How Does Our Brain Tell ‘Doorbells’ from ‘Eggs’?
Many of us take our ability to read, write, listen, and comprehend for granted. Although these tasks appear effortless, our brains engage in complex processes to accomplish them.
I am sure you can agree that human communication is a remarkable feat. Yet today, neuroscientists are still struggling to fully comprehend how the human brain understands language and makes sense of words. Whether we are watching our favorite movie or listening to a friend’s office gossip, our brain engages in a complex and intricate activity known as semantic processing to help us make sense of all the information. While we often take this ability for granted, how exactly does the brain accomplish it? Which parts of the brain help us understand the meaning of words? And how does the brain recognize and process words? Researchers from Harvard may have recently cracked the code[1], potentially bringing us closer to answering these pressing questions.
Brain regions involved in language and comprehension
Decades of research have informed us that different brain regions are responsible for specific tasks. Initial linguistic input is first received and processed by the auditory cortex[2], which is part of the temporal lobe of the brain (Fig. 1). The auditory cortex is essential for perceiving sound (e.g. location of source of sound, frequency of sound, pitch) and even perform high-level processing such as speech recognition. Subsequently, auditory information is further processed in the prefrontal cortex in the frontal lobe. One such region is called the Broca’s area, which is found in the left language-dominated hemisphere of the brain (most people) (Fig. 1).
Different parts of the Broca area focus on various aspects of language. The anterior (front) part deals with word meaning, while the posterior (back) part handles how words sound. The Broca’s area also helps with repeating words, making gestures, forming sentences, speaking smoothly, and understanding others’ actions[3]. Insults and injuries to the Broca’s region have been found to result in Broca’s aphasia, a speech disorder characterized by the inability to speak[4].
Another important brain region for language comprehension is known as the Wernicke’s area, which is located at the posterior (back) part of the temporal lobe (Fig 1). Wernicke’s area is involved in understanding written and spoken language[5]. Insults and injuries to this brain region have been found to result in Wernicke’s aphasia, a speech disorder characterized by the inability to comprehend, making conversations challenging[6].
Knowledge gap — lack of mechanistic insights into semantic processing
Although major speech and language pathways in the brain have been largely identified, how individual brain cells (neurons) work to process and comprehend words and language is unknown. Therefore, this study aims to undertake the daunting task that is to investigate how each neuron or group of neurons in the human brain encodes semantic information during natural language comprehension to represent the meaning of words.
Additionally, we know that context matters in how we understand word sentences; words can mean differently depending on the context. For instance, the meaning of the word ‘bat’ differs in the context of baseball v.s. animals. However, we still do not fully understand how neurons precisely adjust their activity to interpret different word meanings based on varying contexts.
Also, our brain understands words as part of a structured network where words are categorized into hierarchies/ groups based on relationships and similarities[7]. For example, when we think of the word ‘dog’, other related concepts such as ‘bark’, ‘animal’, and ‘pet’ are activated. This efficient organization of networks helps us retrieve information about words quickly. Nonetheless, how individual neurons/ groups of neurons represent these networks and relationships remains unknown.
How is single-neuronal recording performed?
In the study described by Mohsen and colleagues, they used a technique called single-neuronal recording. Specifically, recordings were made from the prefrontal cortex of the left language-dominant hemisphere. The recordings were conducted using microelectrodes arrays (Fig. 2a) which are essentially probes that measure electrical signals in brain tissues. These electrodes are tiny and are arranged in a grid-like pattern, allowing them to record electrical activity (also known as action potential, Fig. 2b) of individual neurons in a brain region. Depending on the type of microelectrode array used, hundreds of neurons can be measured at a given time.
This method of recording neuronal activity differs greatly from imaging-based techniques such as functional magnetic resonance imaging (fMRI). fMRI measures changes in blood flow as a proxy to determine active brain regions. However, fMRI does not provide information at a single-cell (neuron) level and are therefore of lower-resolution. Additionally, imaging-based methods such as fMRI cannot accurately measure brain activities in real-time since time is needed for blood to flow to different brain regions. In contrast, the single-neuronal recording measures as soon as neurons become active and fire, providing more precise, real-time information. Also, unlike fMRI, single-neuronal recording provides data on the timing and pattern of neuron firing, at a resolution that allows us to map neural circuits and pathways.
All ten participants of this study were right-handed native English speakers undergoing planned intraoperative neurophysiology for the treatment of epilepsy. Because the participants were awake during the recordings, they were able to perform language-related tasks. During the recordings, participants listened to a variety of natural sentences played in a random order. On average, they heard about 459 unique words and 1,052 total words across approximately 131 sentences each.
There is a neuron for everything
One of the first things the researchers observed was that some neurons are selective in responding to certain semantic categories (e.g. food, objects, action). This level of selectivity was measured using something called the ‘selectivity index (SI)’. SI ranges from 0–1, 1 means that a neuron responds to only one category, while 0 means that a neuron responds equally to all categories. About 14% of neurons were found to be ‘selective’ in that they could respond specifically to ‘food’ or ‘actions’ (Fig. 3a), with an average score of 0.32. A different recording method yielded an average score of 0.42. In total, about 48 out of 287 neurons across all the participants showed a level of selectivity during word comprehension.
