Adverbs, Surprisingly

Proceedings of the The 12th Joint Conference on Lexical and Computational Semantics (*SEM 2023), pages 512–526 July 13-14, 2023 ©2023 Association for Computational Linguistics Adverbs, Surprisingly Dmitry Nikolaev1 Collin F. Baker2 Miriam R.L. Petruck3 Sebastian Padó1 1 IMS, University of Stuttgart, Germany 2 International Computer Science Institute, Berkeley, USA 3 FrameNet Corresponding Author: dnikolaev@fastmail.com Abstract This paper begins with the premise that adverbs are neglected in computational linguistics. This view derives from two analyses: a literature re- view and a novel adverb dataset to probe a state- of-the-art language model, thereby uncovering systematic gaps in accounts for adverb mean- ing. We suggest that using Frame Semantics for characterizing word meaning, as in FrameNet, provides a promising approach to adverb anal- ysis, given its ability to describe ambiguity, se- mantic roles, and null instantiation. 1 Introduction Adverbs are the part of speech (POS) that has seen the least attention in (computational) linguistics, likely due to its challenging nature (Conlon and Evens, 1992). As Huddleston and Pullum (2002, 563) state, “the adverb is a [...] residual category [...] to which words are assigned if they do not satisfy the more specific criteria for nouns, verbs, adjectives, prepositions, and conjunctions.” Syntactically, they modify many POSs, except nouns (eat porridge quickly, hardly noticeable), or even complete clauses (Probably, I’ll come to- morrow). They are semantically varied (Thoma- son and Stalnaker, 1973), ranging from intensi- fiers/modifiers (absolutely, beautifully) to temporal and spatial specifications (yesterday, forward), to so-called speaker-oriented adverbs yielding infer- ences about speaker attitudes, beliefs, and evalu- ations. Finally, adverbs can occupy different po- sitions in sentences, creating complex issues of scoping and ambiguity (Alexiadou, 2004; Payne et al., 2010). Consider the following sentences:1 (1) a. Happily, they watched TV until dinner. b. They happily watched TV until dinner. c. They watched TV happily until dinner. d. They watched TV until dinner happily. 1Huddleston and Pullum (2002, 575) While language users tend to interpret Ex. 1b–1d as describing the TV watchers’ mental state, Ex. 1a is ambiguous and can also be read as a positive evaluation of the situation by the speaker. In sum, adverbs provide crucial information not just about the where and how of events, but also about attitudes and evaluations. However, relatively little research on adverbs exists in computational linguistics, although lexical factors are generally recognized as central for many NLP tasks (Berger et al., 2000). Lexical information is generally rep- resented either in online dictionaries or by embed- dings extracted from corpora (Turney and Pantel, 2010; Devlin et al., 2019; Peters et al., 2018). As a dictionary, WordNet (Miller et al., 1990) lists ad- verbs but only provides a relatively impoverished account, while lexicons for sentiment analysis (Be- namara et al., 2007; Dragut and Fellbaum, 2014) and hedging detection (Jeon and Choe, 2009; Is- lam et al., 2020) only consider specific subtypes of adverbs as to how they modulate the intensity of adjectives. On the distributional side, adverbs have been considered from a derviational perspec- tive (Lazaridou et al., 2013); …

Content truncated to 500 words