss of a vital organ. An analy- pencil can mean either you own a pencil or you sis of VerbNet classes in terms of their application (possibly temporarily) have use of a pencil. Within to figurative language (Brown and Palmer, 2012) the predicate, slot-1 is reserved for the posses- showed that some classes concern only metaphoric sor and can take thematic roles Source, Recipient, uses of their member verbs (e.g., calibratible cos- Goal, Agent, and Co-Agent. Slot-2 is reserved for 45.6.1), with semantic representations that directly the possession, and can take roles Theme and As- represent the figurative meaning without reference set. to the source domain. Many classes, however, Transfer is now a causative predicate, describ- were shown to refer to literal uses of the verbs, ing an event in which possession of a thing trans- although it was suggested that transformations or fers from one possessor to another. All three par- re-interpretations of the semantic representations ticipants are given as arguments. Slot-1 is reserved could be possible. for the possessor who initiates the transfer, and can 3.1 Previous Representations take thematic roles Agent and Co-Agent. Slot-2 is reserved for the possession (Theme or Asset), and The previous model allowed only three temporal slot-3 is reserved for the other possessor (Source, subevent periods: Start, During, and End. For Recipient/Goal, Agent, and Co-Agent). both Change of Possession and Transfer of In- The order of arguments within this predicate of- formation classes, each possession received one ten aligns with the temporal order of possession, path rel for the Start period and one for the End but this is incidental. Sometimes, an Agent who is period, allowing one clear owner per period. For initiating a transfer is the recipient of that trans- Change of Possession, it was reasonable to as- fer; in these cases, the Agent will still occupy slot- sume that possession transferred fully during the 1, even though they end up with possession last. It event, and as such, information about who did is also possible for an Agent to occupy slot-3 if an- not possess a thing at any point could have been other party (Co-Agent) is initiating the transfer. inferred through a rule. This model was suffi- The subevent numbering of the has possession cient for Change of Possession classes in and of predicates before and after the transfer provide a themselves, but failed to capture any contrast with full description of the temporal order of posses- Transfer of Information classes, for which this as- sion. sumption does not hold. The new basic representation is shown in (3). The cause predicate included arguments for an Agent or Causer (no other thematic roles were al- (3) has possession(e1, Source, Theme) lowed), and the overall event E. This was sufficient ¬has possession(e1, Recipient, Theme)

