Getting BART to Ride the Idiomatic Train: Learning to Represent Idiomatic Expressions

Natural language has a common yet special class of multi-word expressions (MWEs) called idiomatic expressions (IEs) that exhibit semantic non-compositionality, where the meaning of the expression cannot be inferred from that of its constituent words (e.g., the idiom break a leg) (Baldwin and Kim, 2010 ). They are commonly used for specific communicative intents (Moon, 1998; Baldwin and Kim, 2010 ) and are individually rare but collectively frequent, appearing frequently across genres (Moon, 1998; Haagsma et al., 2020 ). They have been classically regarded as a “pain in the neck” to NLP systems (Sag et al., 2002 ) not only because of their non-compositionality, but also because of their contextual semantic ambiguity (used in idiomatic or literal meaning depending on the context). Challenges posed by the presence of IEs have been identified across multiple NLU tasks even with state-of-the-art (SOTA) solutions, including sentiment analysis (Liu et al., 2017; Biddle et al., 2020 ), paraphrase generation (Zhou et al., 2021 ), natural language inference (Chakrabarty et al., 2021 ), and dialog models (Jhamtani et al., 2021 ).

Even the the flagship NLP model GPT-3 (Brown et al., 2020 ) finds idioms challenging. We tested for its idiom comprehension over 75 idioms, covering a spectrum of the most to the least frequent idioms (based on their frequency of occurrence in the BNC (Haagsma et al., 2020 )). We do this in question-answering mode where we ask GPT-3¹ simple questions about idioms that can only be answered correctly if their meanings are understood in context (a sample of questions and answers are shown in Table 1 ) and noted 37 incorrect answers (50.7% accuracy), showing its inability to comprehend many idioms.

Despite the relative success of recent studies using pre-trained language models (PTLMs) for IE processing, including idiom token and type classification (Liu, 2019 ), IE span detection (Zeng and Bhat, 2021; Škvorc et al., 2022 ), and idiomatic to literal paraphrasing (Zhou et al., 2021 ), the emerging knowledge is that PTLMs are unable to produce meaningful representation of IEs in task-specific contexts (e.g., for IE sense disambiguation) (Tayyar Madabushi et al., 2021 ). Noting that solutions to improve IE representation in a task-agnostic manner are largely missing, this study aims to develop targeted solutions to make language models (LMs) idiom-aware with the immediate objective of improving IE representation in large PTLMs yet without relying on astronomically large training corpora and parameters. In this work, we focus on improving IE representation and seek answers to the following questions about IE embeddings.

Are the current PTLMs capable of generating semantically meaningful contextualized representations for IEs? Answering this question, we examine the IEs embeddings produced by a representative SOTA LM, for example, BART (Lewis et al., 2020 ). Specifically, we perform an intrinsic evaluation of IE embeddings by grouping them into semantic classes and observing how they cluster.

Observing the low quality of IE representation, we ask,

How can we expand the capability of these LMs to produce high quality IE embeddings? As a solution, we propose the Generation of Idiom Embedding with Adapter (GIEA) approach that extends the capabilities of the current SOTA LMs by producing quality IE embeddings.

Concretely, unlike prior work that treats each idiom as a new token (Hashempour and Villavicencio, 2020 ), GIEA refrains from new tokenization to represent IEs and uses an adapter (Houlsby et al., 2019; Pfeiffer et al., 2020a ) as a parameter-constrained learner of IE embeddings. Finally, we devise a denoising auto-encoder-style learning objective and train the network to reconstruct selective masked sentence parts. Our use of symbolic knowledge (Yu et al., 2021 ) of IEs to aid the learning of their embeddings results in the model needing a significantly small amount of data (∼60MB) compared to that required for LM pre-training (∼160GB of text for BART).

Our main contributions are as follows.

• (1) 

  We demonstrate the limited ability of SOTA PTLMs for generating semantically meaningful embeddings for IEs via a simple probing task.

• (2) 

  We propose a lightweight solution, GIEA, an adapter built over BART, to produce quality IE embeddings without altering input sentences.

• (3) 

  We evaluate the resulting IE embeddings using intrinsic and extrinsic methods to show that they are meaningful in the embedding space and are task-agnostic and generalizable across different idiom processing tasks (IE sense disambiguation and IE span detection). Compared to BART, GIEA gains 0.19 in homogeneity score (intrinsic evaluation), performs competitively on IE sense disambiguation, and gains 25% in sequence accuracy for IE span detection.

