1. Introduction
Text simplification (TS) mainly refers to the process of reducing linguistic complexity at both syntactic and lexical levels and improving readability and understandability (Devlin Reference Devlin1998; Shardlow Reference Shardlow2014; Alva-Manchego, Scarton, and Specia Reference Alva-Manchego, Scarton and Specia2020b). It is commonly used to increase the readability of documents intended for children (De Belder and Moens Reference De Belder and Moens2010), non-native speakers (Petersen and Ostendorf Reference Petersen and Ostendorf2007) and people with dyslexia. The requirements for simplified outcomes may vary among audiences (Xu, Callison-Burch, and Napoles Reference Xu, Callison-Burch and Napoles2015), for instance, depending on the characteristics of the dataset. The task of simplification has been approached through sentence-level simplification (Nishihara, Kajiwara, and Arase Reference Nishihara, Kajiwara and Arase2019; Martin et al. Reference Martin, de la Clergerie, Sagot and Bordes2020a) and paragraph-level simplification (Sun, Lin, and Wan Reference Sun, Lin and Wan2020; Devaraj et al. Reference Devaraj, Wallace, Marshall and Li2021). In this paper, we focus only on sentence-level simplification, as this is where the majority of work on control tokens exists.
In order to fit the requirements of different user groups, some projects introduced explicit discrete prompts as control tokens to assist TS models to learn specific properties from datasets and adjust the resultant simplifications accordingly (Martin et al. Reference Martin, de la Clergerie, Sagot and Bordes2020a; Agrawal, Xu, and Carpuat, Reference Agrawal, Xu and Carpuat2021). By adjusting the value in different control tokens, researchers can manually adjust the characteristics of the output, such as length, syntactic and lexical difficulty, etc.
The control tokens are added to the beginning of the complex sentences and represent a relationship between that sentence and the desired input (such as the desired compression ratio). In addition, the numerical value also changes with the demands of the outcome. The format of the control token is
$\lt$
Token_value
$\gt$
, where Token is a novel extra-vocabulary token with human interpretable meaning, and value is a numerical value indicating some relationship between the given input and output as shown in Fig. 1 and discussed Section 5.1. The design of the control tokens is based on the need for adjustment. Multiple control tokens can be applied simultaneously, and four control tokens are used in this project.
Figure 1. An instance of preprocessed input and raw output of the system.
Although the control tokens are manually crafted, how the control tokens change the outcome remains unstudied. To explore the mechanisms of control tokens in simplification, this paper proposes the following: (1) Reimplement ACCESS (Martin et al. Reference Martin, de la Clergerie, Sagot and Bordes2020a), the state-of-the-art (SOTA) at the time of writing in Section 4.3. (2) Verify the importance of control tokens in Section 5. (3) Explore the influence and interactions of the variation of control tokens in the format in Section 5.1. (4) Investigate the effects of the tokenization method in Section 5.2. (5) Propose both regressive and multi-classification methods on the prediction of the value of control tokens and compare the performance difference with existing methods in Sections 5.3, 5.4 and 5.5.Footnote a (6) Study example cases to evaluate the newly proposed methods in Section 6.2.
Among the above-mentioned content, we will focus on the following three research questions:
-
• Whether the quantisation of applying control tokens will significantly affect the performance;
-
• Whether the tokenization method of control tokens will significantly affect the performance;
-
• Whether the control token predictor can predict the ideal values and hence significantly improve the performance.
2. Literature review
2.1. General text simplification
Natural language generation (NLG) is a sub-task in natural language processing. There have been attempts to build an NLG system based on hand-crafted rules and to define the problem and features based on knowledge in the last century (Hovy Reference Hovy1990; Reiter and Dale Reference Reiter and Dale1997). With the development of computation power and the re-introduction of neural networks, more neural-network-based statistical methods were applied (Wen et al. Reference Wen, Gasic, Mrksic, Su, Vandyke and Young2015; Dušek and Jurčíček Reference Dušek and Jurčíček2016; Lebret, Grangier, and Auli Reference Lebret, Grangier and Auli2016; Mei, Bansal, and Walter Reference Mei, Bansal and Walter2016). One important change happened with the publishing of the transformer architecture (Vaswani et al. Reference Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser and Polosukhin2017), which inspired the “pre-train and fine-tune” paradigm. Later, due to the outstanding performance of this new architecture in both performance and computation consumption, the transformer architecture and its derivatives occupied a dominant position in the NLG domain (Yang et al. Reference Yang, Dai, Yang, Carbonell, Salakhutdinov and Le2019; Floridi and Chiriatti Reference Floridi and Chiriatti2020; Lewis et al. Reference Lewis, Liu, Goyal, Ghazvininejad, Mohamed, Levy, Stoyanov and Zettlemoyer2020). As a sub-task of NLG, TS can also be regarded as monolingual machine translation (Wubben, van den Bosch, and Krahmer Reference Wubben, van den Bosch and Krahmer2012). With the development of sequence-to-sequence machine translation, TS also drew more attention (Guo, Pasunuru, and Bansal Reference Guo, Pasunuru and Bansal2018; Surya et al. Reference Surya, Mishra, Laha, Jain and Sankaranarayanan2019; Omelianchuk, Raheja, and Skurzhanskyi Reference Omelianchuk, Raheja and Skurzhanskyi2021).
2.2. Controllable text simplification
In recent years, researchers introduced explicit parameters to control the simplified output (Nishihara et al. Reference Nishihara, Kajiwara and Arase2019; Martin et al. Reference Martin, de la Clergerie, Sagot and Bordes2020a; Agrawal et al. Reference Agrawal, Xu and Carpuat2021). Martin et al. (Reference Martin, de la Clergerie, Sagot and Bordes2020a) introduced four hyper-parameters in the AudienCe-CEntric Sentence Simplification (ACCESS): the number of characters, Levenshtein similarity (Levenshtein et al. Reference Levenshtein1966), word rank and dependency tree depth, which are used to control the length, similarity, lexical complexity and syntactic complexity, respectively. With the help of the parameters, users can modify the generated simplification based on their needs. However, these parameters may be less straightforward for untrained users, and Agrawal et al. (Reference Agrawal, Xu and Carpuat2021) replaced the detailed parameters with simplification grades. In addition, a minor change in these parameters may significantly affect the readability and fluency of output. Although the value set that maximises the benchmark scores can be given, it may be of little help to the end-users with specific requirements. Further exploration of the effect and proper parameter preferences needs to be made to guide and help untrained users adjust these parameters based on their needs.