157 transfer(e2, Source, Theme, Recipient) a thing initiates taking that thing from the origi- has possession(e3, Recipient, Theme) nal possessor: berry-13.7, deprive-10.6.2, obtain- ¬has possession(e3, Source, Theme) 13.5.2, rob-10.6.4, and steal-10.5. The example cause(e2, e3) from steal-10.5 in (5) shows how Agent replaces This representation contains an initial state Recipient. subevent, a transfer subevent, and a resulting state (5) They stole the painting from the museum subevent. Cause(e2, e3) tells us that the trans- Agent V Theme Source fer triggers the resulting state. The opposing has possession(e1, Source, Theme) ¬has possession predicates show without a doubt ¬has possession(e1, Agent, Theme) that the Source stops having possession as soon as transfer(e2, Agent, Theme, Source) the transfer occurs, and the Recipient does not take has possession(e3, Agent, Theme) possession until then. This allows for clear auto- ¬has possession(e3, Source, Theme) matic tracking of an entity’s ownership status and cause(e2, e3) provides an important contrast with the new Trans- fer of Information representations. It will also al- Four main classes and two additional subclasses low coverage of cases of shared ownership of pos- belonging to classes listed above demonstrate sessions, if VerbNet expands in that direction. two-way transfers: exchange-13.6.1, get-13.5.1, invest-13.5.4, and pay-68, as well as give-13.1- 3.3 Change of Possession Variations 1 and obtain-13.5.2-1. In the following example Agents as Sources or Recipients: Depending from exchange-13.6.1, note the new handling of on the class, an Agent may function as Source subevents, cause, and the argument structure of or Recipient. In the old representations, some transfer. In e2, the Agent initiates the transfer of classes ended up including as core roles both an the Theme, and in e3, the Co-Agent initiates the Agent and a Recipient, or an Agent and a Source, transfer of the Co-Theme. Subevent e2 causes the even if those roles always overlapped in the syn- resulting possession states of the Theme, and e3 tax. This was likely due to pressure to include causes the resulting possession states of the Co- in the class thematic roles that were projected by Theme. the main predicates, path rel, cause and trans- fer. In the new model, we let Agent stand in for (6) Gwen exchanged the dress for a shirt whichever role it overlaps throughout the repre- Agent V Theme Co-Theme sentation. This eliminates the need for the equals has possession(e1, Agent, Theme) predicate, and has allowed us to eliminate syntac- ¬has possession(e1, ?Co-Agent, Theme) tically redundant roles from the class role invento- has possession(e1, ?Co-Agent, Co-Theme) ries. ¬has possession(e1, Agent, Co-Theme) Six classes demonstrate one-way transfers in transfer(e2, Agent, Theme, ?Co-Agent) which the entity who starts with possession transfer(e3, ?Co-Agent, Co-Theme, Agent) initiates giving that possession away: cheat- has possession(e4, ?Co-Agent, Theme) 10.6.1, contribute-13.2, equip-13.4.2, fulfilling- ¬has possession(e4, Agent, Theme) 13.4.1, future having-13.3, and give-13.1-1. In has possession(e5, Agent, Co-Theme) example (4) from fulfilling-13.4.1, Agent replaces ¬has possession(e5, ?Co-Agent, Co- Source throughout. Theme) (4) Brown presented a plaque to Jones cause(e2, e4) Agent V Theme Recipient cause(e3, e5) has possession(e1, Agent, Theme) Substitute-13.6.2 used to be included in this ¬has possession(e1, Recipient, Theme) group, but since it was specifically split off from transfer(e2, Agent, Theme, Recipient) exchange-13.6.1 to deal with a two-way exchange has possession(e3, Recipient, Theme) of location (i.e., two entities change places with ¬has possession(e3, Agent, Theme) each other), we are now treating it purely as a cause(e2, e3) Change of Location class rather than Change of Five classes demonstrate one-way transfers in Possession. When compared with (6), example (7) which an entity who does not have possession of from substitute-13.6.2 highlights the distinctions

158 we are able to achieve using the new Change of guarantee that another party doesn’t already pos- Location vs. Change of Possession treatments. sess it too.

(7) One bell ringer swapped places with another 4.2 New Representations Theme V Location Co-Theme Two new predicates describe Transfer of Informa- has location(e1, Theme, Location I) tion: has location(e , Co-Theme, Location J) 2 • has information(e, [slot-1], [slot-2]) motion(e¨3, Theme, Trajectory) ¬has location(e3, Theme, Location I) • transfer info(e, [slot-1], [slot-2], [slot-3]) motion(e¨ , Co-Theme, Trajectory) 4 These mirror the predicates used in Change of ¬has location(e , Co-Theme, Location J) 4 Possession in terms of their argument slots and has location(e , Theme, Location J) 5 functions, excepting that slot-2 may take Theme has location(e , Co-Theme, Location I) 6 or Topic but not Asset. Topic is used most com- cause(e¨ , e ) 3 5 monly for verbal information, while Theme is re- cause(e¨ , e ) 4 6 served for non-verbal information, which often re- flects assent or emotional states. The basic representation in (8) differs from Additional predicates: Several subgroups within Change of Possession in terms of the boundaries Change of Possession use additional predicates to on possession before and after the transfer info depict additional semantics. Future-having-13.3 subevent. Here, by leaving the Recipient’s pos- and berry-13.7 both take an irrealis(e) predicate session status underspecified in e1, we make no to show that the transfer and resulting states are claims about whether or not the Recipient already intended, but not guaranteed to have taken place knew the information at the beginning of the event. yet. Irrealis’s single argument is a subevent num- By marking the Source’s possession status with a ber, and one predicate is given per qualifying big E, we assert that the Source maintains posses- subevent. Another additional predicate is used sion of the information throughout the event, even in the get-13.5.1, give-13.1-1, obtain-13.5.2, and after the transfer info communication subevent. pay-68 classes. These all involve two-way trans- fers in which a Theme is exchanged for an Asset, (8) has information(E, Source, Topic) where the Asset is the cost of the Theme, repre- transfer info(e1, Source, Topic, Recipient) sented as cost(Theme, Asset). Finally, rob-10.6.4 has information(e2, Recipient, Topic) and steal-10.5 both involve an Agent/Recipient cause(e1, e2) who initiates taking a possession in an illegal man- ner. The representations include a manner(e, Ille- 4.3 Transfer of Information Variations gal, Agent) predicate which, for this usage, takes One-way transfers: Just as with Change of Illegal as a constant. Possession, Transfer of Information classes may involve an Agentive Source or Agentive Recip- 4 Transfer of Information ient. The basic representations for these types alternate from the basic Transfer of Information 4.1 Previous Representations representation in the same way demonstrated In the old model, the only consistent difference above, with Agent replacing either Source or between Transfer of Information and Change of Recipient throughout. The vast majority of Trans- Possession in terms of predicates and represen- fer of Information classes are of the Agentive tation structure lay within path rel, which con- Source type, including advise-37.9, complain- tained a constant called either TR OF INFO or 37.8, confess-37.10, crane-40.3.2, curtsey-40.3.3, CH OF POSS, respectively. Like Change of Pos- initiate communication-37.4.2, inquire-37.1.2, session, only one path rel was provided per tem- instr communication-37.4.1, interrogate-37.1.3, poral period, allowing only one clear possessor per lecture-37.11, manner speaking-37.3, nonver- period. Unfortunately, this failed to capture the bal expression-40.2, overstate-37.12, promise- important distinctions that knowledge is generally 37.13, say-37.7, tell-37.2, transfer mesg-37.1.1, not lost when communicated, and one party’s pos- and wink-40.3.1. Just one class, learn-14, features session and communication of knowledge is no an Agentive Recipient.