• (4) 

  We conduct detailed analyses on the performance and limitations of GIEA system to provide meaningful insights and future directions.²

Compositionality is a dominant paradigm driving the SOTA in NLP both at the tokenization and architectural levels. The tokenization of most LMs, for example, Byte-Pair Encoding (BPE) (Sennrich et al., 2016 ) and WordPiece (Wu et al., 2016 ), assumes compositionality not only at the phrase-level but also at the word level. This suggests that the meaning of a word is deduced from that of the subword components. At the architectural level, transformer-based LMs implicitly consider all phrases (or even words) as compositional. The self-attention mechanism in transformers considers the embedding of a word to be an attention weighted sum of the word embeddings in its context. This design leads to phrase or even sentence embeddings to be overall compositional. In addition, each IE is individually rare, compounding the difficulty for obtaining good IE representation. This leads us to hypothesize that the inherent notion of compositionality and the rarity of IEs are a hindrance to the representation of the IEs that are inherently non-compositional. We test the validity of this hypothesis by analyzing PTLMs’ representation of IEs.

We first obtain the embeddings for the IEs in the MAGPIE dataset (Haagsma et al., 2020 ), a collection of potentially idiomatic expressions (PIEs), that is, idioms used in a literal and idiomatic sense, and the sentences in which they occur. Focusing on the IEs used idiomatically (thus ensuring their non-compositionality), we first retrieve all the sentences in which they occur. Then, for each sentence, we extract the BART base embeddings corresponding to the IE tokens in the sentence. We then apply mean pooling across the tokens and across all the sentences in which the IE appears. In this manner we generate the embeddings for 1,480 idioms from an average of 22 sentences per idiom.

We then list IEs most similar to a set of IEs in the embedding space produced by the base BART model, computed using the cosine similarity. Table 2 shows examples of this listing including three most similar IEs (second column) to a sample of IEs (first column). As noted from the examples, IEs with superficial token-level (see red vs. see stars) and/or syntactic-level (quick as a flash vs. keen as mustard) matches tend to be most similar according to BART’s embeddings without accounting for their semantic congruence. This suggests that BART considers IEs mostly compositionally, an inadequate approach for representing the non-compositionality of the IEs.

To quantify the above qualitative finding, we manually assigned 129 idioms into 20 distinct meaning groups—“in summary”, “anger/upset”, “easy/relax”, “quick”, “exactly”, “death”, “punish/criticize”, “impress”, “happy”, “to understand”, “fail”, “success”, “close to”, “decline/worsen”, “grief/sad”, “confront/deal with”, “persevere”, “great effort”, “unimportant”, “careful”—averaging 6.4 idioms per group (see Table 3 for example groups and their idioms).

The idiom groups must satisfy the following two requirements: (1) Any two idioms from the same group must have a similar meaning though the idioms may not necessarily be interchangeable; and (2) any two idioms from different groups must not overlap in their meanings, that is, the boundaries between any groups should be clear. Moreover, we selected idioms that are idiomatically monosemous (excluding their literal interpretations) according to our dictionaries.³ To group the idioms, we first created a few candidate groups based on commonly occurring idiom meanings, such as “anger/upset” and “happy”. Then, for each idiom we either assigned it to an existing group or to a newly created meaning group. We only retained groups with more than three idioms and stopped the process once we had 20 groups. Using the aforementioned requirements, the validity of the groups and the idiom assignments were verified by two annotators, one with native and the other with near-native English abilities (one of whom was not associated with this study), using an idiom dictionary as needed. Only idiom assignments that were judged as correct by both the annotators were considered.

First, we generate the embeddings for these idioms based on their dictionary definitions using a pre-trained MPNet⁴ (Song et al., 2020 ) for sentence embeddings, referred to as definition embeddings. As a contrast, we generate their BART IE embeddings, referred to as BART embeddings, following the procedure discussed above. Then, we run agglomerative clustering⁵ to produce 20 clusters with complete linkage using the pairwise cosine similarity between the embeddings (definition and BART embeddings separately) as the distance metric. Finally, we measure the clustering quality using the homogeneity score as an index of the embedding quality, which is 1.0 if all the clusters contain only data points that are members of a single class. The homogeneity score for definition embeddings is 0.68, whereas the score for BART embedding is only 0.45. This suggests that BART embeddings are more scattered in the embedding space with less than half of the IEs from each cluster having the same meaning.

Toward producing higher quality IE embeddings by PTLMs, we propose GIEA; given a set of idiomatic sentences (i.e., sentences that each contains an IE), GIEA freezes the base PTLM and trains an adapter that specializes in IE representations. This is done by reconstructing idiomatic sentences that are corrupted with an idiom-aware noising function and meeting a dictionary definition-aided objective. GIEA’s overall framework is illustrated in Figure 1. In this work, we select BART as our base PTLM.

