- Center for Cognitive Interaction Technology, Bielefeld University, Bielefeld, Germany
Sequence labeling is pervasive in natural language processing, encompassing tasks such as Named Entity Recognition, Question Answering, and Information Extraction. Traditionally, these tasks are addressed via supervised machine learning approaches. However, despite their success, these approaches are constrained by two key limitations: a common mismatch between the training and evaluation objective, and the resource-intensive acquisition of ground-truth token-level annotations. In this work, we introduce a novel reinforcement learning approach to sequence labeling that leverages aggregate annotations by counting entity mentions to generate feedback for training, thereby addressing the aforementioned limitations. We conduct experiments using various combinations of aggregate feedback and reward functions for comparison, focusing on Named Entity Recognition to validate our approach. The results suggest that sequence labeling can be learned from purely count-based labels, even at the sequence-level. Overall, this count-based method has the potential to significantly reduce annotation costs and variances, as counting entity mentions is more straightforward than determining exact boundaries.
1 Introduction
Sequence labeling represents a pervasive framework in Natural Language Processing (NLP), encompassing tasks such as Named Entity Recognition (NER), Part-Of-Speech tagging (POS), and Semantic Role labeling (SR), as well as Question Answering (QA) and Information Extraction (IE). These tasks have frequently been addressed using supervised learning approaches that require a labeled dataset with ground-truth sequences. Notable examples of supervised approaches for tackling sequence labeling include conventional Hidden Markov Models (HMM) (Kupiec, 1992), Conditional Random Fields (CRF) (Sha and Pereira, 2003), and (neural) sliding windows (Gallo et al., 2008), as well as deep neural networks such as Recurrent Neural Networks (RNN) (Graves, 2012) and, more recently, the Transformer architecture (Vaswani et al., 2017; Devlin et al., 2019). However, despite the continually improving performance in sequence labeling (Li et al., 2020; Zhang et al., 2023), supervised approaches are constrained by two technical limitations:
• Training vs. evaluation: There exists a common disparity between the training objective, typically a differentiable loss function, and the task-specific, possibly discrete evaluation metric, such as the F1-score. Consequently, minimizing the loss function might not directly optimize the evaluation measure, resulting in off-target training.
• Labeling datasets: In (standard) supervised machine learning, a labeled dataset with fine-grained, task-specific ground-truth annotations y1, ..., yn is normally required, but the process of acquiring such ground-truth annotations can be resource-intensive, depending on the application.
In reinforcement learning (RL), the learning progress is achieved by maximizing an arbitrary, possibly discrete reward-function R (the training objective). Since R is arbitrary, it exhibits two important properties: (a) R can directly represent and therefore optimize the evaluation measure (e.g. F1-score), and (b) training does not explicitly demand a labeled dataset . Thus, RL naturally overcomes the above-mentioned limitations for supervised machine learning approaches (Keneshloo et al., 2020).
Yet, despite achieving notable success in gaming (OpenAI et al., 2019; Vinyals et al., 2019; Ye et al., 2020), robotics (Zhu et al., 2020; Akalin and Loutfi, 2021; Raffin et al., 2022), and planning (Zhu et al., 2021; Hamrick et al., 2021; Esteso et al., 2022), the utilization of reinforcement learning in natural language processing remains limited. While RL methods have been employed in, e.g., question answering (Choi et al., 2017; Buck et al., 2018) and information extraction (Narasimhan et al., 2016; Qin et al., 2018), the approaches considered are specifically engineered for rephrasing questions (Buck et al., 2018), denoising datasets (Qin et al., 2018), and assembling or condensing information (Choi et al., 2017; Narasimhan et al., 2016), as opposed to directly tackling the objective as an RL problem. As noteworthy exceptions, RL methods have been adopted to directly optimize policies in dialogue systems (Li et al., 2016; Lu et al., 2019; Liu et al., 2020) and paraphrase generation (Li et al., 2018; Qian et al., 2019; Siddique et al., 2020), including the fine-tuning processes of large language models such as InstructGPT (Ouyang et al., 2022) and GPT-4 (OpenAI, 2023), albeit with supervised pre-training.
Similarly, various methods have been proposed to address NER with RL (Wang et al., 2018; Yang et al., 2018; Wan et al., 2020; Peng et al., 2021a,b), but the proposed techniques only formulate secondary operations from an RL perspective. In some works, the RL methods are employed to pre-process incomplete or inaccurate annotations to accommodate strong(er) supervision by detecting and removing, sampling or cleaning negative and noisy instances (Yang et al., 2018; Peng et al., 2021a,b). In other works, the RL methods are instead utilized to complement a (traditional) supervised tagging approach by identifying and correcting invalid predictions (Wang et al., 2018; Wan et al., 2020).
The limited adoption of RL in NLP could, depending on the application (Uc-Cetina et al., 2022), be explained by the challenge of expressing the environment as an appropriate and well-defined sequential Markov decision process, as well as the notorious instability in training and low sample-efficiency when addressing complex learning problems or environments (Yu, 2018). In addition, designing a suitable reward function for effective learning can be challenging and is oftentimes accompanied by delayed rewards (Eick, 1988) or sparse rewards (Minsky, 1961), resulting in the well-known (temporal) credit assignment problem (Minsky, 1961). To mitigate this, meticulous reward shaping (Eschmann, 2021) or extensive exploration (Amin et al., 2021) may be necessary.
In this work, we present a novel RL-based approach that (a) considers sequence labeling exclusively from an RL perspective, and (b) does not strictly require token-level annotations for training. To accomplish this, we condense (or aggregate) standard token-level labels to summarize the ground-truth annotations by counting entity mentions. Then, we generate feedback for training by comparing the predicted and annotated entity counts. We experiment with combinations of feedback aggregation (i.e., multiple predictions are assigned a single reward signal) and reward functions, both count-based and standard (that is, with direct access to token-level labels), while evaluating our approach on the NER datasets CoNLL-2003 (Sang and Meulder, 2003), OntoNotes 5.0 (Hovy et al., 2006), and BC5CDR (Wei et al., 2016). In multiple instances for standard feedback, we obtain results that are competitive with a standard supervised baseline (i.e., that minimizes the cross-entropy loss), even outperforming the baseline by 2.33 points in F1-score on BC5CDR. For count-based feedback at the sequence-level, we obtain results that are only 11.37 and 9.56 points behind the standard baseline for CoNLL-2003 and BC5CDR, respectively. In summary, our findings indicate that learning sequence labeling tasks, such as NER, simply by counting entity mentions is possible and feasible, achieving remarkably solid performance. Such count-based methods could significantly reduce annotation costs as well as variances between annotations, as counting specific entity mentions is more straightforward and less subjective than determining precise entity boundaries.