159 Two-way transfers: The two-way Transfer of voluntary bodily motion named by the verb, in- Information classes, chit chat-37.6 and talk-37.5, cluding classes crane-40.3.2, curtsey-40.3.3, and differ from the two-way Change of Possession wink-40.3.1. In addition to the basic trans- classes in several ways. Most notably, they are fer info predicates, these classes take a Patient not limited to a single transfer in each direction; role that is shown to be a body part of the Agent instead, a sequence of transfers repeats back and with a part of(Patient, Agent) predicate. Dur- forth between the two participants an unspeci- ing the course of the transfer info subevent, the fied number of times. The subevent ordering is Agent moves the Patient into a verb-specific po- changed so that the state resulting from one trans- sition, represented using has position(e, Patient, fer info occurs before the next transfer info be- V Position) and body motion(e,¨ Agent). These gins. The repeated turn-taking is expressed using classes have a more nuanced take on the posses- the repeated sequence predicate, which may take sion boundaries than the basic representation in as many subevent arguments as necessary to cap- (8). In example (10) from wink-40.3.1, the Theme ture the full span of the repeated behavior. The is a nonverbal emotional state conveyed through a example in (9) is from chit chat-37.6. bodily motion. We can generally assume that the Recipient does not have prior access to this type of (9) Susan chitchatted with Rachel about the information, and we make this explicit in e1. problem Agent V Co-Agent Topic (10) Linda nodded her agreement has information(E, Agent, Topic I) Agent V Theme has information(E, Co-Agent, Topic J) has information(E, Agent, Theme) ¬has information(e , ?Recipient, Theme) transfer info(e1, Agent, Topic I, Co-Agent) 1 ¬has position(e , ?Patient , V Position) has information(e2, Co-Agent, Topic I) 1 transfer info(e¨ , Agent, Theme, ?Recipi- transfer info(ee3, Co-Agent, Topic J, 2 Agent) ent) body motion(e¨ , Agent) has information(e4, Agent, Topic J) 2 has position(e , ?Patient , V Position) cause(e1, e2) 2 has information(e , ?Recipient, Theme) cause(e3, e4) 3 part of(?Patient , Agent) repeated sequence(e1, e2, e3, e4) cause(e¨2, e3) Additional predicates and selectional restric- tions: Several subgroups within Transfer of In- The second group involves potentially invol- formation capture further semantic details us- untary nonverbal expressions of an internal state, ing either additional predicates or specialized and includes classes animal sounds-38 and non- selectional restrictions on class roles. Two verbal expression-40.2 (11). As part of this re- classes feature verbs of asking: inquire-37.1.2 and lease, we have added a new Stimulus thematic interrogate-37.1.3. These classes take a Topic role role to these classes. The previous release in- with a selectional restriction [+question], which cluded frames for constructions using a Recipient, helps clarify that the communication event taking like Paul laughed at Mary and The dog barked place regards the question and never the response. at the cat, but didn’t cover possible construc- Manner speaking-37.3 and nonverbal expression- tions like Paul laughed at Mary to his friends or 40.2 both feature verbs that describe the manner The dog whimpered to its owner about the rab- of communication. The representations use an- bit in the yard. Adding Stimulus and its usual other manner predicate, this time with a verb- predicate in reaction to(e, Stimulus) to these rep- specific role V Manner in place of a constant. resentations aligns them with the other Stimu- Instr communication-37.4 features verbs that de- lus/Experiencer classes and expands the range scribe an instrument used to communicate (e.g., of frames they cover. These classes reflect the phone), and uses utilize(e, Agent, V Instrument) same assumptions about boundaries on possession to convey this. shown in (10). Two subgroups use Theme with selectional restriction [+nonverbal information]. The first (11) The dog whimpered to its owner at the sight group involves communication via some sort of of the rabbit in the yard