Following the pre-training for BART, our training has a text corruption stage with novel noising functions and a text reconstruction stage. In the text corruption stage, we introduce three noising functions such that one permits predicting masked IEs using the context words—the idiom-aware text infilling transformation—and the other two permit the model to use IEs to predict context words, namely, the copy and the span infilling transformation. In the idiom-aware text infilling transformation, given a sentence containing an IE, the entire IE is replaced with a single [MASK] token. During training, the model is asked to reconstruct the masked IE using the context words. Yet the masking of IEs alone is not sufficient for learning meaningful IE embeddings because the model sees IEs only in the decoder’s input but never in the input sentences, leaving the encoder’s adapter parameters unreachable by the reconstruction loss.

The two additional noising functions, the copy and the span infilling transformation, alleviate this shortcoming by allowing the model to learn to use IEs to infer the context words. In the copy transformation, for each sentence with its IE masked, we also supply its original, uncorrupted sentence as input and thus the model only has to copy the input sentence to the output. In the span infilling transformation, we mask a span of consecutive tokens excluding the IE tokens with a single [MASK], effectively asking the model to reconstruct the masked span using the IE and the remaining context. As in BART pre-training, span lengths are drawn from a Poisson distribution (λ = 3). However, our 0-length spans correspond to the original (input) sentence, identical to that of the copy transformation. Hence, the span infilling technically subsumes the copy transformation.

Ideally, we would like the model to use an IE to predict masked context words that are directly related to the meaning of the IE. For example, as shown in Figure 1, masking the sequence “so tired” helps the learning of the IE, “hit the sack”.

However, since the masked spans are randomly chosen, to guarantee that reconstructing the masked spans contributes to the IE meaning acquisition and inspired by prior success in prompting methods (Liu et al., 2021 ), we inject manually created templates for span infilling (e.g., When people say hit the sack, they mean that [MASK].) by connecting each IE to its dictionary definition as a sentence. We create four such templates per idiom with variations.⁶

During training, that is, the reconstruction stage, we randomly apply the idiom-aware text infilling transformation to 50% of sentences, while applying the copy or span infilling transformation to the remaining sentences in each epoch, and the model is asked to reconstruct the uncorrupted sentences. We experiment with and analyze the use of both the copy and span infilling in Section 5.

We leverage the dictionary definitions of IEs to aid the learning of semantically rich IE embeddings and supplement the small number of idiomatic sentences. To give an idea of the relative paucity of available idiomatic sentences, the number of idiomatic sentences in MAGPIE, the largest dataset for idiomatic sentences to date, is less than 30K, which is several orders of magnitude smaller than the BART pre-training corpus. Although collecting more sentences with IEs from other corpora is a way to directly enlarge the existing collection, isolating the truly idiomatic instances of potentially idiomatic expressions requires manual annotation, an exercise that we leave for future work.

Specifically, during training, we use MPNet to generate definition embeddings for each IE as before. MPNet is used because it empirically outperforms BART, as we will show in Section 5. We also generate IE embeddings by mean pooling the BART’s final layer output embeddings corresponding to the IE tokens. Note that these IE embeddings are generated from BART and the adapter being trained and thus correspond to a non-compositional representation. We then include the learning objective of increasing the cosine similarity between the IE embeddings and their corresponding definition embeddings. We refer to this learning objective as similarity forcing, which is intended to facilitate the learning of the IE embeddings by making the embedding space be more semantically meaningful, that is, locating IEs with similar meanings closer to each other.

The final loss during training is the weighted sum of the cross-entropy loss from reconstruction and the cosine similarity loss from similarity forcing. In our experiments, we set the two losses to be equally weighted and leave other weighting schemes for future explorations.

Instead of fine-tuning the full model on our new learning objective, we added an adapter with the Pfeiffer architecture (Pfeiffer et al., 2020a ) to the base BART model for conditional generation. This is so that during training, only the parameters of the adapter are trainable while those of the underlying language model are fixed, thus making our solution lightweight. Intuitively, because the added adapter is trained with the added objective of producing meaningful embeddings for non-compositional phrases (IEs), the adapter can be considered to be an expert in processing non-compositional language.

We use MAGPIE (Haagsma et al., 2020 ), a recent and the largest-to-date dataset of potentially idiomatic expressions in English, to train GIEA and evaluate the baseline models. We sample a subset of the dataset by selecting idioms with a single idiomatic meaning according to our IE dictionary (referencing Google dictionary and Wiktionary) and their corresponding sentences that are unambiguously labeled as being idiomatic (indicated by a perfect confidence score). The resulting collection has sentences drawn from a diverse set of genres from the British National Corpus (BNC) with 1,480 idioms with 32,693 sentences (77.4% idiomatic) in the train set and 1,001 idioms with 4,102 (77.57% idiomatic) sentences in the test set.

