minor fix for "L"inear and parentheses for citation

nncf/experimental/torch/sparsify_activations/ActivationSparsity.md

@@ -20,7 +20,7 @@ $$
 
 `sparsify_activations` automates the process of identifying the pruning thresholds based on user-specified layers, target sparsities and input dataset.
 
-> Note: This feature is **experimental** and intended solely for evaluation of sparsity-task performance. While activation sparsity can improve inference efficiency of decoding phase for Large Language Models (LLMs) [Liu et al., 2023](https://arxiv.org/abs/2310.17157), it neccessitates optimized runtime kernels, which are in development.
+> Note: This feature is **experimental** and intended solely for evaluation of sparsity-task performance. While activation sparsity can improve inference efficiency of decoding phase for Large Language Models (LLMs) ([Liu et al., 2023](https://arxiv.org/abs/2310.17157)), it neccessitates optimized runtime kernels, which are in development.
 
 #### Example Usage
 
@@ -70,7 +70,7 @@ In this example, we first conduct data-free INT8 asymmetric weight quantization
     }
     ```
 
-- `ignored_scope`: Optional. If specified, it should be an instance of `nncf.IgnoredScope` class that defines the nodes in the model graph to be ignored by this algorithm. Note that unsupported layer types are already filtered out internally, so there is no need to mention them in `ignored_scope`. The algorithm currently only supports Linear layers, as they benefit most from dynamic sparse activations by reducing memory read bandwidth for the large Linear weights used in LLMs.
+- `ignored_scope`: Optional. If specified, it should be an instance of `nncf.IgnoredScope` class that defines the nodes in the model graph to be ignored by this algorithm. Note that unsupported layer types are already filtered out internally, so there is no need to mention them in `ignored_scope`. The algorithm currently only supports Linear layers, as they benefit most from dynamic sparse activations by reducing memory read bandwidth for the large Linear layer weights used in LLMs.
 
 #### Evaluation results
 
@@ -143,4 +143,4 @@ Here is the word perplexity for different language models on a subset of [wikite
 
 1. Currently activation sparsity only supports Torch backend. Consequently, this restricts the available compression modes to 8-bit integer modes when using `nncf.compress_weights()` before activation sparsification. More information on supported modes can be found at [Weights Compression](../../../../docs/usage/post_training_compression/weights_compression/Usage.md#limitations).
 2. Actual activation sparsity during inference is dynamic and per input basis, deviation from the target should be expected. In our local experiments, the statistical mean of actual activation sparsity aligned to the target when thresholds are calibrated on datasets similar to the final task.
-3. Similar to other compression methods, model accuracy and activation sparsity are trade-off at play. For LLMs like [Llama](https://llama.meta.com), it is recommended to start with 30%~50% sparsity for the linear layers in feed-forward networks.
+3. Similar to other compression methods, model accuracy and activation sparsity are trade-off at play. For LLMs like [Llama](https://llama.meta.com), it is recommended to start with 30%~50% sparsity for the Linear layers in feed-forward networks.