Another novel research on the training datasets is multilingual unsupervised sentence simplification (MUSS) (Martin et al. Reference Martin, Fan, de la Clergerie, Bordes and Sagot2020b). They fine-tuned Bidirectional and Auto-Regressive Transformer (BART) (Lewis et al. Reference Lewis, Liu, Goyal, Ghazvininejad, Mohamed, Levy, Stoyanov and Zettlemoyer2020) on their mined paraphrases datasets instead of complex-simple parallel corpora and found that with the help of ACCESS, the unsupervised model outperformed the other unsupervised TS models and became the latest SOTA. As an extension of ACCESS, the authors improved the design of control tokens and changed the tokenization strategy. They showed that performance differences between the two types of datasets might be acceptable only if the mined paraphrase dataset is good enough. Training on paraphrase datasets provides more options than training solely on the supervised datasets and there is a nearly unlimited amount of unlabelled data. They also found that the performance of the combination of unsupervised and supervised training is the best, which is very similar to the pre-train and fine-tune paradigm. Although MUSS also includes multilingual models and tests, the models of different languages were trained separately and had little interference with the aim of this project. Thus, there is no need to focus on other languages for our study of control tokens.
2.3. Metrics
The metrics also play a vital role in evaluating the performance of models. Although current metrics do not compete with human evaluations, they can still partially reflect the performance in certain indexes. Among the popular metrics, there are reference-based metrics like Bilingual evaluation understudy (Papineni et al. Reference Papineni, Roukos, Ward and Zhu2002) and Recall-Oriented Understudy for Gisting Evaluation (Lin Reference Lin2004) and non-reference-based metrics like Flesch-Kincaid Grade Level (Flesch Reference Flesch1948). Currently, the most popular metric for TS is the system output against references and the input sentence (SARI) (Xu et al. Reference Xu, Napoles, Pavlick, Chen and Callison-Burch2016). SARI is designed especially for TS tasks, which evaluates the outputs in aspects of adding, keeping and deleting. Although it is found to have some deviation from human judgement, SARI is still a valuable metric to evaluate simplicity (Alva-Manchego, Scarton, and Specia Reference Alva-Manchego, Scarton and Specia2021). As for the non-reference-based metrics, the BERTScore is a BERT-based metric that evaluates the similarity between the output and references by calculating the correlation in the embedding space (Zhang et al. Reference Zhang, Kishore, Wu, Weinberger and Artzi2019). It is found to have a high correlation with human judgement (Scialom et al. Reference Scialom, Martin, Staiano, de la Clergerie and Sagot2021). By combining the metrics, the performance can be evaluated more comprehensively.
3. Methodology
In this section, we demonstrate the methodology in Fig. 2. There are 4 main steps in the whole system: preprocessing, fine-tuning, optimisation and evaluation.
Figure 2. The reimplementation methodology is represented as a flow chart. We fine-tune BART-base on the preprocessed WikiLarge training set so that the model learns to simplify under the control token or control tokens. After optimisation on the target dataset, the model can apply the optimal value of control tokens to the input and generate desired simplifications.
3.1. Preprocessing
The preprocessing step followed the MUSS project (Martin et al. Reference Martin, Fan, de la Clergerie, Bordes and Sagot2020b). The authors defined four types of prompts used as control tokens to manipulate the features of the outputs. Each control token is designed to represent one character of the sentence. The
$\lt$
DEPENDENCY TREEDEPTH_x
$\gt$
represents the syntactic complexity; The
$\lt$
WORDRANK_x
$\gt$
represents the lexical complexity; The
$\lt$
REPLACEONLYLEVENSHTEIN_x
$\gt$
represents the inverse similarity of input and output at the letter level; The
$\lt$
LENGTHRATIO_x
$\gt$
represents the length ratio of input and output. The value of each control token is calculated based on the reference complex-simple pairs in the training dataset, which is WikiLarge in this project (Zhang and Lapata Reference Zhang and Lapata2017). The WikiLarge dataset (Zhang and Lapata Reference Zhang and Lapata2017) is one of the biggest parallel complex-simple sentence datasets based on various existing corpora and contains 296,402 sentence pairs in the training set. After the calculation, these control tokens will be added to the beginning of complex sentences, and the model will be trained on this preprocessed dataset. In addition to the combined control tokens, this project also explored the effects of a single-control token. In these experiments, only the corresponding control tokens are kept in that dataset.
3.2. Fine-truning
The next step is training. It follows the majority of fine-tuning processes for pre-trained language models. By feeding the preprocessed complex-simple sentence pairs to the model, the model is expected to learn how to simplify texts based on the instruction of control tokens. In our experiments, we explore the use of different tokenization strategies for the control tokens, proposing three possible methods. We also explore the contribution of individual control tokens under each tokenization strategy. In total, 16 models are fine-tuned in the experiment: 1 baseline model, 1 model with all control tokens for each tokenization strategy (3 models) and a separate model for each of our four control tokens under each tokenization strategy (12 models).
3.3. Optimisation
The next step is optimisation. As mentioned in previous sections, the value of control tokens is limited to a small range. All options fall between 0.2 to 1.5 except the Levenshtein, whose upper boundary is limited to 1 due to the calculation method that divides the minimum replacement steps to change from the original sentence to the target sentence by the maximum possible steps of replacement. Only these options are provided during optimisation, and the optimisation problem is reduced to finding the best value combination of control tokens within the optimisation budget. Even though only finite combinations can be applied to the model, the optimisation algorithm is still supported by the Nevergrad (Rapin and Teytaud Reference Rapin and Teytaud2018) API to compare with the current SOTA. Within budget as a limitation to repeat the optimisation process 64 times, the algorithm can find a relatively optimised result in all 334,611 combinations of control token values. Although we have half the optimisation budget compared to MUSS, the system still manages to achieve better performance than MUSS. A sample result of SARI in the optimisation procedure is shown in Appendix A.1. In order to ensure the reliability of the score under the optimised combination, a bootstrapping on the Abstractive Sentence Simplification Evaluation and Tuning (ASSET) (Alva-Manchego et al. Reference Alva-Manchego, Martin, Bordes, Scarton, Sagot and Specia2020a) test dataset will be executed by resampling the dataset 200 times and hence generate a 95% confidence interval.