The overview of the intrinsic and extrinsic evaluation tasks are illustrated in Figure 2. The first task is an intrinsic evaluation of IE embeddings.

We follow the same procedure as described in Section 2 to perform clustering on the 20 distinct idiom groups with IE embeddings from the testing models. Note that we only use the sentences from the test set here to generate the IE embeddings. We use agglomerative clustering with complete linkage and pairwise embedding cosine similarity as the affinity metric.

The following two idiom-related tasks serve as extrinsic evaluations of the IE embeddings.

This is a common probing task used to probe if IE embeddings can differentiate the literal (compositional) from the idiomatic (non-compositional) uses of the IEs (Tayyar Madabushi et al., 2021; Adewumi et al., 2021 ). Many IEs can be used both figuratively or literally depending on the context. For example, the phrase “behind closed doors” can be interpreted literally as in The valuable items are locked behind closed doors and can be understood figuratively as in They avoided any publicity and made all deals behind closed doors. To account for this contextual ambiguity, these phrases are often refer to as potentially idiomatic expressions (PIEs) (Haagsma et al., 2020 ). The IE sense disambiguation task aims to classify each IE usage into idiomatic and literal class. To create a disambiguation classifier, we appended a single linear layer after the trained baseline embedding model. Given a sentence with a PIE and the location of the tokens belonging to the PIE, the baseline embedding model generates the embeddings for every token in the sentence. Then, the token embeddings corresponding to the PIE are mean pooled and fed to the linear layer to generate a binary classification. Only the linear layer is trainable when training the classifier. Given that nearly 78% percent of IEs are used figuratively in MAGPIE test data, the majority-class baseline predicts idiomatic label for all instances.

This is a more demanding task compared to IE sense disambiguation and studies focusing on this task are only emerging (Zeng and Bhat, 2021 ). Given a sentence with a PIE, a model is expected to classify every token as idiomatic or literal; when the PIE is used idiomatically, the tokens from the PIE will be tagged as idiomatic; when the PIE is used literally, all its tokens will be tagged as literal. To succeed in this task, a model must identify the presence of an IE and then precisely predict its boundary. To create such a classifier, we append a two-layer MLP that reduces the number of hidden neurons by a factor of 2 after each layer and uses ReLU activation between the layers. Only the MLP is trainable. Because the tokens are overwhelmingly literal, the majority-class baseline predicts each token to be literal.

Note that for both the above tasks, more powerful classifiers exist, as shown in prior works (Liu and Hwa, 2017, 2019; Zeng and Bhat, 2021; Škvorc et al., 2022 ). However, we deliberately constrain the complexity of the classifiers to linear layers (or MLPs) to ensure the performance differences reflect primarily the effect of different IE embeddings rather than that of additional modeling.

For intrinsic evaluation, that is, the embedding clustering task, we evaluate performance using homogeneity score to evaluate the clustering quality. Given that two idioms from different groups should have distinct meanings, we also measure the mean cosine distance between the embeddings for IEs from different groups; the larger the distance the better. For the IE sense disambiguation task, because it is a binary classification problem, we use accuracy and F1 score to evaluate the performance. For IE span detection, given that this is a sequence tagging task, we use three evaluation metrics, namely, sequence accuracy, token-level recall score, and token-level accuracy. In sequence accuracy, an instance is considered as correct if and only if all the tokens in the sequence are tagged correctly, making this the strictest metric. However, by only considering sequence accuracy, one may underestimate the performance of models that can tag most of the tokens from the positive (idiomatic) class correctly. Hence, we also consider the token-level recall and the accuracy score to complement the strict sequence accuracy metric. For token-level recall and accuracy, we compute the recall and accuracy for each predicted sequence and the final scores are averaged across all sequences.

Due to the lack of directly related prior work, we include only the majority-class baseline, BART, and variations of GIEA to demonstrate the effect of different components of our method detailed below.

is a naïve baseline that chooses the majority class for any classification problem.

is the original pre-trained BART-base model.

is the fine-tuned full pre-trained BART-base model using dictionary definition template sentences mentioned in Section 3 in addition to the MAGPIE train data with the idiom-aware text infilling and span infilling objective.

is a baseline that trains the adapter with only the idiom-aware text infilling transformation.

is a baseline that trains the adapter with both the idiom-aware text infilling and span infilling transformations.

is GIEA that trains the adapter with the idiom-aware text infilling transformation and similarity forcing learning objective.

We include two competing versions of GIEA using different noising functions:

is GIEA that trains the adapter with the similarity forcing objective and both the idiom-aware text infilling and copy transformations.

is GIEA that trains the adapter with the similarity forcing objective and both the idiom-aware text infilling and span infilling transformations.