2 Method
2.1 Preliminaries
(Deep) reinforcement learning algorithms are conventionally implemented through a sequential Markov decision process (MDP)—a mathematical framework used to determine a suitable environment E to be interacted with – and is denoted by with state space , action space , transition function T (potentially stochastic) and reward function R. Subsequently, an agent (i.e, the learning system), whose actions on states are dictated by a (typically stochastic) policy function π(st), interacts with the environment E over a sequence of discrete time-steps via state-action pairs (s0, a0), (s1, a1), ..., (st, at), and, in turn, observes rewards rt = R(st, at, st+1) upon each transition st+1~T(·|st, at) as feedback from the environment. Ultimately, the objective function to be optimized by the policy function π is the expected cumulative discounted reward with a discount factor γ, when following the policy function π, i.e., by selecting the action at~π(·|st) that maximizes the expected cumulative discounted reward at each time-step t.
2.2 Framework
We begin by formalizing the well-known framework of sequence labeling as a straightforward Markov decision process. Let denote a sequence of tokens x = x1, ..., xn with ground-truth annotations y = y1, ..., yn from a labeled dataset (e.g., CoNLL-2003). We comprehend the sequence x = x1, ..., xn with respective predictions ŷ = ŷ1, ..., ŷn (i.e., the actions a1, ..., an chosen by the agent) as an episode, and therefore construct the state space from the starting state space , intermediate state space , and terminal state space as , with:
where the states denote that position t (token xt) in sequence x shall be processed next.
Notice that, by construction of the state space , an inherent disregard (or independence) for previously generated predictions ŷ1, ..., ŷt−1 and subsequent predictions ŷt+1, ..., ŷn, respectively, is entailed, therefore suggesting (the application of) a memoryless policy function π. As further implied by the state space , once an action is selected, each non-terminal state st = (x, t) is deterministically transformed into an intermediate (or terminal) state st+1 = (x, t+1) by the transition function T, regardless of the assigned prediction, i.e., T describes a bijection from to . The token labels task-specific to named entity recognition (e.g., O, B-PER, and I-MISC) must, of course, be accordingly represented by the action space , and shall be characterized by non-negative integers . Lastly, we implement a framework of reward functions R to evaluate a sequence of consecutive predictions ŷi, ..., ŷj (i.e., actions ai, ..., aj) against the ground-truth annotations yi, ..., yj, permitting any evaluation measure, and subsequently communicate the aggregated reward (or feedback) through the environment E.
2.3 Agent
We proceed by describing the architecture and behavior of our learning system (i.e. the agent) when operated by the policy function π. In value-based RL, such as Q-Learning, we choose some action based on the estimated state-action value Qπ(st, a) given state at time-step t. Specifically, this Q-value estimate represents the expected, long-term cumulative discounted reward Eπ[Rt] when choosing action a at time-step t while being in state st, and greedily following π thereafter. Thus, we estimate the state-action values Q(st, ·), where st = (x, t) and x = x1, ..., xn, as follows:
The Encoder is assumed to generate the contextualized representations (or hidden states) , with d∈ℕ, corresponding to the sequence x1, ..., xn, and the weight matrix and bias term generate the state-action value predictions from h1, ..., hn.
To address the dilemma of balancing exploration and exploitation (thereby defining our policy function π), we simply pursue an ϵ-greedy strategy, due to the relatively compact action-space . Therefore, π(s) can be expressed as:
where denotes uniform sampling from .
2.4 Reward schemes
We continue by establishing the partitioning mechanism by which the reward signals are delayed and aggregated. Each episode (i.e., a sample to be labeled) is segmented into independent subsections, each of which, once traversed and processed by the learning algorithm, is evaluated with an aggregated (and singular) reward signal. We segment an episode according to (a) the respective ground-truth annotations, and (b) the currently active reward scheme, which governs the breadth of a segment and, as a consequence, directly controls the degree by which the reward signals are delayed and aggregated. We implement the following reward schemes:
• By action: each prediction ŷ1, ..., ŷn is evaluated separately by R(yt, ŷt).
• By region: a sequence of predictions ŷi, ..., ŷj corresponding to a homogeneous (and maximal) sub-sequence of annotations yi, ..., yj (i.e., named entities and non-entities) is evaluated in aggregate by R(yi...j; ŷi...j).
• By entity: a sequence of predictions ŷi, ..., ŷj generated by separating the annotations y1, ..., yn behind each named-entity is evaluated in aggregate by R(yi...j, ŷi...j).
• k-grouped: The concatenation of k completed sequences of predictions ŷ1, ..., ŷk with corresponding annotations y1, ..., yk is evaluated in aggregate by R(y1...k, ŷ1...k), whereby the frequency of reward-signals decreases from rewards-per-sample to samples-per-reward.
We provide an example for the By-Action, By-Region, By-Entity, and 1-Grouped reward scheme in Figure 1. The example assumes a gold-label sequence y1, ..., y7 with one LOC-type entity and one PER-type entity, spanning one LOC and two PER token labels, respectively. Obviously, the underlying input sequence x1, ..., x7 is irrelevant when calculating the rewards. For predictions ŷ1, ..., ŷ7, the By-Action scheme evaluates each individual prediction ŷt via the corresponding token label yt through R. The By-Region scheme, in contrast, aggregates successive token labels into sub-sequences based on label class, e.g., consecutive LOC, O, and PER token labels in Figure 1. For predictions ŷi, ..., ŷj over any such homogeneous sub-sequence, the feedback (one reward per section) is calculated in aggregate via the corresponding token labels yi, ..., yj through R. As an extension, the By-Entity reward scheme combines two adjacent sections (or regions) as outlined in By-Region – e.g., by merging the initial O-region with the subsequent PER-region in Figure 1—thus providing one reward per two regions. Lastly, the k-Grouped reward scheme evaluates k prediction sequences ŷ1, ..., ŷk (using k = 1 in Figure 1) via the corresponding gold-label sequences y1, ..., yk, thus communicating one reward per k input sequences x1, ..., xk.
Figure 1. The considered reward schemes. Here, visually clustered sections (i.e., partitioned annotations) are to be evaluated in aggregate.