3.4. Evaluation
The last step is evaluation. Thanks to Easier Automatic Sentence Simplification Evaluation, multiple evaluation metrics can be applied at the same time easily (Alva-Manchego et al. Reference Alva-Manchego, Martin, Scarton and Specia2019). The SARI score is adopted as the primary metric to compare with the current SOTA, while the BERTScore is added as a second reference. Different from the common applications in other projects, the BERTScore in this project is the correlation between the output and references. One coefficient array can be used to combine different evaluation metrics and give a weighted score. However, in this project, we also follow the operations in MUSS and maximise the SARI score, so only the SARI score is taken into account, and the corresponding coefficient is set to 1. The models will be evaluated on the ASSET (Alva-Manchego et al. Reference Alva-Manchego, Martin, Bordes, Scarton, Sagot and Specia2020a) test dataset, which contains 359 complex-simple pairs, and each complex sentence has ten reference simplifications.
3.5. Metrics
SARI is an operation-specific metric for evaluating the performance of the simplicity gained through TS system (Xu et al. Reference Xu, Napoles, Pavlick, Chen and Callison-Burch2016). It offers a comprehensive evaluation by comparing the system output to both the reference and input sentences and considering the add, keep, and delete operations. The final SARI score is calculated in Equation (1) as follows:
\begin{equation} SARI = d_1F_{\text{add}} + d_2F_{\text{keep}} + d_2F_{\text{del}} \end{equation}
where
\begin{equation} d_1 = d_2 = d_3 = 1/3 \end{equation}
and
\begin{align} P_{\text{operation}} & = \frac{1}{k} \sum _{n=1}^{k} \text{p}_{\text{operation}}(n) \quad R_{\text{operation}} = \frac{1}{k} \sum _{n=1}^{k} \text{r}_{\text{operation}}(n) \nonumber\\[9pt] F_{\text{operation}} & = \frac{2 \times P_{\text{operation}} \times R_{\text{operation}}}{P_{\text{operation}} + R_{\text{operation}}} \quad \text{operation} \in \{\text{del}, \text{keep}, \text{add}\} \end{align}
In Equation (1), the different operations have different algorithms, which take into account output (O), the input sentence (I), references (R), and the binary indicator for the occurrence of n-grams
$g$
in a given set (
$\#_g(.)$
). The n-gram precision
$p(n)$
and recall
$r(n)$
are calculated in the following equations:
\begin{equation} \begin{aligned} p_{\text{add}}(n) = \frac{\sum _{g\in O} \min (\#g(O\cap \overline{I}), \#g(R))}{\sum _{g\in O} \#g(O\cap \overline{I})}, \quad \#g(O \cap \overline{I}) = \max (\#g(O) - \#g(I), 0); \\[6pt] r_{\text{add}}(n) = \frac{\sum _{g\in O} \min (\#g(O\cap \overline{I}), \#g(R))}{\sum _{g\in O} \#g(R\cap \overline{I})}, \quad \#g(R \cap \overline{I}) = \max (\#g(R) - \#g(I), 0), \\[6pt] p_{\text{keep}}(n) = \frac{\sum _{g\in I} \min (\#g(I\cap O), \#g(I\cap R_0))}{\sum _{g\in I} \#g(I\cap O)}, \quad \#g(I \cap O) = \min (\#g(I), \#g(O)), \\[6pt] r_{\text{keep}}(n) = \frac{\sum _{g\in I} \min (\#g(I\cap O), \#g(I\cap R_0))}{\sum _{g\in I} \#g(I\cap R_0)}, \quad \#g(I \cap R_0) = \min (\#g(I), \#g(R)/r), \\[6pt] p_{\text{del}}(n) = \frac{\sum _{g\in I} \min (\#g(I\cap \overline{O}), \#g(I\cap \overline{R_0}))}{\sum _{g\in I} \#g(I\cap \overline{O})}, \quad \#g(I \cap \overline{O}) = \max (\#g(I) - \#g(O), 0), \\[6pt] \#g(I \cap \overline{R_0}) = \max (\#g(I) - \#g(R)/r, 0). \end{aligned} \end{equation}
The other automatic metric we applied is the BERTScore (Zhang et al. Reference Zhang, Kishore, Wu, Weinberger and Artzi2019), which is a metric used for evaluating the similarity between the system output and reference sentences. This metric employs cosine similarity to measure the distance in contextual embeddings created by Bidirectional Encoder Representations from Transformers (BERT) (Devlin et al. Reference Devlin, Chang, Lee and Toutanova2019) in a likelihood matrix, applying greedy matching to maximise the matching similarity score. BERTScore comprises two components:
$BERTScore_{precision}$
, which matches the system output with the reference, and
$BERTScore_{recall}$
, which compares the reference against the system output. The two parts are then combined in
$BERTScore_{F1}$
as shown in Equation (5), where
$x$
and
$\hat{x}$
are the references and candidates.
\begin{align} BERTScore_{recall} & = \frac{1}{|x|} \sum _{x_i \in x} \max _{\hat{x}_j \in \hat{x}} x_i^\top \hat{x}_j, \nonumber\\[5pt] BERTScore_{precision} & = \frac{1}{|\hat{x}|} \sum _{\hat{x}_j \in \hat{x}} \max _{x_i \in x} x_i^\top \hat{x}_j, \nonumber\\[5pt] BERTScore_{F1} & = \frac{2 \cdot BERTScore_{precision} \cdot BERTScore_{recall}}{BERTScore_{precision} + BERTScore_{recall}} \end{align}
4. Experiment design
In this section, we will focus on the detailed settings and changes in this project compared to the original ACCESS (Martin et al. Reference Martin, de la Clergerie, Sagot and Bordes2020a) and the experiment settings for the above-mentioned research questions.
4.1. Tokenization strategies
We introduced two new tokenization methods (compared to the original implementation of ACCESS/MUSS) to explore the effects of tokenization strategies. As shown in Table 1, the default tokenization method in the MUSS project is regarding the control tokens as plain text. In comparison, we added 2 more tokenization strategies: The Joint tokenization strategy is to regard the whole control token as one token in the tokenizer; the Separate tokenization strategy is to break the control token into a combination of type and value and add them separately to the tokenizer. These 2 strategies are achieved by manually adding all possible control tokens to the dictionary of the tokenizer. This will affect not only the evaluation and optimisation process but also the training process, thus each tokenization strategy requires an independent fine-tuned model.