For the adapters in all baseline models, our adapter implementation is based on Pfeiffer et al. (2020a ). The BART-base model is implemented and maintained by Huggingface (Wolf et al., 2020 ). The definition embeddings are generated by an MPNet hosted and maintained by the Sentence-Transformers package (Reimers and Gurevych, 2019 ). For the adapters, we trained all baseline GIEA models for 220 epochs with a batch size of 16. We trained a set of IE sense disambiguation and IE span detection classifiers for each baseline model except for the majority-class baseline. For IE sense disambiguation, we trained the classifier for 55 epochs with a batch size of 32 and for IE span detection, we trained it for 100 epochs with a batch size of 16. The linear layer and the MLP in the respective classifiers were trained with a dropout rate of 0.2. For all training, we used the Adam optimizer with a learning rate of 1e-5. For all models, checkpoints with the best validation performances were used in the experiments. All the other hyperparameters were in their default values. We only use MAGPIE’s idiomatic sentences to train GIEA and the baseline models, but we use both the idiomatic and the literal sentences to train the probing models for evaluation.

One of the defining characteristics of a good representation is that the embedding space should be semantically meaningful, that is, the embeddings of similar meaning IEs should be closer to each other in the embedding space via some distance metric (e.g., cosine similarity). As shown in Table 2, it is clear that after training with our ITI+SF+SI objective, the IE embeddings no longer cluster based on mere superficial similarities, instead, their meaning is the driving factor in determining their proximity in the embedding space. As shown in Table 4, the ITI+SF+SI method achieves the best homogeneity score and is significantly higher than the original BART embeddings by 0.19. Also, the mean cosine distance between the embeddings for the IEs from different meaning groups is merely 0.0379 for BART, indicating the BART embeddings are inadequate in discriminating between meanings; yet, the averaged distance is 0.2284 for ITI+SF+SI, which is very close to the distance of 0.2394 by the definition embeddings. To provide a more direct comparison, we also normalized the baseline performances using the BART embedding score as the lower bound and the definition embedding score as the upper bound. Comparing ITI+SF+SI and ITI+SF+Copy reveals that the more sophisticated SI noising function enabled the model to learn an embedding space that is semantically richer, as the normalized homogeneity score and cosine distance of ITI+SF+SI is higher than that of ITI+SF+Copy by 23.9 and 15.1.

Though commonly used by prior work, IE sense disambiguation is a relatively simple probing task in idiom processing. As shown in Table 5, though ITI+SI achieves the best performance numerically, all methods compared achieve competitive performances with respect to F1 and accuracy. This shows that BART embeddings already capture the idiosyncratic properties of IEs, in line with the findings from recent papers (Tayyar Madabushi et al., 2021; Adewumi et al., 2021 ). However, we believe that one cannot judge the quality of IE embeddings via this task alone, because IE senses can be distinguished correctly without the semantic knowledge of IEs. As evidence, under the same setting, we trained another disambiguation classifier with BART but replaced all the IEs from the sentences with single mask tokens for the classifier to make predictions based on just the embeddings of the mask tokens, thus removing all possible IE-related semantic information. We found that such a classifier still performs with an 86% accuracy, operating only on non-IE contextual information. So, IE comprehension ability and IE embedding quality cannot be fully assessed by probing the IE sense disambiguation ability, suggesting that the intrinsic embedding quality and performances on more difficult IE processing tasks must also be considered.

IE span detection is more difficult than IE sense disambiguation as it requires detecting the presence of IEs and precisely identifying their locations. The performance in this task showcases the superiority of our IE embedding methods. ITI+SF+Copy achieves the best performance that is 25.6 points higher than BART in sequence accuracy, our strictest metric. For token-level recall and token-level accuracy, ITI+SF+SI achieves the best performance with a 15-point gain in recall and 1.66 higher in accuracy than BART. The gain in token accuracy is small because the tokens are overwhelmingly literal; the majority-class baseline already achieves a 91% accuracy. The fact that ITI+SF+SI has better token-level performance than ITI+SF+Copy signifies that though ITI+SF+SI detects the span less precisely, it recovers the tokens from within IEs better than ITI+SF+Copy does.