We simplify our notation by assuming that, supposing a sequence of predictions ŷi, ..., ŷj to be evaluated in aggregate, the environment E communicates a reward signal rt following each interaction at such that ri, ..., rj−1 are valueless (e.g., ⊥) and rj represents the evaluation of the predictions ŷi, ..., ŷj against the ground-truth labels yi, ..., yj. By design, the aggregated reward schemes are both delayed and sparse, because a singular non-empty reward (that is, our feedback for training) is only communicated through the environment E once a partition, as dictated by the current scheme, has been processed by the agent.
2.5 Reward functions
We consider various types of reward functions in our framework. Beyond Exact Match, Accuracy, and F1-score as a reward function R, we further experiment with several variants of the Cosine Similarity to compute a similarity between the predicted and target sequence. Apart from measuring the Cosine Similarity between the ground-truth annotations yi, ..., yj (again, represented by non-negative integers ) and the generated predictions ŷi, ..., ŷj, we further compare (i.e., calculate the similarity) of the entity counts per class between the predictions and ground-truth labels. This aggregate reward abstracts from the actual token-level annotations as, in contrast to standard reward functions such as Accuracy and Exact Match, these count-based reward functions only consider the amount of entities annotated in yi, ..., yj and predicted in ŷi, ..., ŷj, see Figure 2.
Figure 2. The method for calculating similarities by comparing the overall number of entities annotated in y (in green) and predicted in ŷ (in blue).
Note that, when calculating count-based feedback, the function R is actually computed over vectors of entity counts count(yi...j) and count(ŷi...j) rather than token labels yi...j and predictions ŷi...j directly. In practice, count(yi...j) would, of course, be obtained from x via annotation. To simplify our notation, we assume that R handles the counting whenever necessary. For further details, see Section 4.1.
A cardinal problem with employing Cosine Similarity as a reward function R, however, becomes apparent when the generated predictions ŷi, ..., ŷj (or amount of inferred entities) represent a multiple of the ground-truth annotations yi, ..., yj, because only the directions of vectors A and B are considered. As a consequence, a sequence of sub-optimal predictions ŷi, ..., ŷj might be recognized as an optimal solution, as R(y, ŷ) = R(y, y). To address this problem, we incorporate a modification to the original formula to account for deviations in magnitude by which, therefore, a perfect reward is only achieved, if and only if A = B:
2.6 Algorithm
We outline our single-episode learning procedure in Algorithm 1 where, because our reward schemes are considered a characteristic of the environment E (to simplify our notation), the reward schemes and functions are indirectly addressed via E(at|R). In Line 11, the model parameters θ are optimized according to the Mean Squared Error between the predicted and resulting state-action values and q1, ..., qn, respectively:
We implemented several modifications to standard Deep Q-Learning (Mnih et al., 2013):
Firstly, we eliminated the experience replay (Lin, 1992; Fedus et al., 2020), because a sequence of continuous predictions ŷi, ..., ŷj (as determined by the reward scheme) might, in consequence, not be evaluated in aggregate, since the evaluation of a particular prediction ŷt is dependent on the evaluation of the associated sequence ŷi, ..., ŷj containing ŷt. Additionally, by discarding the experience mechanism, each prediction ŷ1, ..., ŷn can be computed from the same contextualized representation h1, ..., hn, requiring only a single encoding per sequence x1, ..., xn, as the parameters θ are yet to be updated.
Secondly, because the aggregated subsections are evaluated separately (predictions are independent by design of our framework), we introduce gated discounting (via γt) to encourage short-term strategies within aggregated subsections and discourage long-term strategies across aggregated subsections. To accomplish this, we condition γ on the received feedback:
Note that γt is always 0 whenever a non-empty reward-signal is observed by the learning algorithm, effectively separating two consecutive subsections (as seen by the agent).
Thirdly, we replace the original Q-value estimates (Mnih et al., 2013) with non-terminal ground-truth Q-values qt to propagate the upcoming, non-empty reward-signals directly within their respective partitions.
By introducing this modification, we associate a single (yet discounted) evaluation rj with a complete sequence of predictions ŷi, ..., ŷj and, as opposed to producing purely local estimates, encourage the agent to estimate the sectional evaluation of ŷi, ..., ŷj through each Q-value estimate .
2.7 Experiments
We utilize a comparatively lightweight BERT checkpoint (bert-base-cased1) sourced from HuggingFace as our base-model. This checkpoint is configured with 12 transformer blocks, a hidden dimension of 768, and 12 attention heads, totaling approximately 110 million pre-trained parameters. As a consequence, the output-layer W (which is used for classification) is composed of parameters, which we randomly initialize from , where .
The individual experiments are conducted over 400 rounds, during each of which 250 updates are performed on the model-parameters θ, amounting to 100,000 updates per experiment. The updates are performed over batches of 8 sequences, sampled uniformly at random. The exploration-exploitation dilemma is addressed by selecting ϵ = max (0.005, 0.5round−1), such that ϵ is never below 0.5%, while discounting is handled with γ = 1. We maintain a constant learning rate α of 1e-5 and utilize AdamW with standard parameters for optimization. We calculate the learning system's performance using seqeval,2 an open-source framework for sequence labeling evaluation.
We implement Exact Match, Accuracy, F1-score, and the enhanced Cosine Similarity from Equation 7 as standard reward functions. Additionally, we use the enhanced Cosine Similarity function for comparing the entity counts contained in the ground truth annotations and predictions, utilizing four configurations: Counts, Counts +O, Counts +P, and Counts +OP. The Counts function enumerates all annotated or predicted named entities, while +O variants also consider the contiguous O-intervals (regions of consecutive O token labels) in a sequence, and +P variants may only enumerate the true-positive entity counts (and O-intervals, if +O), directly assigning a 0-reward to predictions ŷt that are impossible considering the annotated entity counts, see Figure 3. We consider the following reward schemes: By-Action, By-Region, By-Entity, 1-Grouped, 2-Grouped, and 4-Grouped.
Figure 3. The method by which +O and +P counting variants (in yellow) calculate and provide feedback. In this sketch, the reward signal r (in red) is communicated at the sequence-level 1-Grouped. The +O counting variant includes contiguous O-intervals. The +P counting variant provides 0-rewards for predictions yt (in blue) that are impossible given the ground-truth annotations y (in green).
We evaluate our approach on the following datasets for named entity recognition in English: CoNLL-2003, OntoNotes 5.0, and BC5CDR. CoNLL-2003 (Sang and Meulder, 2003), a dataset sourced from news articles, encompasses four categories of named entities: person (PER), location (LOC), organization (ORG), and miscellaneous (MISC). OntoNotes 5.0 (Hovy et al., 2006), compiled from news articles, weblogs, and dialogues, presents a wider array of named entities, featuring 18 categories that include 11 entity classes (such as building, event, and product) and 7 value types (such as percent, time, and quantity). The original BC5CDR dataset (Wei et al., 2016) consists of biomedical documents annotated for mentions of diseases and chemicals. However, for our purposes, we utilize the sentence-based version pre-processed for T-NER (Ushio and Camacho-Collados, 2021). We employ the default dataset splits (see Table 1).