Table 1. Tokenization under differing strategies for the input starting with: ’
$\lt$
DEPENDENCYTREEDEPTHRATIO_0.6
$\gt$
’
4.2. Quantisation strategy
We implemented a quantisation strategy for the control tokens in the optimisation step and the corresponding algorithm. In the preprocessing step, there are 4 control tokens calculated and added to the beginning of complex sentences in the complex dataset. As mentioned in the literature review, the 4 types of control tokens are
$\lt$
DEPENDENCYTREEDEPTH_x
$\gt$
(DTD),
$\lt$
WORDRANK_x
$\gt$
(WR),
$\lt$
REPLACEONLYLEVENSHTEIN_x
$\gt$
(LV) and
$\lt$
LENGTHRATIO_x
$\gt$
(LR). As an augmentation to the control tokens, the calculated values are rounded to the nearest 0.05. However, in the optimisation process as originally proposed by Martin et al. (Reference Martin, de la Clergerie, Sagot and Bordes2020a), the calculated values by the algorithm provided by the Nevergrad API (Rapin and Teytaud Reference Rapin and Teytaud2018) are continuous and have long verbose digits, 0.249452…for example. During our reimplementation, we noticed that only the first one or two digits were recognised as input values and the remaining digits did not provide any meaningful instruction. On the contrary, it brought unnecessary information to the system and even lowered the performance of the model. Thus we replaced the continuous values with discrete values (0.2, 0.25, 0.3,
$\ldots$
, 1.0) and changed to the corresponding discrete algorithm in Nevergrad (Rapin and Teytaud Reference Rapin and Teytaud2018).
4.3. Prediction of optimal control tokens
In contrast to the method of finding an optimised set of control tokens at the corpus level adopted in MUSS (Martin et al. Reference Martin, Fan, de la Clergerie, Bordes and Sagot2020b), we propose both regressive and multi-classification methods of predicting the value of each control token at the sentence level. This method is based on two hypotheses: There is a unique ideal simplification for every sentence; By feeding the ideal value of control tokens from the ideal simplification, the model can generate the ideal simplification. Unlike optimising the value of control tokens on the test set, we assume the ideal value is an intrinsic property and can be predicted. In addition, the problem can be regarded as a regression problem to predict the optimal value of control tokens and how to fine-tune autoregression models to do so for every sentence. In the experiment, we chose the average value of multiple reference sentences as an approximation to the ideal value and compared it with the predicted values and the optimised value. Although the median value of reference sentences shows slightly higher SARI and BERTScore on the ASSET test set, the predicted median values are not as good as the calculated values. In addition, both the average and median meet the requirements as an approximation of ideal values and choosing one or the other makes no difference in the conclusion.
In the training process of predictors for each control token, we fine-tuned the BERT-base-uncased model on the filtered WikiLarge dataset (Zhang and Lapata Reference Zhang and Lapata2017), targeting the common end-users and the ASSET test set as well. We filter the sentences with values in the range of 0.2–1.5 and keep the model with the lowest root mean square error within 10 epochs. For each control token, we report the normalised mean absolute error (MAE) and root mean square error (RMSE). We show the distribution difference with the average value or the expectations and values from all reference sentences on the ASSET test set. In comparison to the optimisation method, we test the predictor model with the corresponding single-control token models, which are fine-tuned in previous sections, on SARI and BERTScore. Then we assembled and tested them as a whole system and reported the SARI and BERTScore.
4.4. Reimplementation of ACCESS
We focus our reimplementation on the ACCESS model (Martin et al. Reference Martin, de la Clergerie, Sagot and Bordes2020a). Although MUSS (Martin et al. Reference Martin, Fan, de la Clergerie, Bordes and Sagot2020b), outperforms ACCESS, it does so by using additional training data, as opposed to any innovation concerning the control tokens. As such, we do not use the additional mined paraphrases as used in MUSS for training.
Whereas ACCESS uses the original base Transformer architecture, we use BART as our base model (Lewis et al. Reference Lewis, Liu, Goyal, Ghazvininejad, Mohamed, Levy, Stoyanov and Zettlemoyer2020), which is also adopted in the MUSS project. The original project can be divided into the following sections: data mining, preprocessing, training, evaluation and optimisation.
The original library used for training is fairseq. This project replaced it with another open-source library—Huggingface. Huggingface provides a collection of the most popular pre-trained models and datasets, including BART (Lewis et al. Reference Lewis, Liu, Goyal, Ghazvininejad, Mohamed, Levy, Stoyanov and Zettlemoyer2020) and a unified, advanced and user-friendly API to achieve the most common applications, which makes it easier for future upgrading and modification. The hyper-parameters of models in the reimplementation, including the learning rate and weight decay, are set to be identical to the original project so that the influence of irrelevant factors can be lowered. The last difference between the reimplementation and the original project is the tokeniser. The tokeniser in the reimplementation is the BART-base byte-pair encoding (BPE) tokeniser instead of the GPT2 BPE tokeniser (Radford et al. Reference Radford, Wu, Child, Luan, Amodei and Sutskever2019). Both tokenisers serve the same purpose and perform very similarly to each other. The new one consumes fewer computing resources, which presumably causes only a little effect on the results. Due to the variation of control tokens, the optimisation algorithm has also changed. The original algorithm is the OneplusOne provided by Nevergrad (Rapin and Teytaud Reference Rapin and Teytaud2018), and the current one is the PortfolioDiscreteOnePlusOne, which fits the discrete values better. As for metrics, the SARI score is kept as the primary evaluation method (Xu et al. Reference Xu, Napoles, Pavlick, Chen and Callison-Burch2016), and BERTScore is introduced as a co-reference.
However, due to the limitation of computation resources and mass fine-tuning demands of models with different tokenization strategies, this project also downgraded the training scale and limited the epochs in both baseline and reimplementation. Here are the changes applied to both the reimplementation and the baseline as follows:
-
• All results are from models trained in BART-base instead of BART-large;
-
• All training processes are set to 10 epochs only;
-
• All models are trained on WikiLarge (Zhang and Lapata Reference Zhang and Lapata2017) only.
As explained earlier, each tokenization strategy corresponds to one model and 16 models in total need to be fine-tuned. This is why only BART-base is applied and the training epochs are limited. The reason for choosing 10 as the targeting epoch number is because the training loss for models with combined control tokens has reached 0.85 and decreased very slowly between epochs, while the validation loss started increasing. If continuing training, the over-fitting problem may occur.
