Short-Term Meaning Shift: A Distributional Exploration

Abstract

We present the first exploration of meaning shift over short periods of time in online communities using distributional representations. We create a small annotated dataset and use it to assess the performance of a standard model for meaning shift detection on shortterm meaning shift. We find that the model has problems distinguishing meaning shift from referential phenomena, and propose a measure of contextual variability to remedy this.

Room to Glo: A Systematic Comparison of Semantic Change Detection Approaches with Word Embeddings

“…We experiment with three approaches: (1) Training continuously by initializing the embeddings for a given time-step t with the embeddings trained at the previous time-step t − 1 (e.g., Kim et al, 2014); (2) Training embeddings for each time-step independently and posthoc aligning them (e.g., Hamilton et al, 2016b;Kulkarni et al, 2015) using orthogonal Procrustes 4 We only report results using CBOW in this paper. We found similar trends when using the skip-gram model, which has been used in previous works on semantic change (e.g., Kim et al, 2014;Kulkarni et al, 2015;Hamilton et al, 2016b;Stewart et al, 2017;Tredici et al, 2018).…”

“…A major challenge in developing semantic change detection systems is obtaining ground truth data (Kutuzov et al, 2018), which has so far prevented a systematic evaluation of different approaches. Many studies rely on hand-picked examples (e.g., Wijaya and Yeniterzi, 2011;Rodda et al, 2017) or human judgements (e.g., Tredici et al, 2018). Some studies have performed evaluations based on dictionary data (e.g., Cook et al, 2014;Basile and McGillivray, 2018), manual annotation of dictionary senses in corpora (McGillivray et al, 2019), and manual annotation of word types (Kenter et al, 2015), but this approach is not well-suited for recent, yet-to-berecorded changes.…”

Distributional Semantics and Linguistic Theory

“…When the amount of data is scarce, spurious effects easily arise. For instance, Del Tredici et al (2019), in a study of meaning shift in a community of soccer fans with data from 2011 to 2017, find that reference to specific people or events causes changes in cosine similarity that do not correspond to semantic change; an example is stubborn, which in 2017 was mostly used when talking about a new coach. 1 Effects like this challenge the Distributional Hypothesis, as a change in context does not signal a change in meaning, and call for more nuanced methods.…”

“…Distributional methods started being used for semantic change around the 2010s, with initial works using classic distributional methods (Sagi et al 2009;Gulordava & Baroni 2011) and Kim et al (2014) introducing neural network representations, which have been predominant in later work (Hamilton et al 2016;Szymanski 2017;Del Tredici et al 2019). Distributional approaches are based on the hypothesis that a change in context of use mirrors a change in meaning, which can be seen as a special case of the Distributional Hypothesis.…”

Mining the UK Web Archive for Semantic Change Detection

“…Recently, word embeddings have become the common practice for constructing word representations in NLP (Mikolov et al, 2013). A typical process followed in the context of semantic change is to learn the representations of a word over different time intervals and then compute its shift, by employing some distance metric over the resulting representations (Kim et al, 2014;Hamilton et al, 2016;Del Tredici et al, 2018).…”