160 Agent V Recipient Stimulus lect among multiple valid frames, we select the has information(E, Agent, ?Theme) frame with highest total number of roles among ¬has information(e1, Recipient, ?Theme) the VN frames with the fewest unmapped roles. transfer info(e2, Agent, ?Theme, Recipi- This approach to VN using multiple in- ent) dependent systems represents a simple baseline manner(e2, Agent, V Manner) approach. We leave a more sophisticated, unified in reaction to(e2, Stimulus) approach to VN semantic parsing to future work. has information(e3, Recipient, ?Theme) cause(e2, e3) 6 Conclusion

The fine-grained semantic representations pre- 5 Automatic VerbNet Parsing sented here improve the consistency and precision of VerbNet’s verb semantics, offering a more use- To facilitate immediate use of the new VerbNet se- ful modeling for the subevent structure of partic- mantic representations, we are releasing a seman- ular event types. This should improve VerbNet’s tic parser that predicts the updated semantic rep- utility for human-robot and human-avatar interac- resentations from events in natural language input tion, and lend enhanced richness to applications sentences. For a given predicative verb in a sen- aimed at temporal event sequencing. tence, we define VerbNet semantic parsing as the All of the resources described in this paper task of identifying the VN class, associated the- are freely available. An online, browsable matic roles, and corresponding semantic represen- version of all the semantic representations tations linked to a frame within the class. is available through the Unified Verb In- We approach VerbNet semantic parsing in three dex at https://uvi.colorado.edu/uvi search. distinct steps: 1. Sense disambiguation to iden- A downloadable version can be accessed at tify the appropriate VN class, 2. PropBank se- https://uvi.colorado.edu/nlp applications. A mantic role labeling (Gildea and Jurafsky, 2002; demo of the VerbNet Parser is at http://verbnet- Palmer et al., 2005) to identify and classify argu- semantic-parser.appspot.com/. ments, and 3. Alignment of PropBank semantic roles with VN thematic roles within a frame be- Acknowledgments longing to the predicted VN class. After align- ing arguments from the PropBank SRL system’s We gratefully acknowledge the support of DTRAl output with the thematic roles in a particular VN -16-1-0002/Project 1553695, eTASC - Empirical frame, the frame’s associated semantic predicates Evidence for a Theoretical Approach to Seman- can be instantiated using the aligned arguments. tic Components and DARPA 15-18-CwC-FP-032 Communicating with Computers, C3 Cognitively For sense disambiguation, we use a supervised Coherent Human-Computer Communication (sub verb sense classifier trained on updated VN class from UIUC) and Elementary Composable Ideas tags (Palmer et al., 2017). For semantic role (ECI) Repository (sub from SIFT). Any opinions, labeling, we use a variation of the system de- findings, and conclusions or recommendations ex- scribed in He et al.(2017) and Peters et al. pressed in this material are those of the authors (2018) using solely ELMo embeddings (without and do not necessarily reflect the views of DTRA any pre-trained or fine-tuned -specific vec- or the U.S. government. tors) trained on a combination of three PropBank annotated corpora described in (O’Gorman et al., In addition, we thank our anonymous reviewers 2019): OntoNotes (Hovy et al., 2006), the En- for their constructive feedback. glish Web (Bies et al., 2012), and the BOLT corpus (Garland et al., 2012). For align- References ment, we begin by applying updated SemLink mappings (Palmer, 2009) to map PropBank roles Omri Abend, Roi Reichart, and Ari Rappoport. 2008. to linked VN thematic roles for the identified VN A supervised algorithm for verb disambiguation into classes. In Proceedings of the 22nd Inter- class. Remaining arguments are then mapped us- national Conference on Computational Linguistics- ing heuristics based on the syntactic and selec- Volume 1, pages 9–16. Association for Computa- tional restrictions defined in the VN class. To se- tional Linguistics.