5. Results
We firstly report the SARI score of influential supervised TS models with our reimplementations in Table 2. Although MUSS (with mined data) (Martin et al. Reference Martin, Fan, de la Clergerie, Bordes and Sagot2020b) is slightly lower than our reimplementation, our reimplementation stays within the 95% confidence interval of MUSS (with mined data). To verify the significance of the difference in the SARI score, we conducted significance studies against the official output of MUSS (without mined data) with a student’s t-test of the SARI score of the two groups and reported the p-value for the bottom four models. As shown in the table, our reimplementation required fewer resources and training data, while maintaining a significant difference. In addition, we propose the prediction rather than optimisation and get a higher BERTScore in Section 5.4. The bottom two methods in Table 2 are described in Section 5.5.
Table 2. The SARI score of supervised text simplification systems (p in the brackets refers to the p-value of the SARI score against the multilingual unsupervised sentence simplification (without mined data))
5.1. Combined performance
The baseline we obtained by rerunning the original code of the current SOTA is 43.83 on the ASSET (Alva-Manchego et al. Reference Alva-Manchego, Martin, Bordes, Scarton, Sagot and Specia2020a) test dataset. This value is consistent with our reported score on MUSS without minded paraphrasing data, which is 43.63
$\pm$
0.71. The reason we compare this model is that this is the test scenario closest to our reimplementation. There is no confidence interval and BERTScore in the baseline because the baseline is generated by rerunning the code in MUSS by altering specific settings only. The actual output lacks these 2 features. As shown in the 4 rows in Table 3, the SARI score with 95% confidence in the reimplementation is slightly higher than the baseline. However, significance testing shows that the improvement in score according to the different tokenization strategies is not significantly different from the baseline result for any of the cases we examined. This implies that tokenization strategy does not affect control token performance.
Table 3. Results on SARI and BERTScore under differing tokenization strategies, with comparison to the baseline (p in the brackets refers to the p-value of the SARI score against the baseline and shows no statistical significance improvement)
Different from the optimisation target in MUSS, we tried optimising control tokens on both the validation and test set of ASSET to find out the peak performance of the system and the results are shown in Table 4. The top 3 rows show the best SARI score with optimised options of control tokens optimised on the validation set. With optimised control tokens on the validation set, the separate tokenization strategy achieved the highest score within the optimisation budgets, while the joint tokenization method has the highest BERTScore. The middle 3 rows are the results directly optimised on the test set, which shows the upper limit of the model. Among the 3 methods, the separate tokenization strategy had the highest SARI score. Interestingly, the BERTScore is not always proportional to the SARI score, but the BERTScore of optimal value is still quite high. The optimised values of control tokens are pretty close in all situations except the DTD. Similar to the results in Table 3, the p-values show no significant difference between the baseline and the reimplementations. We continue to record the results for each of our three tokenization strategies, however, significance testing again shows that changes in the tokenization strategy have not led to gains which are significantly improved from the baseline in these cases. The bottom 3 rows show the performance difference under a unified value of control tokens. The unified value is the average value of all possible values for each control token. Under the unified condition, the separated one outperformed the other two, and the default tokenization method still performs worse. As for the BERTScore, the joint tokenization method still outperforms the other two.
Table 4. SARI and BERTScore on the ASSET test set under different optimisation targets (p in the brackets refers to the p-value of the SARI score against the baseline and shows no statistical significance improvement). The first three rows optimise on the validation set, the middle three rows optimise on the test set and the bottom three rows show the performance under average value of control tokens
Table 5. Results on SARI and BERTScores of peak points in different control tokens (We choose the points with the highest SARI score in the three strategies and corresponding points in the other two strategies)
5.2. Effects of single-control tokens
In order to verify the effects of each single-control token, a more detailed investigation of the SARI score was done on control tokens, respectively, and the results are shown in Table 5 and Fig. 3. Except for Fig. 3b, all 3 tokenization methods show a high consistency in the curves and have a common minimum at the value of 1. As shown in Table 6, it is mainly caused by the low score in both deletion and adding operations.
Figure 3. The effect of varying control tokens with different tokenization strategies on SARI Score.
Table 6. SARI score by operation at turning points in Fig. 3
Table 5 shows the scores in pairs under a unified value. Similar to the combined control tokens, although the optimised SARI score of either separate or joint tokenization methods is slightly higher than the default tokenization method, there is no significant improvement in the peak performance.
In addition to the peak points, the differences in tokenization methods show little effect on the scores while the value of control tokens can change the performance significantly in the whole curves. In Fig. 3a and c, the separate tokenization method shows the highest peak point, while in Fig. 3b and d, the joint tokenization method has the best performance. More about the BERTScore is shown in Fig. A1 in Appendix A.2.
The Table 6 is designed to help readers better understand the reason for variations in Fig. 3. It shows some local minimum or maximum points within the domain and the corresponding SARI score by operations. The addition score is much lower than the keeping and deletion. It is because there is only limited adding operation in the references and many more expression options to carry a similar meaning, which leads to a low hit rate of the addition operation. At the same time, the keep and deletion are chosen from the existing input and thus have a much bigger hit rate and score.
5.3. Performance of control token predictors
Table 7 shows the performance of regression and expectation and median of multi-classification predictors along with the average variance from the reference sentences. The expectation is the weight product of all predictions with probability, while the median is the prediction that splits the probability distribution into half. The value of the MAE and RMSE is low (0.063–0.149) because the unit distance between the two adjacent control tokens is 0.05.
Table 7. Performance of regressive control token predictors on the average value of ASSET test set (Average Variance means the average value of the variances calculated from different sentence pairs in ASSET test set)
Figure 4. The density distribution of predictions, average values and values of all reference sentences.
Figure 5. The box plot of distributions of predictions, average values and values of all reference sentences for the four control tokens.
Upon examining the performance of the predictors for different control tokens, notable variations can be observed. The DTD_predictor exhibits the highest
$R^2$
score among all four predictors, indicating the strongest ability to capture the relationship between the value of the control token and the corresponding input sentences. Following closely is the LR_predictor, which also demonstrates a respectable
$R^2$
score. In contrast, the LV_predictor in the regression method performs noticeably worse compared to its counterpart in the classification method. The
$R^2$
scores are positive for the predictors of dependency tree depth ratio, replace-only Levenshtein ratio and length ratio in both methods, indicating their proficiency in capturing the desired relationship. However, the
$R^2$
score is negative for the predictor of word rank ratio in both regression and classification methods, suggesting the failure of the proposed predictor to effectively capture the corresponding feature. Lastly, the performance of all classification predictors surpasses that of the regression predictors in the direct evaluation.