Finally, to establish a benchmark for comparison, we introduce a standard baseline. This baseline is obtained by training our base-model (BERT) via standard supervised learning, minimizing the cross-entropy loss. In contrast to our RL approach, the baseline has direct access to the token-level annotations. The experimental procedures and evaluations are otherwise identical.
3 Results
We showcase our main experimental results in Table 2. The reported results are clustered by standard feedback, count-based feedback, and the standard baseline, and represent the maximum observed F1-scores on the validation split (averaged over 5 runs). In the following, we distinguish between standard and count-based feedback.
Table 2. The average maximum F1-scores (computed over validation datasets) with respective standard deviations.
3.1 Standard feedback
Unsurprisingly, the results reached by training the learner using standard reward functions that calculate feedback based on token-level annotations are relatively consistent when combined with sub-sequence reward schemes like By-Action, By-Region, and By-Entity. For CoNLL-2003 and OntoNotes 5.0, the highest results (using Accuracy as feedback) are generally competitive with the standard supervised baseline of 94.16 and 86.06 points in F1-score, respectively. Notably, the highest results for BC5CDR outperform the standard baseline (83.45 F1-score) by 1.08 to 2.33 points in F1-score for reward schemes By-Action to By-Entity, even remaining competitive for feedback aggregated to sequence-level 2-Grouped (81.31 F1-score).
As we transition from reward scheme By-Action to sequence-level 1-Grouped, performance naturally deteriorates as feedback becomes more aggregated; this decrease is especially noticeable for OntoNotes 5.0, with the highest results falling by 30.69 points in F1-score, whereas the decrease for CoNLL-2003 and BC5CDR is limited to 5.82 and 2.74 points, respectively. Notably, when switching from 1-Grouped to 2-Grouped, the highest results are relatively stable for BC5CDR, only dropping by 0.48, while results for CoNLL-2003 and OntoNotes 5.0 decrease by 11.07 and 18.00 points in F1-score.
When providing feedback via the By-Action reward scheme, Exact Match and Accuracy as reward functions (whereby each individual prediction ŷt is assigned a 0/1-reward) produce the highest results across all evaluated scenarios and datasets, even outperforming the standard supervised baseline. Notice that when the feedback is conveyed as an F1-score, performance drops significantly. While the reward signals for By-Action are communicated as reward-per-action, the F1-score, unlike Exact Match or Accuracy, is generally not applicable to NER when calculated over singular token-labels. To illustrate this, we propose a scenario where the ground-truth label is yt = I-PER and the prediction is ŷt = B-PER. In this case, the F1-score calculated by seqeval yields F1(yt, ŷt) = 1.0, whereby the learner is unable to distinguish between B-PER and I-PER.
For aggregate feedback via By-Region and By-Entity, Accuracy mostly yields the highest results for all datasets, with performance shrinking by 0.66, 3.47, and 0.50 points for CoNLL-2003, OntoNotes 5.0, and BC5CDR, respectively. In comparison, feedback produced by the much less informative Exact Match achieves only slightly worse results for CoNLL-2003 and BC5CDR. However, results plummet by 24.48 points for OntoNotes 5.0. Once feedback is provided at sequence-level 1-Grouped to 2-Grouped, we observe a notable decrease in the results yielded by Accuracy and Exact Match. In general, F1-score achieves the highest results for feedback provided at sequence-level k-Grouped, remaining remarkably stable as k increases, with performance dropping by at most 23.78, 25.88, and 26.40 points for CoNLL-2003, OntoNotes 5.0, and BC5CDR.
3.2 Count-based feedback
In general, count-based feedback is expected to facilitate a reduced performance when compared to token-based supervision, as it provides less concrete feedback to the learning system. For instance, when considering the sub-sequence reward schemes By-Action, By-Region, and By-Entity, we observe a substantial decrease in performance between conventional and count-based reward functions.
However, when considering sequence-level reward aggregation 1-Grouped, the difference in performance between standard feedback (e.g., F1-score) and count-based feedback (e.g., Counts +OP) is surprisingly low. Specifically, metrics only decrease by 4.64 and 7.90 points for CoNLL-2003 and BC5CDR, and increase by 4.79 points for OntoNotes 5.0, when switching from F1-score to Counts +OP. This is remarkable given the stark contrast in training regimes and, even more so, the supposedly unreliable information conveyed by count-based feedback over feedback directly computed from token-level annotations. Furthermore, the results gained via Counts +O and Counts +OP are competitive with token-based feedback for sequence-level reward aggregation k-Grouped, even outperforming the strongest standard reward functions for CoNLL-2003 and OntoNotes 5.0 (i.e., F1-score and Accuracy) by 16.94 and 24.71 points. Obviously, the highest scores are generally obtained via informed counting with Counts +OP, as it provides the most nuanced feedback to the learning system, whereas naïve (or uninformed) counting, as executed in Counts, consistently yields diminished performance for all experiments.
In addition, results remain reasonably consistent across By-Action to 4-Grouped for count-based feedback, exhibiting a maximum difference (over highest scores) of 13.10, 13.74, and 35.78 points for CoNLL-2003, OntoNotes 5.0, and BC5CDR, respectively. In contrast, results for token-based feedback have relatively high variance, displaying a maximum difference (again, over highest scores) of 33.16, 60.84, and 60.52 for CoNLL-2003, OntoNotes 5.0, and BC5CDR.
Notice that, by design of the reward schemes By-Region and By-Entity (and, trivially, for By-Action), even when considering count-based annotations, the learner is implicitly provided with information about the underlying token-level annotations, as partitions from By-Region and By-Entity are constructed such that they comprise at most one named entity. The reward schemes By-Region and By-Entity thus provide an interesting perspective on the differences in performance when transitioning from By-Region to By-Entity to 1-Grouped. For instance, looking at the highest results for count-based feedback, we observe a significant decrease in performance from By-Region to By-Entity, suggesting that recognizing the boundaries of singular named entities is particularly challenging when provided only with count-based feedback. However, when feedback aggregation is elevated from By-Entity to 1-Grouped (i.e., to sequence-level), results decrease only slightly for OntoNotes 5.0 and BC5CDR, even increasing by 8.51 points for CoNLL-2003, indicating that detecting (the boundaries of) multiple named entities is relatively straightforward when viewed from By-Entity.