Out of the selective neurons, a majority (84%) responded to words from one out of nine categories (also known as ‘domains’), while 16% responded to words from two out of nine categories (Fig. 3b). To make sure that this level of selectivity was robust and reproducible, the researchers first randomly chose a subset of words from the original word list, and secondly, selected words that clearly belonged to specific domains (e.g. ‘apple’, ‘banana’ for the food domain’). Through this method of validation, they were able to show that the neurons behaved consistently in that they were tuned specifically to distinct semantic domains.
Neurons can distinguish real words from gibberish
The researchers also observed that neurons can tell real words from made-up ones like ‘blicket’, ‘florp’, or ‘plooing’ that sound. like words but do not have any meaning. Out of 48 neurons that were tested, 27 were able to distinguish between groups of real words and non-words. This ability was not limited to just the semantically selective neurons, suggesting that a diverse cluster of neurons work together (even neurons that were nonselective) to process word meanings and differentiate sense from nonsense. This interesting observation speaks to the notion that neural networks for language involve the complex interplay between ‘specialized’ and ‘less specialized’ neurons.
Context is important!
Our ability to understand words is largely based on the context of the sentences in which they are used. For example, the phrase ‘He picked a rose…’, suggests that ‘rose’ is a noun (an object/ flower), while ‘He finally rose to….’, suggests that ‘rose’ is a verb (an action). If you are able to understand the meaning of ‘rose’ in both contexts, this is all thanks to your ability to distinguish homophones, words that sound the same but have different meanings. Another example of homophones is ‘sun’ and ‘son’.
To uncover how reliant neurons are on the context of the sentence to represent meaning, the researchers introduced participants to sentences constructed with random words such as ‘to pirate with in bike took is one’ and ‘birthday nap found bake for to girl dog’, where the context is essentially stripped away. The researchers found that the SI dropped significantly from an average of 0.34 to 0.19. This suggests that the response selectivity of the neurons is greatly influenced by the word’s context.
Next, to determine how neurons represent word meanings independent of how similar they are phonetically. They did so by introducing homophones e.g. ‘sun’ and ‘son, ‘knew’ and ‘new’, ‘read’ and ‘red’. Interestingly, they found that neurons displayed high levels of activity when processing homophones compared to words that sounded different but belonged to the same semantic domains. This shows that neurons are capable of encoding word meanings regardless of how they sound, demonstrating our neurons’ remarkable ability to discern the complicated relationship between context, meaning, and even sound.
Processing complex word hierarchies and relationships
The next question the researchers aimed to answer was how neurons make sense of higher-level semantic relationships. They do so by representing the responses of 133 neurons to words in a high-dimensional space (300 dimensions) mathematically (Fig. 4). Each word is represented by a point in this 300-dimensional space. Thereafter, a model showing how neurons might organize words is created (Fig. 4). Then, principal component analysis (PCA) was applied to simplify complex data by reducing it to its key components; the first 5 principal components (PCs) accounted for 81% of selective neuron activity.
When words are projected back into this PC space (first 5 PCs), the distances between words (based on neuronal responses) matched their semantic relationships (Fig. 4) e.g. ‘cat’ and ‘dog’ are closer together in this space compared to ‘cat’ and ‘rocket’.
To further analyze word relationships, ‘cophenetic distance’ in the 300-dimensional space between words was measured and compared with firing activity for each word pair. ‘Cophenetic distance’ refers to how closely the two words are related in a hierarchy. The closer two words mean, the more similar the neuron response is i.e. ‘ducks’ and ‘eggs’ trigger a more similar neuronal activity (smaller differences) compared to ‘eggs’ and ‘doorbells’ (Fig 5). In summary, neurons respond depending on how closely related words are. Words that are closer in the hierarchy elicit more similar neuronal activity compared to words that are further apart.
What’s next?
With a clearer picture of how individual neurons contribute to language processing, we can begin exploring even deeper questions. How does this system change when someone learns a new language? How does it differ in people with language-related disorders?
Having a clearer understanding of how neurons work to process language and represent words is also important in paving the way to develop new tools and treatments for speech disorders. Identifying specific neurons or clusters of neurons allows for more targeted therapies.
Lastly, these new insights can help advance brain-computer interface (BCI) technology such that BCIs can be designed to better interpret users’ intentions in relation to communication and language, thereby improving the quality of speech for people with speech impairments.
Conclusions
This breakthrough study describes how neurons in the prefrontal cortex of our brain process semantic information by tapping on advanced single-neuronal recordings. They found individual neurons that respond selectively to specific words and categories of words. Neuronal responses were also found to be highly dependent on context; neurons decipher word meanings depending on the sentence the words are found in even when they sound similar. Altogether, we now know that single neurons encode word meanings during language comprehension, allowing us to extract meaning from speech.
This is an extremely exciting step forward in neuroscience, with the potential to change how we treat speech disorders. As scientists continue to explore this frontier, we may soon unlock even more secrets of the brain’s incredible ability to understand language.