161 Norm Badler, Martha Palmer, and Rama Bindiganavle. Luheng He, Kenton Lee, Mike Lewis, and Luke Zettle- 1999. Animation control for real-time virtual hu- moyer. 2017. Deep : What mans. Communications of the ACM, 42(7):65–73. Works and What’s Next. Proceedings of the 55th Annual Meeting of the Association for Computa- Ann Bies, Justin Mott, Colin Warner, and Seth Kulick. tional Linguistics (Volume 1: Long Papers). 2012. English web treebank. Linguistic Data Con- sortium, Philadelphia, PA. Eduard Hovy, Mitchell Marcus, Martha Palmer, Lance Ramshaw, and Ralph Weischedel. 2006. OntoNotes: Claire Bonial, Susan Windisch Brown, William Cor- the 90% solution. In Proceedings of the human lan- vey, Martha Palmer, Volha Petukhova, and Harry guage technology conference of the NAACL, Com- Bunt. 2011. An exploratory comparison of thematic panion Volume: Short Papers, pages 57–60. Associ- roles in verbnet and lirics. In Workshop on Interop- ation for Computational Linguistics. erable Semantic Annotation, page 39. Jena D Hwang. 2014. Identification and representation Johan Bos. 2008. Wide-coverage semantic analysis of caused motion constructions. with boxer. In Proceedings of the 2008 Conference on Semantics in , pages 277–286. Association for Computational Linguistics. Daisuke Kawahara and Martha Palmer. 2014. Sin- gle classifier approach for verb sense disambigua- Susan Windisch Brown, Dmitriy Dligach, and Martha tion based on generalized features. In LREC, pages Palmer. 2014. Verbnet class assignment as a wsd 4210–4213. task. In Computing Meaning, pages 203–216. Springer. Karin Kipper, Anna Korhonen, Neville Ryant, and Martha Palmer. 2006. Extending verbnet with novel Susan Windisch Brown and Martha Palmer. 2012. Se- verb classes. In Proceedings of LREC, volume 2006, mantic annotation of metaphorical verbs with verb- page 1. Citeseer. net: A case study of climband poison. In Workshop on interoperable semantic annotation, page 72. Karin Kipper and Martha Palmer. 2000. Representa- tion of actions as an interlingua. In Proceedings of Susan Windisch Brown, James Pustejovsky, Annie Za- NAACL-ANLP Workshop on Applied Interlinguas: enen, and Martha Palmer. 2018. Integrating gener- practical applications of interlingual approaches to ative lexicon event structures into verbnet. In Pro- NLP,, pages 12–17. Association for Computational ceedings of the Eleventh International Conference Linguistics. on Language Resources and Evaluation (LREC- 2018). Beth Levin. 1993. English Verb Classes and Alterna- tions: A Preliminary Investigation. University of Danilo Croce, Roberto Basili, Alessandro Moschitti, Chicago Press. and Martha Palmer. 2012. Verb classification us- ing distributional similarity in syntactic and seman- Edward Loper, Szu-Ting Yi, and Martha Palmer. 2007. tic structures. In Proceedings of the 50th Annual Combining lexical resources: mapping between Meeting of the Association for Computational Lin- and verbnet. In Proceedings of the 7th In- guistics: Long Papers-Volume 1, pages 263–272. ternational Workshop on Computational Linguistics, Association for Computational Linguistics. Tilburg, the Netherlands.