4 Discussion
The findings outlined in Section 3 demonstrate that learning sequence labeling tasks, such as NER, with aggregate feedback is feasible, even when the feedback is derived entirely by counting entity mentions per class, although with some obvious caveats. In comparison to feedback computed from token-level annotations, the count-based rewards facilitate a reduced learning capacity, providing relatively imprecise and, in part, unreliable information to the learning algorithm. By design, reward signals derived from entity counts over generated predictions ŷ1, ..., ŷn and ground-truth annotations y1, ..., yn only communicate information pertaining the existence of an entity, not its respective boundaries. Nevertheless, overall results are remarkably solid considering these constraints.
In Section 1, we briefly explore the advantages of utilizing RL methods over standard supervised learning techniques, especially pertaining to the implications of an arbitrary reward function R dictating the learning progress. This function R can directly represent and therefore optimize the evaluation measure, including the F1-score. Looking at Table 2, the experiments on standard reward functions, which calculate feedback from token-level annotations, support this assumption for sequence-level feedback 1-Grouped (and beyond), as designing the function R to compute the current F1-score between the gold-labels y1, ..., yn and the predictions ŷ1, ..., ŷn) is indeed shown to outperform the token-based alternatives, such as Accuracy. However, while token-based feedback at the sequence-level k-Grouped achieves its greatest potential when representing the F1-score between y1, ..., yn and ŷ1, ..., ŷn, we observe that count-based feedback often surpasses token-based feedback (including the F1-score) while achieving more consistent performance.
As detailed in SubSection 3.2, our results reflect the significance of informed counting, as demonstrated by Counts +OP versus Counts. While Counts +O (considering contiguous O-intervals, i.e., non-entities) and Counts +P (providing 0-rewards on false-positive counts) both provide some contrasting information to the learning system, the resulting increase in performance from Counts +O overshadows the improvements gained from Counts +P. Furthermore, when integrated as Counts +OP, a significant and consistent improvement in performance (and standard deviation) is achieved over both configurations, especially for CoNLL-2003 and BC5CDR.
Notably, our results suggest that learning progress for count-based feedback may be negatively influenced by the number of entity types, that is, the cardinality of the action space . For instance, even when considering the reward scheme By-Action, the difference in performance to the standard baseline is 6.78 points for CoNLL-2003 and -0.11 points for BC5CDR, which have and , respectively. In comparison, this difference is exacerbated to 29.64 points for OntoNotes 5.0, where , thus raising questions regarding the suitability of count-based feedback at the sequence-level when handling sequence labeling tasks with a relatively large action space .
As an alternative explanation, the divergence in performance could instead be caused by class label imbalances during training. In fact, OntoNotes 5.0 (k = 18 classes) exhibits the greatest label imbalance, with 4 entity types constituting more than 66% of overall entity mentions, whereas CoNLL-2003 (k = 4) and BC5CDR (k = 2) feature a more balanced class distribution, see Table 3. Hence, our overall results might be improved by employing a more sophisticated algorithm for sampling training instances from (as opposed to uniform random sampling). We provide further material for this correlation in Section 4.1.
Table 3. The absolute (ci) and relative (ci/n) entity counts per label class and dataset (train split).
4.1 Case study
In this Section, we provide some concrete model predictions to demonstrate how count-based feedback is calculated and distributed over predictions, thus promoting a more comprehensive understanding of our methodology. In addition, considering the imprecision and granularity of count-based feedback, where no boundary-related information is communicated, the exemplary predictions shall emphasize the remarkable performance on boundary detection.
The following examples were generated for CoNLL-2003 and obtained by training with count-based feedback (Count +OP) at the sequence-level 1-Grouped as described in Section 2.7. The example predictions and resulting per-class and absolute model performance are presented in Tables 4, 5, respectively. Notably, although no token-level information is communicated via count-based feedback, the learner reaches an astonishing performance on boundary detection, as demonstrated by examples (a) through (f) in Table 4.
Table 4. A selection of predictions (and common mistakes) from a model trained on CoNLL-2003 via Counts +OP with sequence-level feedback 1-Grouped.
Table 5. The overall metrics achieved by training on CoNLL-2003 via Counts +OP with sequence-level feedback 1-Grouped.
We observe that incorrect label predictions are almost always encountered in one of the following cases: scenario (1), wherein the learning system (entirely) ignores MISC- and ORG-class tokens, as illustrated by examples (c) and (d), or scenario (2), wherein the learning system only detects the LOC-related segment in MISC- and ORG-class entities, as illustrated by examples (e) and (f). Note that scenario (1) can result from scenario (2). To support this observation, we present the resulting confusion matrix (per token) in Table 6. Additionally, as anticipated in Section 4, we observe an apparent decrease in performance for infrequent token labels, specifically the previously mentioned MISC-class entities, see Tables 5, 6. However, looking more closely at Tables 3, 5, this correlation is only partially supported. The results indicate that the overall PER and MISC-class mentions are roughly proportional to the corresponding F1-Scores, yielding PER-to-MISC ratios of 1.92 (mentions) and 2.03 (F1-Score). In contrast, despite the notably more frequent mentions of LOC over ORG-class entities (7140 vs. 6321 mentions, ratio 1.13), we observe an unexpected, marginal decrease in performance (75.01 vs. 75.68 F1-Score, ratio 0.99). A similar pattern can be observed when comparing LOC with PER-class entities, having ratios of 1.08 (mentions) and 0.82 (F1-Score). Overall, these results suggest that absolute mention frequency does not consistently correspond with performance.
Table 6. The confusion matrix obtained by training on CoNLL-2003 via Counts +OP with sequence-level feedback 1-Grouped.
4.1.1 Reward calculation
To demonstrate the procedure for calculating feedback as illustrated in Figure 3, we manually determine the count-based feedback under reward scheme 1-Grouped for examples (b) and (e) in Table 4. Let countc(y) denote the overall entity counts for class c in a sequence y. For CoNLL-2003, we further denote by count(y) the ordered sequence of (entity) counts countc(y) for c=PER, LOC, ORG, MISC, and O (whenever +O counting variants are considered). We obtain the following ordered sequences for examples (b) and (e):
where the values , , and indicate the number of contiguous O-intervals (beware the punctuation) for the respective gold-label sequence y and predictions ŷ. Subsequently, we calculate the modified cosine similarity σ between count(y) and count(ŷ) to obtain our (global) reward signal r over the predictions ŷ for the learning system (see Equation 7):
Further, when +P counting variants are considered, we assign a (local) 0-reward to predictions ŷt that are impossible given the ground-truth counts. For instance, in example (e), we observe that countLOC(y) = 0 and countLOC(ŷ) = 2, thus resulting in 0-rewards for any prediction ŷt that matches the label class LOC.