We next examine the usefulness of the copy transformation and span infilling transformation in the noising function. Without copy and span infilling, the ITI+SF suffers in both intrinsic and extrinsic evaluation. For embedding clustering, the homogeneity score of ITI+SF is lower than ITI+SF+SI by 0.15 and lower than ITI+SF+Copy by 0.21, performing even slightly worse than the original BART’s embeddings. For IE span detection, ITI+SF’s sequence accuracy is lower than that of ITI+SF+SI and ITI+SF+Copy by 21.0% and 21.4%, respectively. Notably, without copy and span infilling transformation, ITI+SF performs barely better than BART, gaining only 4.2% in sequence accuracy. To a lesser degree, ITI+SI also demonstrates the usefulness of the span infilling transformation when compared with ITI, gaining 5.22% in sequence accuracy. Thus, copy or span infilling transformation is necessary and beneficial during the training of the embedding model.Moreover, even though ITI+SF+Copy and ITI+SF+SI performs competitively on the extrinsic evaluation tasks, ITI+SF+SI outperforms ITI+SF+Copy in the intrinsic evaluation task by a meaningful margin demonstrating ITI+SF+SI’s superiority over ITI+SF+Copy.

By comparing ITI+SF and ITI or ITI+SF+SI and ITI+SI, we examine the effect of similarity forcing. While ITI+SF performs similarly or even slightly worse than ITI on evaluation tasks, the performance gain of ITI+SF+SI over ITI+SI is noteworthy, for example, it gains 15.8% in sequence accuracy for IE span detection and 0.20 points in homogeneity score for embedding clustering. Considering the effect of copy and span infilling noising function, we see that ITI+SF+SI shows better performance than either ITI+SI or ITI+SF. This leads us to infer that similarity forcing is only useful when combined with the copy and span infilling transformation. In addition, we also compare the performance between ITI+SF+SI and BART-FT to demonstrate the usefulness of similarity forcing. BART-FT is a BART model fine-tuned on the same training data as ITI+SI. Though BART-FT has significantly more trainable parameters and the same access to external knowledge from the IE definition template sentences during training, BART-FT under-performs ITI+SF+SI by 14.48 points in sequence accuracy for span detection and 0.18 points in homogeneity score for embedding clustering. Therefore, we conclude that using similarity forcing in combination with copy- or span infilling transformation can boost the performance by a significant margin.

Though the MPNet’s definition embeddings and BART’s IE embeddings are in different spaces, we believe minimizing the cosine similarity between them to improve IE embeddings’ semantic meanings is a valid exercise because (1) the idiomatic meanings of IEs and the meaning of their component words are not related; hence relating their idiomatic meanings to the definition meanings from MPNet’s space will not affect the embeddings of the original words; and (2) prior research suggests that minimizing cosine similarity can even help relate the meanings between image embeddings and natural language embeddings (clearly not in the same embedding space) (Radford et al., 2021 ), hence the space difference between MPNet and BART should not present a problem. Moreover, using MPNet for the definition embedding results in an overall better empirical performance because MPNet produces higher-quality sentence embeddings than BART. We experimented training the ITI+SF+SI model but replaced the MPNet’s definition embeddings with that from BART. Comparing the results of the resulting model with those of ITI+SF+SI with MPNet embeddings, shown in the second row of Table 6, we see the resulting model achieves competitive performance for disambiguation but inferior performances in both span detection and embedding clustering with a sequence accuracy that is lower by 1.46% and a homogeneity score that is lower by 0.18. In fact, even the definition embedding, when generated by BART, only obtains a homogeneity score of 0.55 (not shown in tables) which is even lower than the ITI+SF+SI by around 0.10. This justifies our use of MPNet for definition embeddings.

In our case, encoder-decoder LMs (e.g., BART) are more suitable than an encoder-only LMs (e.g., BERT) because the decoder allows the use of the idiom-aware text infilling objective that asks the model to reconstruct the entire idiom from a single mask token. To empirically demonstrate the benefit, we trained an ITI+SF+SI model with BERT as the base LM and modified the idiom-aware text infilling objective by using one mask token per idiom token. As shown in the third row of Table 6, the BERT-based model under-performs its BART-based counterpart in all evaluation tasks by large margins.

Here, we further examine the quality of the definition embeddings and ITI+SF+SI’s IE embeddings (named GIEA embeddings). We compute precision at k (P@k) score for each idiom from the 129 idioms in the 20 meaning groups as follows. Given the embedding for an IE, E, we first find the k = 3 closest IEs using pairwise cosine similarity and n, the number of k closest IEs that are from the same group as E; then, P@3 is computed as n/k. The mean score for definition embeddings is 0.64. Meanwhile, the mean score for GIEA embeddings is 0.52, that is, each IE has about half of the 3-closest IEs from the same group. We found a large disparity among the groups with respect to the mean score for each meaning group. While most groups have a mean score around 0.5, groups such as “anger/upset”, “quick”, and “success” have scores higher than 0.6, and those of others, such as “punish/criticize”, “decline/worsen”, “persevere” are lower than 0.2.