Jennifer Garland, Stephanie Strassel, Safa Ismael, Anjali Narayan-Chen, Colin Graber, Mayukh Das, Zhiyi Song, and Haejoong Lee. 2012. Linguistic Md Rakibul Islam, Soham Dan, Sriraam Natarajan, resources for genre-independent language technolo- Janardhan Rao Doppa, Julia Hockenmaier, Martha Workshop gies: user-generated content in bolt. In Palmer, and Dan Roth. 2017. Towards problem Programme , page 34. solving agents that communicate and learn. In Proceedings of the First Workshop on Language Daniel Gildea and Daniel Jurafsky. 2002. Automatic Grounding for Robotics labeling of semantic roles. Computational Linguis- , pages 95–103. tics, 28(3):245–288. Tim O’Gorman, Sameer Pradhan, James Gung, and Ana-Maria Giuglea and Alessandro Moschitti. 2006. Martha Palmer. 2019. The unified propbank land- Semantic role labeling via , verbnet and scape. Ms. propbank. In Proceedings of the 21st International Conference on Computational Linguistics and the Martha Palmer. 2009. Semlink: Linking propbank, 44th annual meeting of the Association for Compu- verbnet and framenet. In Proceedings of the gen- tational Linguistics, pages 929–936. Association for erative lexicon conference, pages 9–15. Pisa Italy. Computational Linguistics. Martha Palmer, Daniel Gildea, and Paul Kingsbury. Adele E Goldberg. 2006. Constructions at work: The 2005. The proposition bank: An annotated cor- nature of generalization in language. Oxford Uni- pus of semantic roles. Computational linguistics, versity Press on Demand. 31(1):71–106.

162 Martha Palmer, James Gung, Claire Bonial, Jinho Choi, Orin Hargraves, Derek Palmer, and Kevin Stowe. 2017. The Pitfalls of Shortcuts: Tales from the word sense tagging trenches. Springer series Text, Speech and Language Technology, Essays in Lexical Semantics and Computational Lexicography - In Honor of Adam Kilgarriff. Matthew Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word rep- resentations. In Proceedings of the 2018 Confer- ence of the North American Chapter of the Associ- ation for Computational Linguistics: Human Lan- guage Technologies, Volume 1 (Long Papers), pages 2227–2237.

J. Pustejovsky. 1995. The Generative Lexicon. Brad- ford Book. Mit Press. James Pustejovsky. 2013. Dynamic event structure and habitat theory. In Proceedings of the 6th Interna- tional Conference on Generative Approaches to the Lexicon (GL2013), pages 1–10. ACL. James Pustejovsky and Jessica Moszkowicz. 2011. The qualitative spatial dynamics of motion. The Journal of Spatial Cognition and Computation. James Pustejovsky, Martha Palmer, Annie Zaenen, and Susan Brown. 2016. Verb meaning in context: In- tegrating verbnet and gl predicative structures. In Proceedings of the LREC 2016 Workshop: ISA-12, Potoroz, Slovenia, volume 2016. J. Allbeck N. Badler A. Joshi R. Bindiganavale, W. Schuler and M. Palmer. 2000. Dynami- cally, altering agent behaviors using natural lan- guage instructions. In Proceedings of the 4th International Conference on Autonomous Agents (AGENTS), pages 293–300.

Karin Kipper Schuler. 2005. Verbnet: A broad- coverage, comprehensive verb lexicon. Lei Shi and Rada Mihalcea. 2005. Putting pieces to- gether: Combining framenet, verbnet and for robust semantic parsing. In International con- ference on intelligent text processing and computa- tional linguistics, pages 100–111. Springer.

Annie Zaenen, Daniel G Bobrow, and Cleo Condo- ravdi. 2008. The encoding of lexical implications in verbnet predicates of change of locations. In LREC.

163