Note: In this work, we compute singular rewards (per partition) based on overall entity counts, jointly. One could, however, provide one reward signal per entity class instead, e.g., by computing and evaluating the deviation between countc(y) and countc(ŷ) directly, thus assigning feedback at the sub-sequence level without requiring token-level annotations. This modification would naturally extend and generalize the +P counting variants.
4.2 Related work
As suggested in Section 1 and discussed in Section 4, reinforcement learning techniques can—by virtue of an arbitrary reward function R—potentially overcome the aforementioned limitations of (standard) supervised machine learning, namely the prevalent mismatch between the training objective and the evaluation measure, as well as the requirement of a labeled dataset D with fine-grained annotations (e.g., at token-level).
Yet, despite this apparent potential, and although various RL methods have been proposed to complement sequence labeling approaches for weakly supervised learning—where training is conducted on approximate annotations, meaning incomplete, inexact, or inaccurate (Zhou, 2018), as fine-grained, high-quality annotations are generally expensive to assemble—we notice that RL techniques are generally not utilized to (a) address NLP tasks directly, that is, without extending or requiring a pre-trained model, and (b) overcome the aforelisted limitations for supervised machine learning (in NLP), particularly the reliance on fine-grained annotations.
For instance, Yang et al. (2018) propose an approach that involves partial annotation learning to address the incomplete annotations, followed by an RL-based instance selector that identifies positive (or clean) samples for training, thus handling the inaccurate annotations. In similar fashion, Peng et al. (2021a,b) propose an RL-based instance selector for pre-training a classifier, followed by (a) training on negative samples (Peng et al., 2021a), or (b) an adversarial training mechanism (Peng et al., 2021b) to improve the classifier's robustness against incomplete or inaccurate annotations. In contrast, Wang et al. (2018) and Wan et al. (2020) employ an RL-based system for detecting and rectifying (a) incorrect predictions generated by some pre-trained tagging system (Wang et al., 2018), or (b) incorrect token-labels from annotations auto-generated via distant supervision (Wan et al., 2020).
In this work, we have thus introduced (a) a framework for sequence labeling that directly addresses the problem from a standalone and value-based RL perspective, without requiring a pre-trained model, and (b) the utilization of count-based rewards for training that are obtained by counting entity mentions at the sequence-level (as opposed to considering token-level annotations).
Notably, token-label counts have previously been leveraged to formulate a consistency loss function to maintain consistent entity mentions across paraphrased sequences (Chen et al., 2020). Beyond this, we are not aware of comparable count-based approaches in NLP. However, count-based learning has been investigated in various computer vision settings, such as Weakly Supervised Object Detection (Hsu and Li, 2020), where object-counts are considered over ground-truth candidate proposals (e.g., object classes and specific locations), and Crowd Counting (Savner and Kanhangad, 2023), where count-based annotations are utilized instead of point-level annotations. In other works, a clustering framework for Multiple Instance Learning is presented (Oner et al., 2020), where the training approach relies solely on collection-level annotations that indicate the number of distinct classes within a collection of instances, labeled unique class counts. Count-based learning has also been employed for weakly-supervised temporal localization (Schroeter et al., 2019), specifically the localization and detection of instantaneous event occurrences (lasting for one time-step) in sequential data, with training being conducted on occurrence-counts only. Unfortunately, when transferred to NER, this method requires token-level annotations, since the problem definition assumes that event occurrences (named entities) are instantaneous (composed of a single token).
5 Conclusion and future work
In this work, we presented a unique method to sequence labeling that leverages count-based annotations, e.g., obtained by counting (rather than marking) specific entity mentions in a text, for training. Therefore, we introduced a framework that directly formulates the sequence labeling task from an RL perspective. To validate our approach for NER, we experimented with various degrees of feedback aggregation (multiple predictions are assigned a single reward) in combination with standard and count-based reward functions, where standard feedback is calculated via token-level labels, and count-based feedback is calculated solely by comparing the entity counts per class between the predictions and ground-truth labels. The results indicate that learning sequence labeling tasks, such as Named Entity Recognition, with aggregate feedback is feasible, even from count-based annotations. Furthermore, our findings suggest that informed counting can significantly increase performance.
We acknowledge that the experimental results have potential for considerable improvements, especially regarding the method by which count-based feedback is calculated and attributed to individual label predictions, even when feedback is provided at sequence-level. Although our approach does not completely eliminate the need for labeled datasets, we demonstrate that learning from count-based (or aggregate) annotations can achieve reasonable performance for Named Entity Recognition. By proposing this training approach, we are pushing toward more general and less biased annotations, e.g., counting instead of marking specific entities may lower inter-annotator disagreement. In further studies, the effectiveness of aggregate labels should be explored for more advanced NLP tasks, such as Question Answering or Event Extraction.
Data availability statement
Publicly available datasets were analyzed in this study. This data can be found here: https://huggingface.co/datasets/.
Author contributions
MG: Writing – original draft, Writing – review & editing, Conceptualization, Methodology, Software, Validation. PC: Supervision, Writing – review & editing, Conceptualization.
Funding
The author(s) declare financial support was received for the research, authorship, and/or publication of this article. The authors acknowledge the financial support of the German Research Foundation (DFG) and the Open Access Publication Fund of Bielefeld University for the article processing charge.
Conflict of interest
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Publisher's note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
Footnotes
References
Akalin, N., and Loutfi, A. (2021). Reinforcement learning approaches in social robotics. Sensors. 21:1292. doi: 10.3390/s21041292
Amin, S., Gomrokchi, M., Satija, H., van Hoof, H., and Precup, D. (2021). A survey of exploration methods in reinforcement learning. arXiv [Preprint]. arXiv:2109.00157v2. doi: 10.48550/arXiv.2109.00157
Buck, C., Bulian, J., Ciaramita, M., Gajewski, W., Gesmundo, A., Houlsby, N., et al. (2018). Ask the right questions: Active question reformulation with reinforcement learning. arXiv [Preprint]. arXiv:1705.07830v3. doi: 10.48550/arXiv.1705.07830
Chen, J., Wang, Z., Tian, R., Yang, Z., and Yang, D. (2020). “Local additivity based data augmentation for semi-supervised NER,” in Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), eds. B. Webber, T. Cohn, Y. He, and Y. Liu (Stroudsburg, PA: Association for Computational Linguistics), 1241–1251.