Figs. 4 and 5 show the density histogram and box plot of the predicted value, the mean and all values from reference sentences. Fig. 4a shows the histogram of all values in red dots, which exhibits a highly centralised impulse around the value of 1, with a noticeable gap on both sides. The remaining values predominantly occupy the left section of the distribution, ranging from 0.2 to 0.9. In contrast, the distribution of mean values, indicated by the green dashed line, demonstrates less centralisation and a closer approximation to a normal distribution. In Fig. 5a, similar to Fig. 4a, all values have the most outliers and the largest range, while the predicted methods show higher concentration. The classification model overlaps more with mean values in the quartiles compared to the regression model.
Fig. 4b shows the histograms of all values in the red dots and the mean values in the green dashed line, which exhibit a similar shape, with the median values tending to centralise around 0.9. However, both the distribution of the regression model and the classification model appear more concentrated around the value of 0.95. These observations regarding the distribution align with the
$R^2$
scores presented in Table 7, indicating that the predictors struggle to capture the relationship between input sentences and the word rank ratio when compared to the overall means. Similar to Fig. 4b, Fig. 5b shows that the box plot for all values overlaps more with mean values. However, there is a notable long tail due to outliers.
Fig. 4c shows the histogram of all values in red dots, which exhibits a characteristic long tail, while the regression model and mean values share a similar shape. However, considering the
$R^2$
scores provided in Table 7, it becomes evident that the classification model aligns more closely with the mean value. This discrepancy may be attributed to the left section of the regression model, which registers values lower than 0.7. The observed differences in the histograms clarify the performance of the predictors and their relation to the LV control token. Consistent with Fig. 4c, Fig. 5c shows that the classification model has a more centralised distribution, closely overlapping with mean values. The regression model resembles the mean values, but the classification model performs better. There are numerous outliers in all values, indicating extreme data points.
Fig. 4d shows the histogram of all values in red dots, which exhibits a less centralised distribution in comparison to the other three histograms. However, despite this relatively dispersed nature, the mean value itself displays a more centralised shape compared to the predictors. Notably, the histogram of the classification model aligns closely with the mean value distribution, further reinforcing the consistency between these two distributions. This aligns with the distribution characteristics and the
$R^2$
scores presented in Table 7. Fig. 5d presents box plots for control token LR. Although it is difficult to distinguish differences between the classification and regression models, the gap in performance compared to mean values is evident. The box plot for all values aligns with the histogram in Fig. 4, showcasing diversity and variance.
Table 8. Performance of single-control token models with predictors on ASSET test (Regression and Classification methods refer to the situation that single-control models working with the control token predictors, average method means the control token value is the average value of reference sentences and optimization method is the same as shown in previous sections)
5.4. Performance with control token predictors
Table 8 shows the SARI and BERTScore of single-control token models with different methods. The performance gap in SARI and BERTScore among the regression, expectation and median models aligns with the gap in the direct numeric evaluation in Table 7, which demonstrates both the reliability of our chosen metric SARI and BERTScore and competence of mean value as an approximation to the ideal value. As a comparison, the scores of the average and median values calculated from the reference sentences and the optimised value found on the test set are added. Although the SARI scores of predictors and the average value are lower than the optimised ones, BERTScore remains higher. Among all the control tokens, the LV has the closest gap between the two types of methods.
Similar to Table 8, Table 9 shows the SARI and BERTScore with 4 predictors working cooperatively with the combined control token model. The performance difference still aligns with the previous table. However, the SARI score difference between the average value and optimization method is very low, while the average value maintains a higher BERTScore, which shows the average value of multi-reference can be a proper approximation of the ideal value for each sentence.
Table 9. Performance of control token model with predictors on ASSET test. The top three rows are predicted values, and the bottom three rows are calculated or optimised values
Table 10. The performance with only one static value of the control token (The value is referred from Table 4)
5.5. Performance of hybrid method
Considering the poor alignment and performance of single-control token models for some of the predictors, we conducted a hybrid method of both optimization and prediction methods in Tables 10 and 11. In Table 10, we replace the one predictor with the fixed value, referred from the optimization results in Table 4, and found the impact of different control token predictors. The DTD control token predictor has the most significant negative impact on the SARI score, while the LR has the least. By replacing the DTD predictor with a fixed value of 0.35, we increased the SARI score from 42.76 to 43.97.
In Table 11, we replace one of the optimised values with predicted values from the classification method and find the performance differences with control token predictors. Notably, the one with the DTD predictor still shows the largest drop in the SARI score and the one with the LR predictor outperforms the optimization method in both the SARI score and BERTScore.
6. Case study
6.1. Effect and limitation of control tokens
In this section, we show the effect of each control token in examples. Table 12 lists the outputs of one sample sentence with the Length Ratio varying from 1.2 to 0.2 while the other 3 control tokens stay at 1. The difference between adjacent sentences is underlined. The design of this control token is about the sentence length. The output with LengthRatio_1.2 has the highest character length, while the output with LengthRatio_0.2 has the smallest, and there is no variation when the LengthRatio is set to 1. The effect of LengthRatio is significant and well-learned by the model. However, the desired meaning of control tokens in this table is varying the sentence length while maintaining the character, lexical complexity and syntactical complexity identical to the source sentence, which is impossible. Although the model is well affected by the control tokens, the influence of the control token is not absolute and there are still limitations since the control token is not the only constraint applied to the model.
Table 11. The performance with only one dynamic value of the control token and the static value for the remaining control tokens are DTD:0.35, WR:0.7, LV:0.85 and LR:0.9, respectively (The value is referred from Table 4)
Table 12. Effect of varying Length ratio with the others remaining at 1.0
In Table 13, we show a set of outputs of the same source sentence with different ReplaceOnlyLevenshtein ratios while the other 3 control tokens stay at 1. Since the ReplaceOnlyLevenshtein ratio can not be larger than 1, here lists the sentences with ReplaceOnlyLevenshtein ratio varying from 1 to 0.2. The design of this control token is about the similarity in character level between the input and output. In output with LV_1.0, the sentence is identical to the source sentence. In output with LV_0.8 and LV_0.6, the two sentences are identical because the model can only generate meaningful sentences with correct grammar that approach the requirements of control tokens. The output for LV_0.2 shows the highest difference, but the meaning is far from the source sentence, which also shows the limitation of controllable TS with the control tokens that there is no guarantee of the factual reality.
