For irregular polysemes (WIRE: cable; surveillance), whose meanings cannot be determined by a productive rule, we tested the shared-features model developed in Brocher et al. (2016). This model assumes a shared and an unshared portion in the semantic representation of an irregular polyseme. Using ambiguous words normed for similarity, frequency, and type of ambiguity (see Eddington & Tokowicz, 2015), we conducted a continuous priming lexical decision experiment and a sentence reading eye-tracking experiment. We tested irregular polysemes with one frequent and one infrequent meaning (biased) and irregular polysemes with about equally frequent meanings (balanced) and compared them to homonyms. In Expt1, individual words were presented until participants made a lexical decision. The ISI between prime and target words was 50 ms or 200 ms (between subjects). In experimental trials, an ambiguous word (biased/balanced homonym or biased/balanced polyseme) or a pronounceable nonword was followed by either a dominant-meaning or subordinate-meaning related target word (e.g., WIRE or GINDER – CABLE; WIRE or GINDER – POLICE). Facilitation was observed for all conditions at 50 ms ISI (t = 4.35), with no differences between dominant and subordinate meanings of irregular polysemes or homonyms. At 200 ms ISI, we found robust priming for dominant but not subordinate meanings of biased homonyms, t = 2.21, and for both meanings of balanced homonyms and balanced irregular polysemes, t = 3.96. There was no priming for either meaning of biased polysemes. These data suggest that both meanings of homonyms are accessed at a short ISI, while the dominant meaning is selectively accessed at a longer ISI (Simpson & Burgess, 1985). For irregular polysemes, the claims that their meanings overlap in representation and that activation of the shared portion precludes frequency effects were supported. Importantly, for balanced polysemes, activation of the unshared portions leads to between-meaning competition: readers need to make a choice as to which meaning to retrieve, just as is the case for balanced homonyms. For biased polysemes, activation of unshared portions decays over time, which explains the lack of priming with the longer ISI: it is the unshared portions that are linked to disambiguating target words. In Expt2, participants read single sentences with a biased or balanced homonym, a biased or balanced irregular polyseme, or a matched control word in the first clause (e.g., When Mr. Jordon discovered the wire (bomb) in the lamp, the FBI aborted the top secret mission). The second clause contained context information that supported the ambiguous word’s less frequent meaning. In the first clause, gaze durations were longer for balanced homonyms and balanced polysemes than for controls, t = 2.07. In the second clause, we found shorter gaze durations, regression path times, and total reading times for disambiguations of biased and balanced homonyms, ts > 2.17. There were no reliable differences between biased and balanced polysemes and controls. These data converge with the results from Expt1. Between-meaning competition during retrieval led to longer reading times for balanced words than controls. For biased words, readers retrieved the dominant meaning when meanings were unrelated (homonyms) and a shared representation when meanings overlapped (irregular polysemes). At subordinate-bias disambiguation, readers often needed to reanalyze their initial interpretation when they had encountered a homonym in the first clause (Duffy et al., 1988), but not when they had encountered a polyseme (Brocher et al., 2016). For balanced homonyms and balanced polysemes, readers took longer in selecting a meaning during retrieval. However, when encountering disambiguation, readers only needed to undergo a costly reanalysis in the case of balanced homonyms. In comparison, for balanced irregular polysemes, the availability of shared features helped readers switching from one interpretation to another in case they had fully accessed the unsupported meaning before. ——————————————————