Choi, E., Hewlett, D., Uszkoreit, J., Polosukhin, I., Lacoste, A., and Berant, J. (2017). “Coarse-to-fine question answering for long documents,” in Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Vancouver, BC: Association for Computational Linguistics), 209–220.
Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. (2019). “BERT: Pre-training of deep bidirectional transformers for language understanding,” in Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, eds. J. Burstein, C. Doran, and T. Solorio (Minneapolis, MN: Association for Computational Linguistics), 4171–4186.
Eick, S. G. (1988). The two-armed bandit with delayed responses. Ann. Statist. 16, 254–264. doi: 10.1214/aos/1176350703
Eschmann, J. (2021). “Reward function design in reinforcement learning,” in Reinforcement Learning Algorithms: Analysis and Applications (Cham: Springer), 25–33. doi: 10.1007/978-3-030-41188-6_3
Esteso, A., Peidro, D., Mula, J., and nero, M. D.-M. (2022). Reinforcement learning applied to production planning and control. Int. J. Prod. Res. 61, 1–18. doi: 10.1080/00207543.2022.2104180
Fedus, W., Ramachandran, P., Agarwal, R., Bengio, Y., Larochelle, H., Rowland, M., et al. (2020). “Revisiting fundamentals of experience replay,” in International Conference on Machine Learning (New York, NY: PMLR), 3061–3071.
Gallo, I., Binaghi, E., Carullo, M., and Lamberti, N. (2008). “Named entity recognition by neural sliding window,” in 2008 The Eighth IAPR International Workshop on Document Analysis Systems (Nara: IEEE), 567–573.
Graves, A. (2012). Supervised Sequence Labelling with Recurrent Neural Networks. Berlin: Springer Berlin Heidelberg.
Hamrick, J. B., Friesen, A. L., Behbahani, F., Guez, A., Viola, F., Witherspoon, S., et al. (2021). On the role of planning in model-based deep reinforcement learning. arXiv [Preprint]. arXiv:2011.04021v2. doi: 10.48550/arXiv.2011.04021
Hovy, E., Marcus, M., Palmer, M., Ramshaw, L., and Weischedel, R. (2006). “OntoNotes: the 90% solution,” in Proceedings of the Human Language Technology Conference of the NAACL, Companion Volume: Short Papers (New York, NY: Association for Computational Linguistics), 57–60.
Hsu, C. Y., and Li, W. (2020). “Learning from counting: leveraging temporal classification for weakly supervised object localization and detection,” in 31st British Machine Vision Conference, BMVC 2020 [British Machine Vision Association (Online)]. Available at: https://www.scopus.com/record/display.uri?eid=2-s2.0-85136319201&origin=inward&txGid=82b81a3455cc6aba7a15bbf9d48f1d09
Keneshloo, Y., Shi, T., Ramakrishnan, N., and Reddy, C. K. (2020). Deep reinforcement learning for sequence-to-sequence models. IEEE Trans. Neural Netw. Learn. Syst. 31, 2469–2489. doi: 10.1109/TNNLS.2019.2929141
Kupiec, J. (1992). Robust part-of-speech tagging using a hidden markov model. Comp. Speech & Lang. 6, 225–242. doi: 10.1016/0885-2308(92)90019-Z
Li, J., Monroe, W., Ritter, A., Jurafsky, D., Galley, M., and Gao, J. (2016). “Deep reinforcement learning for dialogue generation,” in Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, eds. J. Su, K. Duh, and X. Carreras (Austin, TX: Association for Computational Linguistics), 1192–1202.
Li, X., Feng, J., Meng, Y., Han, Q., Wu, F., and Li, J. (2020). “A unified MRC framework for named entity recognition,” in Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, eds. D. Jurafsky, J. Chai, N. Schluter, and J. Tetreault (New York, NY: Association for Computational Linguistics), 5849–5859.
Li, Z., Jiang, X., Shang, L., and Li, H. (2018). “Paraphrase generation with deep reinforcement learning,” in Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, eds. E. Riloff, D. Chiang, J. Hockenmaier, and J. Tsujii (Brussels: Association for Computational Linguistics), 3865–3878.
Lin, L.-J. (1992). Self-improving reactive agents based on reinforcement learning, planning and teaching. Mach. Learn. 8, 293–321. doi: 10.1007/BF00992699
Liu, Q., Chen, Y., Chen, B., Lou, J.-G., Chen, Z., Zhou, B., et al. (2020). “You impress me: Dialogue generation via mutual persona perception,” in Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, eds. D. Jurafsky, J. Chai, N. Schluter, and J. Tetreault (New York, NY: Association for Computational Linguistics), 1417–1427.
Lu, K., Zhang, S., and Chen, X. (2019). “Goal-oriented dialogue policy learning from failures,” in Proceedings of the Thirty-Third AAAI Conference on Artificial Intelligence and Thirty-First Innovative Applications of Artificial Intelligence Conference and Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, AAAI'19/IAAI'19/EAAI'19 (Washington, DC: AAAI Press), 2596–2603.
Minsky, M. (1961). Steps toward artificial intelligence. Proc. IRE 49, 8–30. doi: 10.1109/JRPROC.1961.287775
Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., et al. (2013). Playing atari with deep reinforcement learning. arXiv [Preprint]. arXiv:1312.5602v1. doi: 10.48550/arXiv.1312.5602
Narasimhan, K., Yala, A., and Barzilay, R. (2016). “Improving information extraction by acquiring external evidence with reinforcement learning,” in Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, eds. J. Su, K Duh and X. Carreras (Austin, TX: Association for Computational Linguistics), 2355–2365.
Oner, M. U., Lee, H. K., and Sung, W.-K. (2020). Weakly supervised clustering by exploiting unique class count. arXiv [Preprint]. arXiv:1906.07647v2. doi: 10.48550/arXiv.1906.07647
Open AI (2023). Gpt-4 technical report. arXiv [Preprint]. arXiv:2303.08774v6. doi: 10.48550/arXiv.2303.08774
Open AI, Berner, C., Brockman, G., Chan, B., Cheung, V., Dębiak, P., et al. (2019). Dota 2 with large scale deep reinforcement learning. arXiv [Preprint]. arXiv:1912.06680. doi: 10.48550/arXiv.1912.06680
Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., et al. (2022). “Training language models to follow instructions with human feedback,” in Advances in Neural Information Processing Systems, eds. S Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh (New York, NY: Curran Associates, Inc), 27730–27744.