Also, we found that the per group P@3 scores of the definition embedding are positively correlated with those of GIEA embedding with a Pearson correlation coefficient of 0.76. Based on these observations, we infer that the difficulty of learning IE meanings depends on the specific meaning group and the quality of the definition embedding directly affects the learned GIEA embedding. Improving the definition embeddings through better sentence embedding methods (e.g., by training specifically on dictionary definitions) may further improve the performance of our method. We also leave the important aspect assessing the quality of original compositional embeddings after learning IE embeddings to a follow-up study.

Here, we analyze the error of the best performing ITI+SF+SI model on the tasks of span detection and disambiguation. For span detection, we sampled 300 incorrect instances with imperfect sequence accuracies (30.5% of all incorrect samples) and categorized them into the six error types defined in Zeng and Bhat (2021 ). Among the sampled errors, we found that 3.7% were attributable to identifying one of the IEs when multiple IEs are present, 57% to detecting only a portion of the idiom span, 1% to identifying figurative expressions other than the ground truth idiom, 25% to identifying a PIE as idiomatic when actually used in the literal sense, 8.3% for failing to recognize the presence of an idiom, and another 5% for returning random tokens that are not meaningful nor part of any PIEs, that is, over 60% of the errors were in the detection of figurative tokens. In fact, over 40.8% of test idioms had their spans precisely tagged in all of their test instances. For disambiguation, over 82.8% of the test PIEs were classified with 100% accuracy and only less than 6% of the test PIEs had an accuracy less than 50%. For both disambiguation and span detection, the per-idiom accuracies were weakly correlated with the number of training instances per idiom (Pearson correlation coefficient of −3.84e-4 for disambiguation and 0.26 for span detection), suggesting that the performance discrepancy among idiom types is caused by factors other than their frequency in the train set. Future studies should consider the characteristics of the hard-to-learn idioms to improve the embeddings of the under-performing idioms.

An obvious limitation of GIEA is that it cannot generalize its representation ability to idioms unseen during training. From the results in Section 2 and Section 5, it is evident that the meanings of IEs cannot be learned from general corpora alone (even when there is a collection of sentences with IEs), rather, external knowledge (e.g., IE definitions) is a fundamental to providing the strong supervising signal (i.e., similarity forcing loss) needed for training. Taking this into consideration, we believe that it is impractical to generalize the representation ability to the unseen idioms because (1) intuitively, each IE has a unique origin, metaphorical linkage, and interpretation, so, the meaning of IEs have to be learned on a case-by-case basis; and (2) from our error analysis, even with the same training data and objective, the learning difficulty is highly idiom dependent, a point that is also corroborated by Nedumpozhimana et al. (2022 ). Therefore, we do not currently see a practical way to generalize GIEA to idioms that are unseen. However, we argue that this does not hinder the utility of GIEA, since our training data, MAGPIE, already contains idiomatic sentences for idioms (and metaphors) that occur in sources such as the Oxford Dictionary of English Idioms (Ayto and Press, 2009 ) and Wiktionary. Thus, we expect GIEA to cover most frequently used idioms. Besides, even though expanding an IE lexicon to include new idioms may be easy, gathering idiomatic sentences for those new idioms requires human input. So, an important future study is to consider methods that generalize GIEA to idioms with known identities but with limited or no idiomatic sentences.

Classically, two main idiom-related processing tasks, namely, idiom type classification and idiom token classification, have been studied (Cook et al., 2008; Liu and Hwa, 2019; Liu, 2019 ). Idiom type classification aims to decide if a set of MWEs can be used as IEs without considering additional context (?Fazly and Stevenson, 2006; Tabossi et al., 2008, 2009; Shutova et al., 2010; Reddy et al., 2011; Cordeiro et al., 2016 ). Idiom token classification determines if a given PIE is used in a literal or figurative sense in a sentence and solutions include those that mostly assume the knowledge of the location and/or identify of the PIEs (Fazly et al., 2009; Feldman and Peng, 2013; Peng and Feldman, 2016; Salton et al., 2016; Taslimipoor et al., 2018; Peng et al., 2014; Liu and Hwa, 2019 ), build per-idiom classifiers (Liu and Hwa, 2017 ), extract embeddings based on PIE positions (Liu and Hwa, 2019 ), or focus on only PIEs with specific syntactic structures (Taslimipoor et al., 2018 ). Due to the impracticality of acquiring this prior knowledge in real-world applications, most recent works (Zeng and Bhat, 2021; Škvorc et al., 2022 ) study the idiomatic expression identification problem, jointly the detecting and localizing a PIE without requiring PIE identity or position. This problem is related to the MWE identification task in STREUSLE (Schneider and Smith, 2015 ) but with a focus on expressions with semantic idiomaticity. In-line with prior state-of-the-art, we use the IE token classification and IE identification, dubbed as IE sense disambiguation and IE span detection, as the extrinsic evaluation tasks to our IE embeddings.