Table 13. Effect of varying ReplaceOnlyLevenshtein ratio with the others remaining at 1.0
Table 14. Effect of varying WordRank ratio and some other ratios with the others remaining at 1.0
In Table 14, we show 2 sets of outputs with different WordRank ratios along with some other ratios and the control tokens unmentioned remain at 1. The design of this control token is about the complexity of words used in the sentence. In the second and third rows, the model replaces the ‘inadvertently’ with ‘deliberately’ and ‘does not follow’. However, when WR is set to 0.4, the model generates identical outputs. In the fifth and sixth rows, we set the LV to 0.8, which allows more variation in the output, and WR to 0.4 and 0.2. The model further replaces the ‘initiates’ with ‘starts’ and ‘migration’ with ‘move away’. Similarly, when WR set 0.2, the model recovers the ‘move away’ to ‘migration’. The examples in this table show the limitation of the interrelationship among all control tokens.
In Table 15, we show the effect of the DependencyTreeDepth ratio. The design of this control token is about the complexity of sentence structure, which normally splits long sentences into shorter sentences. The output sentence with DTD_1.2 is the same as the input sentence. When the DependencyTreeDepth ratio decreases to 0.8, the model splits the long sentence into two shorter sentences. With a lower DTD_0.6, the model further reduces the syntactical simplicity by changing the demonstrative pronoun ‘they’. However, when the DependencyTreeDepth ratio further decreases to 0.4, there is no difference in the output, due to the limitation of other control tokens. This example also shows the inconsistency of the effect of control tokens and the importance of the proper value of control tokens that can affect the sentence significantly.
Table 15. Effect of varying DependencyTreeDepth ratio with the others remaining at 1.0
Based on the observation and comparison among the examples in Tables 12–15, we can confirm that in the given examples, the control tokens work as designed. As shown in Table 14, the control tokens also affect one another. The interference among the control tokens made them work as a whole rather than four separate ones, causing difficulties in tweaking one of the control tokens to achieve the desired outputs. In addition, as shown in the bottom row of Table 13, the misuse of the control token may lead to false content and mislead the users.
6.2. Effect and limitation of optimisation and prediction methods
Tables 16–18 show several examples of simplifications produced by optimisation and prediction methods and the output with average value as the reference on the test set of the ASSET. All three outputs are from the same model with different values of control tokens listed in the second column.
Table 16. Examples of limitations of the optimisation method (changed the meaning in the third row)
Table 17. Examples with mispredicted values (missing ‘or’ clause in the third row)
Table 18. Examples with properly predicted values
In Table 16, the optimisation method shows a typical factual error because of the fixed values in control tokens. There is no relationship between ‘website’ and ‘magazine’ related to ‘taken over’ in the source sentence. However, due to the LR and LV ratios being fixed on the whole test set, the simplification from the optimisation method has to maintain a longer and more different sequence than the prediction method in this case, which makes it tend to generate extra content to fulfil the requirements. As a result, it generates false content and changes the meaning of the source sentence, which will mislead readers. This could be more common in larger and more diverse datasets. Although the output of the prediction method also changed the original meaning, it just loses some meaning without adding false content. The output of the average value best preserves the meaning in this example.
In Table 17, we show the consequence of misprediction. The output of the optimisation method and the average value is quite similar except the length of the optimisation method is a bit longer. In the prediction method, the output sentence is incomplete due to the lower length ratio compared to the average value. Although there is also a gap in the DTD ratio in optimisation and prediction methods, there seems to be no obvious change in the syntactical complexity, which is aligned with the limitations mentioned in previous sections.
In Table 18, we show an example of desired predictions. Although there are some small variances in the value of control tokens, the output of the prediction method is identical to the output of the average value, which is used as the reference sentence. However, the optimisation method changes the original meaning in the source sentence.
We have included SARI and BERTScore at the instance level for each row in Tables 16–18. It can be seen from the variation in these scores that there is sometimes a discrepancy between the accuracy of the simplified text and the value of the score. As such these scores are best used when aggregated over a large dataset, where instance-level variations have less effect.
7. Discussion and future directions
7.1. Effect of quantisation
One phenomenon found during the optimisation section in the original project is that the score of recommended optimisation is even lower than the default values of control tokens at 0.8. A hypothesis emerged that continuous optimisation is not an ideal option to maximise the score. As shown in the four rows in Table 3, the score in reimplementation is higher with equivalent values in the four control tokens. There could be several reasons: the algorithm is not working as expected or the optimisation budget is not large enough to find better optimisations. The default tokenization method in the MUSS project that breaks the control tokens into pieces brings more noise and probably lowers the performance. Apart from the verbosity in optimal values, the long tokenization of the control token is another concern of noisy input. Although the results above show signs of such problems, they may become more serious with the increasing length of control tokens, especially for short sentences. It would be wiser to limit the unnecessary noise in the input to a lower level.
7.2. Similarities and differences among control tokens
Fig. 3 and Table 6 expose the reason for variation with the control token and provide a good illustration of nature in each control token. In single-control tokens, the peak points mainly fall between 0.6 and 0.7, and the score decreases with the value deviating from the peak point. However, there are still some differences among the control tokens. In the DependencyTreeDepth Ratio and Length Ratio, the reduction is more dramatic than the other two. In both graphs, the SARI_add decreases with the value deviating from the peak point and increases slowly when the value is bigger than 1. The SARI_keep and SARI_del fluctuate in the form of two half-phase shifted sine functions and the maximum sum is found in between the peaks. The graph of the WordRank Ratio shows some diversity in both Figs. 3b and 7b among the tokenization methods. Although there is no explanation for the deviations, the deviations show the potential of combining different tokenization methods. When focusing on the main section from 0.5 to 1, the graph shows characteristics similar to the graphs in the previous two control tokens. As for the ReplaceOnlyLevenshtein Ratio, the slope is milder on the left side and it seems to have less effect on the SARI score. Unlike the other three control tokens, this control token can only indicate the intensity of change but not the direction of change. Although the combined effects are still under research, a more effective control token could be a better solution.
As for the optimal value, they are the optimal values within the budget. When rerunning the code, it is quite common to have a different set of optimal values. This is one limitation of the current SOTA system and we propose the predictor to improve this drawback. In addition, the optimal values for one control token and the combined control tokens are different. The correlation among the control tokens presumably causes this variation. If the four control tokens can be designed to work independently, the graph on a single-control token can be directly used to find the optimal value. However, the graph of combined control tokens is bound to have some distortions for now. Based on the detailed graph, it is also clear that the value of control tokens can significantly affect the performance of the models trained in this way and should be treated carefully.