Peng, S., Zhang, Y., Wang, Z., Gao, D., Xiong, F., and Zuo, H. (2021a). “Named entity recognition using negative sampling and reinforcement learning,” in 2021 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) (Houston, TX: IEEE), 714–719.
Peng, S., Zhang, Y., Yu, Y., Zuo, H., and Zhang, K. (2021b). “Named entity recognition based on reinforcement learning and adversarial training,” in Knowledge Science, Engineering and Management, eds. H. Qiu, C. Zhang, Z. Fei, M. Qiu, and S. Y. Kung (Cham: Springer International Publishing), 191–202.
Qian, L., Qiu, L., Zhang, W., Jiang, X., and Yu, Y. (2019). “Exploring diverse expressions for paraphrase generation,” in Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) (Hong Kong: Association for Computational Linguistics), 3173–3182.
Qin, P., Xu, W., and Wang, W. Y. (2018). “Robust distant supervision relation extraction via deep reinforcement learning,” in Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, eds. I. Gurevych, and Y. Miyao (Melbourne, VIC: Association for Computational Linguistics), 2137–2147.
Raffin, A., Kober, J., and Stulp, F. (2022). “Smooth exploration for robotic reinforcement learning,” in Conference on Robot Learning (New York, NY: PMLR), 1634–1644.
Sang, E. F. T. K., and Meulder, F. D. (2003). “Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition,” in Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003 (Edmonton, AB: Association for Computational Linguistics), 142–147. doi: 10.3115/1119176.1119195
Savner, S. S., and Kanhangad, V. (2023). CrowdFormer: Weakly-supervised crowd counting with improved generalizability. J. Vis. Commun. Image Represent. 94:103853. doi: 10.1016/j.jvcir.2023.103853
Schroeter, J., Sidorov, K., and Marshall, D. (2019). “Weakly-supervised temporal localization via occurrence count learning,” in International Conference on Machine Learning (New York, NY: PMLR), 5649–5659.
Sha, F., and Pereira, F. (2003). “Shallow parsing with conditional random fields,” in Proceedings of the 2003 Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics (Edmonton, AB: Association for Computational Linguistics), 213–220. doi: 10.3115/1073445.1073473
Siddique, M. A. B., Oymak, S., and Hristidis, V. (2020). “Unsupervised paraphrasing via deep reinforcement learning,” in Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (New York, NY: ACM), 1800–1809.
Uc-Cetina, V., Navarro-Guerrero, N., Martin-Gonzalez, A., Weber, C., and Wermter, S. (2022). Survey on reinforcement learning for language processing. Artif. Intellig. Rev. 56, 1543–1575. doi: 10.1007/s10462-022-10205-5
Ushio, A., and Camacho-Collados, J. (2021). “T-NER: An all-round python library for transformer-based named entity recognition,” in Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: System Demonstrations (New York, NY: Association for Computational Linguistics), 53–62.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., et al. (2017). “Attention is all you need,” in NIPS'17: Proceedings of the 31st International Conference on Neural Information Processing Systems (Long Beach, CA: Curran Associates Inc.). doi: 10.5555/3295222.3295349
Vinyals, O., Babuschkin, I., Czarnecki, W. M., Mathieu, M., Dudzik, A., Chung, J., et al. (2019). Grandmaster level in starcraft ii using multi-agent reinforcement learning. Nature 575, 350–354. doi: 10.1038/s41586-019-1724-z
Wan, J., Li, H., Hou, L., and Li, J. (2020). “Reinforcement learning for named entity recognition from noisy data,” in Natural Language Processing and Chinese Computing, eds. X. Zhu, M. Zhang, Y. Hong, and R. He (Cham: Springer International Publishing), 333–345.
Wang, Y., Patel, A., and Jin, H. (2018). “A new concept of deep reinforcement learning based augmented general tagging system,” in Proceedings of the 27th International Conference on Computational Linguistics (Santa Fe: Association for Computational Linguistics), 1683–1693.
Wei, C.-H., Peng, Y., Leaman, R., Davis, A. P., Mattingly, C. J., Li, J., et al. (2016). Assessing the state of the art in biomedical relation extraction: Overview of the biocreative v chemical-disease relation (cdr) task. Database 2016:baw032. doi: 10.1093/database/baw032
Yang, Y., Chen, W., Li, Z., He, Z., and Zhang, M. (2018). “Distantly supervised NER with partial annotation learning and reinforcement learning,” in Proceedings of the 27th International Conference on Computational Linguistics (Santa Fe: Association for Computational Linguistics), 2159–2169.
Ye, D., Liu, Z., Sun, M., Shi, B., Zhao, P., Wu, H., et al. (2020). “Mastering complex control in moba games with deep reinforcement learning,” in Proceedings of the AAAI Conference on Artificial Intelligence (Washington, DC: AAAI Press), 6672–6679.
Yu, Y. (2018). “Towards sample efficient reinforcement learning,” in Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI-18 (New York, NY: International Joint Conferences on Artificial Intelligence Organization), 5739–5743
Zhang, S., Cheng, H., Gao, J., and Poon, H. (2023). Optimizing bi-encoder for named entity recognition via contrastive learning. arXiv [Preprint]. arXiv:2208.14565v2. doi: 10.48550/arXiv.2208.14565
Zhou, Z.-H. (2018). A brief introduction to weakly supervised learning. Nation. Sci. Rev. 5, 44–53. doi: 10.1093/nsr/nwx106
Zhu, H., Gupta, V., Ahuja, S. S., Tian, Y., Zhang, Y., and Jin, X. (2021). “Network planning with deep reinforcement learning,” in Proceedings of the 2021 ACM SIGCOMM 2021 Conference, Sigcomm '21 (New York, NY: Association for Computing Machinery), 258–271.
Keywords: reinforcement learning, reward functions, annotations, sequence labeling, information extraction
Citation: Geromel M and Cimiano P (2024) Sequence labeling via reinforcement learning with aggregate labels. Front. Artif. Intell. 7:1463164. doi: 10.3389/frai.2024.1463164
Received: 11 July 2024; Accepted: 28 October 2024;
Published: 15 November 2024.
Edited by:
Jie Yang, Delft University of Technology, NetherlandsReviewed by:
Peide Zhu, Fujian Normal University, ChinaDongyuan Li, The University of Tokyo, Japan
Zhen Wang, Tokyo Institute of Technology, Japan, in collaboration with reviewer DL
Copyright © 2024 Geromel and Cimiano. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Marcel Geromel, bWdlcm9tZWwmI3gwMDA0MDt0ZWNoZmFrLnVuaS1iaWVsZWZlbGQuZGU=