Since Sag et al.’s (2002 ) study on the impact of MWE, not only have studies identified the influence of IEs across various NLP applications (Salton et al., 2014; Fadaee et al., 2018; Ganitkevitch et al., 2013; Liu et al., 2017; Biddle et al., 2020 ), recent efforts have also sought ways to mitigate them (Jhamtani et al., 2021; Chakrabarty et al., 2021 ). However, the techniques used either simply enlarge the training data by including idiomatic sentences or paraphrase idiomatic sentences into equivalent literal sentences, completely ignoring the fundamental issue of IE representation. Other works (Tayyar Madabushi et al., 2021 ) have probed how idiomaticity is handled in PTLMs but offer no solution to improve their representation. Efforts to improve IE span detection or IE sense disambiguation include transforming the original representations from pre-trained LMs by incorporating static word embeddings alone (Liu and Hwa, 2017 ), with additional syntactic information (Zeng and Bhat, 2021 ), utilizing contrastive loss to make literal and figurative speech embeddings more distinctive (Lin et al., 2021 ), treating IEs as new tokens during training (Hashempour and Villavicencio, 2020 ), or combining representations from multiple pre-trained LMs (Škvorc et al., 2022 ). Taking a different approach in this work, instead of creating task-specific representations or altering tokenization at the input, we first train an LM that produces better IE embeddings in general and then show their benefit in the idiom processing tasks. In principle, our trained GIEA can be plugged into the prior works for idiom processing tasks, replacing their embedding models and improving their performances, an aspect we leave to future explorations.

Originally developed for computer vision applications (Rebuffi et al., 2017, 2018), adapters are new modules of simple projection layers added between the trained transformer layers, used in NLP as a parameter-efficient and fast fine-tuning method to adapt pre-trained LMs to new tasks or domains (Houlsby et al., 2019; Bapna and Firat, 2019 ). Recently, adapters have shown effectiveness in multi-task and multi-lingual transfer learning as well (Pfeiffer et al., 2020b; Ansell et al., 2021 ). In this work, we utilize an adapter as a lightweight non-compositional language expert that is trained on idiomatic sentences and thus can expand upon the base LM to generate semantically meaningful IE embeddings. The compact Pfeiffer adapter architecture (Pfeiffer et al., 2020a ) is used in GIEA.

The core idea for works on non-compositional phrase embeddings is to avoid treating phrases as purely compositional (by aggregating word embeddings) or non-compositional (treating phrases as single units), but consider both aspects. The approaches have adaptive weights and consider different compositions within a phrase (Li et al., 2018a; Hashimoto and Tsuruoka, 2016; Li et al., 2018b ) or utilize hypernymy information and represent phrases in special embedding spaces (Jana et al., 2019 ). Although related, these embedding methods cannot produce the contextualized phrase embeddings as transformer-based models do, nor can they be combined with PTLMs to aid downstream tasks.

The evaluation of word and phrase embeddings (Hashimoto and Tsuruoka, 2016; Jana et al., 2019 ) is typically via intrinsic methods (e.g., similarity and analogy) and extrinsic methods, e.g., downstream NLP tasks (Schnabel et al., 2015; Ghannay et al., 2016; Hupkes and Zuidema, 2018; Wang et al., 2019 ). A popular alternative evaluation method is probing, where a simple diagnostic classifier is trained to extract information from frozen embeddings and determine the extent to which desired linguistic properties are encoded in the representations (Adi et al., 2016; Warstadt et al., 2019; Alt et al., 2020; Ravichander et al., 2021 ). Our intrinsic and extrinsic evaluation of embeddings follow these prior works.

In this work, we first demonstrate current BART’s inability produce semantically meaningful representations for idioms, then, we propose GIEA, that uses a lightweight adapter, a set of denoising auto-encoder-style learning objectives, and a similarity forcing objective to produce quality IE embeddings without altering the input tokenization. Through both intrinsic evaluation of embedding quality and extrinsic evaluation on their usefulness on idiom-processing tasks, we find that GIEA greatly improves upon embedding quality and usefulness compared to the original pre-trained BART’s embeddings.

Future work should explore means to improve embedding quality for hard-to-learn idioms based on observed performance, IEs other than idioms (e.g., phrasal verbs), and the use of GIEA with other SOTA idiom processing models. Lastly, applying idiom-aware PTLMs to downstream applications that require the IE comprehension, such as dialog modeling and machine translation, would be fruitful pursuits.