7.3. Control token predictors
In Fig. 4a, the distribution of all values with a green dot line shows a very high centralisation at 1.0 and other values are decentralised. This attribute makes it difficult to be regarded as a regression task and hence has the highest MAE and RMSE in Table 9. In Table 7, the prediction of DTD is less applicable in the four control tokens. In addition, in Table 8, both DTD and WR prediction models show a much lower SARI score than the optimisation method, which indicates the average value may not be a proper approximation to the ideal values. In addition, the design of DTD and WR may hardly reflect the features of sentences and probably needs revisions.
In Table 8, we show that the SARI score of the prediction method and average value for control token DTD and WR is quite low compared to the optimisation method. However, the BERTScore is significantly higher, which is reasonable since the goal set in the optimisation method is to maximise the SARI score only. However, we would like to emphasise the limitations of the SARI score. A sentence with a higher SARI score is not necessarily more meaningful at the sentence level, because the SARI score is only related to the word level. A sentence full of similar add, keep and delete operations to the reference with complete unreadable order will still have a high SARI score. If the only goal is to chase the SARI score, the model may generate some meaningless content as shown in Table 16.
7.4. Hybrid method
In Tables 10 and 11, we tested the hybrid methods with either the prediction method or optimisation method to evaluate the effect of certain control token predictors. Both tables show poor performance with the DTD predictor and WR predictor. However, unlike the poor alignment of WR in Table 7, the DTD demonstrates good performance in
$R^2$
score. The main reason for the poor performance with the DTD predictor may be that this control token is not well designed to reflect the attribute of sentences or it is not suitable for the model to learn. In addition, the optimisation method with the LR predictor outperforms the original optimisation method.
7.5. Future work
During the experiments, we found the following points for further exploration in subsequent studies:
-
• As the control tokens can be regarded as customised prompts to the models, the applicability of control tokens or similar ideas is not only limited to TS tasks. With the popularity of prompt tuning, similar ideas can be applied to other tasks which need control ability to the outputs. In addition, the metric-oriented control token might also help.
-
• As the tokenization strategy shows no statistically significant effect on the performance in the control token experiments, a further study on the longer prompts may be worth doing, especially in the prompt-driven generative models with templates or formats.
-
• The SARI score highly relies on the quality and diversity of the references and can reflect only partially on the simplicity of the output. This could be overcome with a new reference-less metric to measure the readability in different audiences’ perspectives or measure the similarity of the outputs with text genre targeting specific audience groups with large language models.
-
• As shown in Sections 5.4 and 5.5, both the predictors and the average values of DTD and WR improve the SARI score marginally, which might be because of the design of these two control tokens. If some of the control tokens cannot fulfil the design purpose, it might be worth building some new control tokens or fulfilling the designed goal differently.
-
• As shown in Section 6.1, there is interference among the control tokens, impairing the usability and practicality of control tokens. One alternative solution is to break the system into several subsystems, build a pipeline to normalise the sentence and divide and conquer the subtasks in steps.
-
• Last but not least, the predictor learns only from the WikiLarge dataset, and cannot serve the purpose of giving precise predictions based on the needs of different user groups. To better achieve the goal of customised simplification for different user groups, further improvement in the control token-predictor mechanism is needed.
7.6. Concluding remarks
In the investigation, we reimplemented an updated version of ACCESS with fewer resources and equivalent performance and uploaded it to Huggingface. We show the effects and interactions of control tokens with different values and their effectiveness and importance, which can be used to balance user intention and performance. We give some suggestions for improvements in quantisation, prove the insignificance of different tokenization strategies of control tokens in our settings and propose a generative method on the value of control tokens to improve the generalisation on different datasets. In the generative method, we test the performance of predictors of different control tokens, increase the whole system’s performance with this newly proposed method and explore the potential of this prediction method. Although there remain limitations in both the design and prediction of control tokens, the prediction method can promote the popularity of control token methods. At last, we verified the insignificance in the performance for each single research question we set, but several small improvements together can achieve significant improvement.
Acknowledgement
Some sections of the content of this manuscript have been presented previously at the Text Simplification, Accessibility, and Readability (TSAR) workshop at the EMNLP2022 (Li, Shardlow, and Hassan Reference Li, Shardlow and Hassan2022). This manuscript expands significantly upon the previously presented work.
Appenidx A
Figure A1. SARI score of 128 times of optimisation.
A.1. Optimisation budget
Fig. A1 shows the SARI score during the optimisation procedure in 128 times attempts. The highest score appears at the35th attempt and four of the top-five scores appear within 64 times. Even though a higher SARI score can be between 65 and 128, there is only a small performance gap between the highest and the second-highest score. Thus it is reasonable to set the optimisation budget to 64.
A.2. Detailed singile control token results
Fig. A2 shows the BERTScore. Nearly all 3 tokenization strategies show high similarity to each other except Fig. A2(b). The figures show that nearly all models have the highest BERTScore around 1. Since the BERTScore calculates the correlation between the output and references, when the control token is set to 1, the model processes nothing, and the output can be quite similar to some of the references. Under this situation, as shown in Table 6, the SARI keep reaching the top. However, the peak of the BERTScore in Fig. A2(c) shifts to the left marginally, which shows that the references and input are not identical.
Figure A2. The effect of varying control tokens with different tokenization strategies on BERTScore.
A.3. More insights into the two outputs
We manually check the system outputs between the official system output of BART trained on the WikiLarge from MUSS and ours. In Table A1, we listed the findings and limitations of both models. The major differences include the following types:
-
• Shorter and separate sentences: We find many examples showing that our system prefers to split a long sentence into several shorter sentences. This may benefit aphasia people, who find it hard to follow very long sentences. On the other hand, it may also set obstacles to people having difficulties understanding the pronouns.
-
• Hallucination and false content: Both systems still generate hallucination and false content, which are inconsistent with the inputs. These false contents will undermine the practicality of the simplification systems. However, some of the false contents are caused by the control tokens themselves and can be alleviated by giving proper input value or a new control mechanism.
-
• Omitting contents: We also found that the two systems might have different preferences in deleted contents. For now, there lack of proper control mechanisms in control tokens to decide which part is the key information in the sentence and needs further improvement.
Table A1. More insights in the two systems
Open access
